FINAL REPORT- SRDC PROJECT DPI021 REMOTE SENSING- BASED PRECISION AGRICULTURE TOOLS FOR THE SUGAR INDUSTRY BY ANDREW ROBSON, CHRIS ABBOTT, ROB BRAMLEY, DAVID LAMB Final Report SRDC Project DPI021 Remote Sensing- based Precision Agriculture Tools for the Sugar Industry ANDREW ROBSON1, CHRIS ABBOTT1, ROB BRAMLEY2, DAVID LAMB3 1DepartmentAgriculture, Fisheries and Forestry, Queensland. Kingaroy. 4610. 2CSIRO Ecosystem Sciences, Adelaide. SA. 5064. 3Precision Agriculture Research Group, University of New England. Armidale NSW. 2351. Project commencement date: Project completion date: SRDC Program: SRDC Strategy: August 2009 March 2013 Regional Futures Farming and harvesting systems Administrator: Helen Kamel Organisation: Dept. of Agriculture, Fisheries and Forestry, Qld Postal Address: PO Box 241 Darling Heights Qld 4350. Ph: +61 7 4631 5380 Fax: +61 7 44631 5378 E-mail: Helen.kamel@daff.qld.gov.au Principal Researcher: Organisation: Postal Address: Ph: +61 7 41600735 Fax: Dr Andrew Robson (Senior Research Scientist) Dept. of Agriculture, Fisheries and Forestry, Qld PO Box 23 Kingaroy Queensland 4610. +61 7 41623238 E-mail: andrew.robson@daff.qld.gov.au Additional Researchers: Chris Abbott (DAAF Qld): David Lamb (UNE): Rob Bramley (CSIRO): chris.abbott@daff.qld.gov.au dlamb@une.edu.au Rob.Bramley@CSIRO.au SRDC Project DPI021 Final Report_without appendices.doc 2 Table of Contents 1. Executive Summary. 2. Publications/ Project extension. 3. Background. 4. Objectives. 5. Methodology. 5.1. Site locations. 5.2. Image pre-processing. 5.2.1. Converting ‘At Sensor’ digital numbers to ‘Top of Atmosphere’ reflectance values. 5.2.2. Geographic registration of imagery and vector data. 5.3. Extracting spectral data from imagery using mill vector data. 5.4. Derivation of vegetation indices. 5.5. Collection of field data for ground truthing imagery. 5.6. Developing image based yield maps from field sampled data: 5.7. Development of a generic yield algorithm: 5.8. Regional predictions of average yield using a ‘generic’ algorithm: 5.9. Derivation and distribution of yield maps at the regional scale using yield prediction algorithms. 5.10. Validation of image derived yield maps. 5.11. Correlating image results against harvester derived yield maps (CSE022). 5 7 9 10 11 11 15 15 15 15 16 17 20 20 21 21 21 22 6. Results. 6.1. Benchmarking and identifying the most feasible and suitable commercial imagery (i.e. spatial resolution, repeat time and economic feasibility) for identifying crop variability and thus directing targeted mid-season management within the Australian cane industry. 6.1.1. Evaluation of the Raptor sensor. 6.2. Identifying the optimum time of image capture that will accurately depict mid- season crop variability whilst avoiding seasonal times most prone to cloud cover, across key Australian cane farming regions.34 6.2.1. Using spatial technologies to identify growth variability in cane crops, likely constraints to production and suggested remedial action. 6.3. Assess the utility of this imagery for explaining the yield variability measured through the CSE022 “A coordinated approach to Precision Agriculture RDE for the Australian Sugar Industry’ project. 6.3.1. CSEO22 sites. 6.3.2. Additional yield validation sites. 6.3.3. Derivation of a generic algorithm. 6.3.4. Validation of the generic algorithm at the regional level. 6.3.5. Validation of the generic algorithm at the block and within block level. 6.3.6. Production of classified yield maps using the generic algorithms. SRDC Project DPI021 Final Report_without appendices.doc 23 23 24 35 46 47 52 55 56 65 68 3 6.4. Implement optimal image processing and delivery protocols for the rapid distribution of classified imagery to agronomists, growers etc. 6.5. Provide recommendations to participating growers, consultants and industry representatives on the potential cost / benefit of implementing RS technologies into current agronomic management practices. 71 72 7. Conclusion. 8. Acknowledgements. 9. References. 10. Appendix 1: Tutorials. Tutorial 1: ENVI: Converting ‘At Sensor’ Digital Numbers to ‘Top of Atmosphere’ reflectance values Tutorial 2: ENVI: Georectification of satellite imagery using an orthorectified base layer and derivation of a GNDVI image. Tutorial 3: ArcGIS: Conversion of Mapinfo (.TAB) files into ArcGIS (.SHP) files. Tutorial 4: ArcGIS: Buffering of polygons and removal of those affected by cloud before the extraction of spectral data. Tutorial 5: Starspan GUI: Extracting average spectral values and associated attribute information for multiple blocks. Tutorial 6: ENVI: Producing a classified vegetation index map of a cane crop from a 4 band satellite image. Tutorial 7: ENVI: Extracting point source spectral information from imagery using regions of interest (ROI’s). Tutorial 8: ENVI: Converting VI pixel values into yield (TCH) using an exponential linear algorithm. Tutorial 9: ENVI: Creating Google Earth KMZ files from Geotiffs. 75 76 76 80 80 87 91 92 95 97 100 101 102 11. Appendix 2: Media/ Publications. 2012 ASSCT presentation. Article in Australian Sugarcane. February- March 2011. 2011 ASSCT Poster presentation. 2010 ASSCT Poster presentation. Article in the Burdekin ‘The Advocate’ Newspaper promoting the presentation at Burdekin Productivity Services Annual General Meeting 16th August 2011. SRDC promotion of DPI021 involvement in the Herbert Resource Information Centre Spatial Community in Action Conference. Ingham (18-19 August 2011) Hand out for BSES cane talks The SRDC sugar cane project (DPI021) featured in a Department of Innovation, Industry, Science, and Research (DIISR): Space Policy Unit (SPU) handout promoting Earth Observation R&D being conducted in Australia Article in the Australian Canegrower, December 2010: 16.“Cane monitoring made easy with new sensor” Flier promoting SRDC Webinar conducted 22 February 2013. 105 105 115 119 122 124 124 125 126 127 128 12. Appendix 3: Project contracts and variations. SRDC Project DPI021 Final Report_without appendices.doc 129 4 1. Executive Summary This project aimed to develop remote sensing applications that were both relevant and of commercial benefit to the Australian sugar industry and therefore adoptable. Such applications included the in season mapping of crop vigour so as to guide future management strategies, the identification of specific abiotic and biotic cropping constraints, and the conversion of GNDVI variability maps into yield at the block, farm and regional level. In order to achieve these applications the project team reviewed an array of remote sensing platforms, timing of imagery capture, software and analysis protocols; as well as distribution formats of derived imagery products, to a range of end users. The project developed strong collaborative linkages with all levels of the industry including mills, productivity services, agronomists, growers and researchers and increased its initial coverage from three individual farms in Bundaberg, Burdekin and the Herbert, coinciding with project CSE022, to include over 33,000 crops grown across 6 growing regions (Mulgrave, Herbert, Burdekin, Bundaberg, ISIS and Condong) during the 2011/2012 season. The remote sensing systems evaluated throughout this project included ALOS (1 capture), SPOT5 (12), SPOT4 (1), RapidEYE (4), IKONOS (21), GeoEYE (1) and Raptor aerial active sensor (3). These sytems provided a range of passive and active sensors, spatial and spectral resolutions, revisit times, cost and processing requirements. Overall, imagery from the French owned SPOT5 satellite, supplied by Astrium (http://www.astrium-geo.com/), was identified to be the most suitable for a range of applications. A single SPOT5 scene (3600km2) encompassed the majority of cane crops within a particular growing region therefore eliminating the need for additional image processing such as mosaicing and colour balancing. The imagery was shown to be cost effective at ~AUS$1 per km2 and owing to its 2-3 day revisit time, provide sufficient operational flexibility to acquire cloud free imagery during periods of continual cloud cover. The spectral resolution of SPOT5 (green, red, near and mid infrared) allowed most accepted vegetation indices to be derived, with a greenness normalised difference vegetation index (GNDVI) exhibiting the strongest correlation with yield in terms of tonnes of cane per hectare (TCH). At the block level, the classified GNDVI ‘zonal’ vigour maps produced from the 10 metre spatial resolution images were comparable to that generated from the higher spatial resolution systems. However they proved unable to discern sub metre constraints such as weed infestations and damage resulting from grubs, soldier fly, rat and feral pig. The pan sharpened IKONOS product (0.8 metre spatial resolution) was shown to be effective in identifying such sub metre constraints and was the most cost effective at AUS$22 per km2 when purchased as three 50 km2 areas of interest, as supplied by Geoimage (http://www.geoimage.com.au/geoimage/) and AAM (http://aamgroup.com/). The active airborne Raptor sensor was also evaluated, with a number of tests conducted to identify the optimal image capture protocols for sugar cane. The optimal flying height was identified to be between 100 and 135 ft above ground level with data collected at transects not exceeding 50 metres apart. Although the direction in which the data was collected i.e. along or across cane rows was shown to have little influence, an internal buffer of 15m was SRDC Project DPI021 Final Report_without appendices.doc 5 required to remove ‘edge’ effects from non cane specific targets. The crop vigour maps derived from the Raptor sensor were found to be highly comparable to those produced from IKONOS imagery, supporting the recommendation that the Raptor sensor, once commercialised could be an effective option for the Australian cane industry. Through this project, remote sensing was demonstrated to be an accurate tool for identifying the spatial variability of crop vigour at the individual block and farm level. By deriving 5 colour class NDVI maps, growers were able to identify the extent of low performing regions and through targeted plant and soil testing determine the likely nature of the constraint. It was initially envisaged that growers supplied with this information could implement remedial action within that growing season. However, consistent cloud cover over most Queensland growing regions between January and March greatly limited the availability of this information within the crucial vegetative stage of the crop. This delay meant that when the information was provided, the crop was at a growth stage and height that made remedial action of little value or mechanically impossible. Imagery captured during the early vegetative growth stage, i.e. during December, may provide useful and timely information that can assist within season remedial action but this was not evaluated through this project. The mid season variability maps that were generated through this project did support alternative management strategies post-harvest, such as the zonal applications of silica, fertilizer, mill mud as well as re- land forming of blocks. It is For the prediction of yield, strong correlations between GNDVI and tonnes of cane per hectare were achieved from imagery acquired between March and May. This time frame was believed to coincide with a stabilisation period of cane growth between vegetative development and maturation, a hypothesis supported by other research. The derivation of a non cultivar and non class specific algorithm for converting GNDVI values into yield achieved regional predictions to within 5% agreement to what was actually achieved for the Bundaberg, Isis, Herbert and Condong growing regions. An additional algorithm was developed for the Burdekin region to account for the vastly different climatic conditions. Although the yield estimates were encouraging, additional research is required to understand and then account for the impact of variable seasonal conditions. This may require the integration of an agro-meteorological model to normalise the seasonal trends or more simply the development of slightly refined algorithms that represent ‘good’ and ‘bad’ years. In an attempt to extrapolate the accuracies of the yield predictions to the sub block level, the project team developed methodologies for the derivation of image based yield maps from both high and mid spatial resolution imagery. Using comprehensive GIS vector data layers provided by each mill and the freeware software Starspan GUI, spectral information for the majority of cane crops within each growing region was extracted. For the 2012 growing season each pixel value (GNDVI) was converted into TCH using the appropriate regional GNDVI algorithm, with each block then segregated into 8 yield classes. These surrogate yield maps were then distributed to industry via a number of formats including GoogleEarth (kmz) files, GeoTiffs, static images or embedded in documents. Although the predictions of average block yield and the spatial trends of crop performance were shown to SRDC Project DPI021 Final Report_without appendices.doc 6 be accurate across some locations, others were highly inaccurate. This varied result indicates that further research is required to ensure a greater degree of precision is achieved. This will likely require the development of algorithms for specific class and variety and possibly subgrowing regions. The software ENVI (http://www.exelisvis.com/ProductsServices/ENVI.aspx) was identified to be suitable for remote sensing data analysis, whilst for the querying and manipulation of vector data ArcGIS (http://www.esri.com/software/arcgis) was used. To assist with the future adoption of these technologies analysis protocols developed through this project are provided as training tutorials within the appendices of this report. The Department of Agriculture, Fisheries and Forestry, Queensland (DAAF Qld) with CSIRO and UNE greatly appreciate the opportunity provided by the SRDC to undertake this research and acknowledge the input and collaboration from a number of industry partners. It is believed that the results of this research will support the increased adoption of remote sensing technologies by the Australian sugar industry. 2. Publications/Project Extension Publications: o Andrew Robson, Chris Abbott, David Lamb, Rob Bramley and Mary Barnes (2012). Deriving sugar cane yield maps from SPOT 5 satellite imagery at a regional scale. Poster Abstract. Proceedings of the International Society of Sugar Cane Technologists workshop. Townsville, Qld. 10 – 14 September 2012. o Andrew Robson, Chris Abbott, David Lamb and Rob Bramley (2012). Developing sugar cane yield algorithms from satellite imagery. Proceedings of the Australian Society of Sugar Cane Technologists. 34th Conference, Cairns, Qld. 1- 4 May 2012. o Andrew Robson, Chris Abbott, David Lamb and Rob Bramley (2011). Paddock and regional scale yield prediction of cane using satellite imagery. Poster Abstract. Proceedings of the Australian Society of Sugar Cane Technologists. 33rd Conference, Mackay, Qld. 4 -6 May 2011. o Andrew Robson, Chris Abbott, David Lamb and Rob Bramley (2011). Satellite remote sensing of sugarcane- some FAQs. Australian Sugarcane. February- March 2011. o A. Robson, J.R. Hughes and R.J. Coventry (2010). Using spatial layers to understand variability in precision agricultural systems for sugarcane production. Poster abstract. In Proceedings of the 32nd conference of the Australian Society of Sugar Cane Technologists. Bundaberg. Qld. 11 – 14 May 2010. o Robson, A.J., Wright, G.W., Bell, M.J., Medway, J., Hatfield, P., and Rao. C.N. Rachaputi (2009). Practical remote sensing applications for the Peanut, Sugar cane SRDC Project DPI021 Final Report_without appendices.doc 7 and Cotton Farming Systems. Poster presentation and abstract. 13th Annual Symposium of Precision Agriculture in Australasia. Armidale, NSW. 10- 11th September 2009. Research referred to in: o Bramley RGV, Trengove S. 2012. Precision Agriculture in Australia: present status and recent developments. In ConBap 2012 - Proceedings Congresso Brasileiro de Agricultura de Precisão, Ribeirão Preto - SP, Brasil. 24 to 26 September. o Bramley RGV. 2012. Precision Agriculture: Opportunities for Improved Management of Sugarcane Production. In: RA Gilbert (Ed) Sustainable sugarcane production. Proceedings of the International Society of Sugar Cane Technologists workshop. Townsville, Qld. 10 – 14 September 2012. o Mike Bell, Steve Walker, Andrew Robson, David Jordan and Dave Murray (2009). The evolution of modern cropping systems In National Water and Environment Bulletin. Spring 2009. Page 43- 47. Project extension/ Professional engagements: o 2013 (22 February): SRDC Webinar presentation of project results. o 2012 (24 October): Project results presented at Sugar Research and Development Corporation (SRDC) seminar series. Queensland University of Technology (QUT), Brisbane. o 2012 (14- 15 May): Project results presented at Sugar Research and Development Corporation (SRDC) research exposition: Mackay, Proserpine. o 2012 (9- 11 May): Project results presented at Sugar Research and Development Corporation (SRDC) research exposition: Maryborough, Murwillumbah and MacLean. o 2012 (1- 4 May): Project results presented at the 34th conference of the Australian Society of Sugar Cane Technologists. Cairns, QLD. o 2011 (25 November): Project results presented to growers/ ISIS mill and Canegrowers representatives at Childers. o 2011 (1 October): Project results presented to a World Congress of Conservation Agriculture (WACCA) tour group, Kingaroy, Qld. o 2011 (31 August): Project meeting phone conference. o 2011 (18- 19 August): Project results presented at Herbert Resource Information Centre Spatial Community in Action Conference. Ingham. Qld. o 2011 (16 August): Guest speaker at Burdekin Productivity Service annual general meeting. Ayr. o 2011 (4- 6May): Poster presentation at the 33rd Conference of the Australian Society of Sugar Cane Technologists. Mackay, Qld. o 2011 (14- 16 March): Project results presented at 6 BSES CPI meetings in the Burdekin region (115 participants). o 2011 (15- 22 February): Project results presented at 6 BSES CPI meetings in the Bundaberg region (81 participants). SRDC Project DPI021 Final Report_without appendices.doc 8 o 2010 (3 September): Poster presentation at the 14th Annual Symposium of Precision Agriculture in Australasia. Albury, NSW. 2- 3 September 2010 o 2010 (11- 14 May): Poster presentation at the 32nd conference of the Australian Society of Sugar Cane Technologists. Bundaberg. Qld. o 2010 (4 March): Project meeting at SRDC head Office, Brisbane. o 2009 (1 December): ENVI training course for industry collaborators, Mackay (5 participants). o 2009 (26 November): Project results presented at ‘Southern Queensland Farming Systems’ field day. Kingsthorpe, Qld. o 2009 (3 November): Project introduction meeting for industry collaborators in Bundaberg (19 participants). o 2009 (16 October): Project introduction meeting for industry collaborators in Townsville (15 participants). o 2009 (15 October): Project introduction meeting for industry collaborators in Mackay (5 participants). o 2009 (10- 11 September): Poster presentation at the 13th Annual Symposium on Precision agriculture in Australia. Armidale, NSW. 3. Background The last decade has seen a major global increase in the development and application of spatial technologies, driven by improved access to high quality, cost efficient remote sensing platforms; intelligent analysis softwares; enhanced computer processing and data storage capacities and image delivery systems including GoogleEarth. In terms of agriculture, decades of research across multiple cropping systems has identified remotely sensed (RS) imagery as an effective tool for identifying mid season spatial variability in crop vigour and yield. This information when incorporated into a Geographic Information System (GIS) has facilitated the adoption of precision agriculture, particularly targeted sampling and then variable rate technologies, applications that can reduce input costs whilst maintaining or increasing productivity. It has also provided an effective tool for in season yield forecasting, supporting decisions regarding harvest management, as well as the handling, storage and forward selling of produce post harvest. The Australian sugar industry has widely adopted GIS as an essential framework for the recording and managing spatial data (Davis et al. 2007). The mills themselves implement comprehensive GIS vector layers that spatially define and detail every crop within their specified growing region. This has greatly increased the integration of mill and productivity datasets, thus enabling greater efficiencies in data retrieval and analysis of client information (Markley et al. 2008). A whole-of-community GIS system developed for the Herbert River sugar district has the capacity to record real-time cane harvester operations via GPS. This has improved the coordination and planning of the cane harvest, the identification of consignment errors and been used to improve the safety and efficiency of the rail transport infrastructure (De Lai et al. 2011). The establishment of such integrated GIS system supports the use of remote technologies, by allowing the rapid extraction of spectral data via crop SRDC Project DPI021 Final Report_without appendices.doc 9 boundaries, data interrogation based on variety and class, and re-distribution of derived imagery products using the existing GIS framework. Globally, remote sensing has been shown to be an effective yield prediction tool (Fernandes et al. 2011; Benvenuti and Weill 2010; Bégué et al. 2010; Simões et al. 2009; Abdel-Rahman and Ahmed 2008; Bégué et al. 2008; Almeida et al 2006; Simões et al. 2005; Krishna Rao et al. 2002; and Rudorff and Batista 1990). However, research in Australia has been limited (Noonan. 1999; Markley et al. 2003; Robson et al. 2011; Robson et al. 2010; Lee-Lovick and Kirchner 1991). Commercially, Mackay Sugar Ltd has been the predominant adopter of satellite imagery as a commercial yield forecasting tool, by utilising yield prediction algorithms derived from SPOT2 imagery (Markley et al. 2003). However, this has not been extrapolated to the other Australian growing regions. This project aimed to address that shortfall. In regards to the use of remote sensing technologies for mid season detection of crop variability there has also been little adoption by the Australian sugar industry. This is believed to be the result of a lack of awareness of the technologies available, insufficient evidence supporting the cost/benefits of adoption, limited expertise to analyse current high resolution data, an inability to image cropping regions due to continued cloud cover or lack of access to the information close to ‘near real time'. This project was developed to produce practical and relevant benchmarks, protocols and recommendations for the adoption of remote sensing technologies for improved in season management and therefore production within the Australian sugar cane industry. 4. Objectives 1. Benchmarking and identifying the most feasible and suitable commercial imagery (i.e. spatial resolution, repeat time and economic feasibility) for identifying crop variability and thus directing targeted mid-season management within the Australian cane industry. 2. Also the optimum time of image capture that will accurately depict mid- season crop variability whilst avoiding seasonal times most prone to cloud cover, across key Australian cane farming regions. 3. Assess the utility of this imagery for explaining the yield variability measured through the CSE022 ‘A coordinated approach to Precision Agriculture RDE for the Australian Sugar Industry’ project. 4. Implement optimal image processing and delivery protocols for the rapid distribution of classified imagery to agronomists, growers etc. 5. Provide recommendations to participating growers, consultants and industry representatives on the potential cost / benefit of implementing RS technologies into current agronomic management practices. SRDC Project DPI021 Final Report_without appendices.doc 10 5. Methodology The following section details a number of methodologies and analysis protocols developed by this project in order to deliver on its objectives. 5.1. Site locations. http://www.canegrowers.com.au/page/Industry_Centre/about-sugarcane/Statistics_facts_figures/ Figure 1. Location of Australian sugar growing regions and associated mills. This project had an initial focus area of one individual farm within each of the Australian intensive cropping regions of Bundaberg, Burdekin and the Herbert (Figure 1). The three farms Pozzebon (Burdekin) (Figure 2a), Hubert (Bundaberg) (Figure 2b), and Tabone (Herbert) (Figure 2c) coincided with SRDC project (CSE022) ‘A collaborative approach to Precision Agriculture RDE for the Australian Sugar Industry’. The rationale behind the mutual site selections was that this project would supply image derived vigour maps indicating within crop spatial variability that could be compared to the spatial trends produced from the yield monitors being evaluated by CSE022. In return the field sampled yield data collected by CSE022 would assist in the calibration of imagery to yield. SRDC Project DPI021 Final Report_without appendices.doc 11 a b c Figure 2. False colour SPOT5 image coverage (3600km2) of three Australian cane growing regions, with each of the CSE022 sites highlighted in yellow a. Pozzebon: Burdekin; b. Bundaberg: Hubert; c. Tabone; Herbert. SRDC Project DPI021 Final Report_without appendices.doc 12 The false colour images presented in Figure 2, are three band composite image (green, red, and near infrared) with the brighter the red/ pink colour the more vigorous the vegetative growth. As well as image analysis undertaken on the individual farms identified in Figure 2, a number of additional properties were also investigated. Classified GNDVI vigour maps were derived for each of the 974 can blocks (Table 1), and distributed to growers via GoogleEarth or as static documents. Table 1: GNDVI vigour and derived yield maps distributed to growers during the project. Grower Location Bullseye Farming Bundaberg Cayley Bundaberg Halpin Bundaberg Hubert Bundaberg Lewis Bundaberg Pegg Bundaberg Relmay Farming Bundaberg Scott Bundaberg Russo Childers Kangas Herbert Morley Herbert Tabone Herbert BSES trials Gordnonvale Bugeja Mackay Attard Burdekin Burrows Burdekin BSES trials Burdekin Cacciola Burdekin Catalano Burdekin Darween Burdekin Davco Burdekin Blocks Grower 30 Fiamingo 29 Fowler 26 Haigh 32 Linton JK 13 Lyons 9 Luckel 56 Mann A 12 Mann B 88 Mann K 18 Populin 3 Pozzebon 20 Setter 2 Scarbossa 1 SISl farming 24 Sorbella 97 Strathdee 20 Ianmurb 11 Jordan 30 Kelly 19 Linton A 53 Linton J Location Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Burdekin Total Blocks 2 6 28 20 27 52 6 12 18 48 11 7 24 40 12 14 11 12 9 9 15 974 In order to deliver on the first project objective, imagery from an array of platforms was acquired on multiple occasions over the Burdekin, Herbert and Bundaberg growing regions. Some additional Australian cane growing regions were also included (Table 2). SRDC Project DPI021 Final Report_without appendices.doc 13 Table 2. Inventory of imagery: Location Sensor Herbert LIDAR Bundaberg RapidEYE Bundaberg RapidEYE Bundaberg ALOS Bundaberg SPOT5 Bundaberg SPOT5 Bundaberg (CSE022 sites) IKONOS Burdekin SPOT5 Burdekin TerraSAR-X Burdekin (CSE022 site) IKONOS Burdekin IKONOS Herbert IKONOS Herbert IKONOS Gordonvale IKONOS Mackay RapidEYE Mackay IKONOS Mackay SPOT4 Mackay TerraSAR-X Bundaberg (CSE022 sites) IKONOS Bundaberg (CSE022 sites) IKONOS South Bundaberg IKONOS South Bundaberg IKONOS Farnsfield (Bundaberg) IKONOS Herbert (CSE022 site) IKONOS Burdekin (CSE022 site) IKONOS Burdekin IKONOS Mona Park (Burdekin) IKONOS Bundaberg SPOT5 Burdekin SPOT5 Herbert SPOT5 Bundaberg (CSE022 sites) UNE Raptor system Bundaberg (CSE022 sites) UNE Raptor system Bundaberg (CSE022 sites) UNE Raptor system Gordonvale RapidEYE Mackay IKONOS Herbert (CSE022 site) IKONOS Herbert SPOT5 Bundaberg (CSE022 site) IKONOS Bundaberg SPOT5 Mulgrave GeoEYE Burdekin SPOT5 Burdekin IKONOS Burdekin IKONOS Burdekin (CSE022 site) IKONOS Northern NSW SPOT5 Spatial resolution 5m 5m 10m 10m 10m 0.8m PS 10m 3m 0.8m PS 0.8m PS 0.8m PS 0.8m PS 0.8m PS 5m 0.8m PS 20m 3m 0.8m PS 0.8m PS 0.8m PS 0.8m PS 0.8m PS 0.8m PS 0.8m PS 0.8m PS 0.8m PS 10m 10m 10m Varying Varying Varying 5m 0.8m PS 0.8m PS 10m 0.8m PS 10m 0.5m PS 10m 0.8m PS 0.8m PS 0.8m PS 10m Acqusition date Area covered 9-Aug-09 sample area 1-Mar-09 sample area 30-Jun-09 sample area 17-Jan-10 5000km2 10-May-10 3600km2 14-Apr-10 3600km2 14-May-10 50km2 14-May-10 3600km2 28-Jun-10 sample area 28-May-10 50km2 11-Jun-10 50km2 16-Aug-10 50km2 22-Jun-10 50km2 18-Mar-10 50km2 6-Mar-10 sample area 19-Jun-10 50km2 11-Apr-10 3600km2 May-10 sample area 23-Mar-11 50km2 30-Apr-11 50km2 23-Mar-11 50km2 30-Apr-11 50km2 25-May-11 50km2 1-May-11 50km2 12-May-11 50km2 20-Apr-11 50km2 28-May-11 50km2 27-Mar-11 3600km2 22-Apr-11 3600km2 5-May-11 3600km2 23-Feb-11 3 individual crops 23-Mar-11 3 individual crops 2-May-11 3 individual crops 12-Apr-11 sample area 25-May-11 50km2 11-Apr-12 50km2 4-Apr-12 3600km2 6-Apr-12 50km2 1-Apr-12 3600km2 4-May-12 50km2 16-May-12 3600km2 11-Apr-12 50km2 25-Mar-12 50km2 28-Mar-12 68km2 29-Feb-12 3600km2 * PS refers to pan sharpening, where the spatial resolution of a multispectral imagery can be increased by using an additional panchromatic band. Pan-sharpened images were not used for algorithm development. At the time of this project, an individual SPOT5 image (3600km2) supplied by Astrium (http://www.astrium-geo.com/) cost AU$3465 per scene (AU$0.96/km2), with additional processing such as orthorectification attracting an additional fee (AU$500). The cost of an IKONOS image was US$1100 (US$22/km2) per scene when purchased under a three * 50km2 image capture deal; again, additional processing attracted additional fees (AU$687.50). The evaluation RapidEYE data was provided by AAM (http://aamgroup.com/): The airborne Raptor data acquired by the University of New England cost on average AU$8,000 per flight. However 75% of the cost covered the ferry of the aircraft/sensor to site locations (~ return flight time from Armidale to Bundaberg). The sensor and flight specifications are as follows: Dual wavelength Red (658 nm: 17 nm fwhm) and NIR (850 nm: 32 nm fwhm); SRDC Project DPI021 Final Report_without appendices.doc 14 irradiance footprint divergence angle of ~14° x 8° from source; 5 Hz GPS (Garmin GPS18x5 Hz, Olathe KA USA) plus interpolation setting on GeoScout providing ~17 Hz position calculation rate. The data was collected at an airspeed of ~ 100 knots with a flight transect width of 3m. Data was collected on three occasions (Table 2) with multiple flights over the Relmay site (23 Feb 2011) to assess the impact of height (100 ft, 135 ft and 180 ft Above Ground Level (AGL)) and flight direction (across rows and with cane rows) on the resultant kriged NDVI image. 5.2. Image pre-processing. 5.2.1. Converting ‘At Sensor’ digital numbers to ‘Top of Atmosphere’ reflectance values. To allow temporal comparison, all satellite imagery was corrected for variable atmospheric conditions using a ‘top of atmosphere’ (TOA) correction. To assist in the future adoption of these analysis protocols, the theory and methodology for applying this correction to a number of commercial satellites is detailed below. An ENVI tutorial for undertaking this correction is also supplied as Appendix 1: Tutorial 1: ENVI: Converting ‘At Sensor’ Digital Numbers to ‘Top of Atmosphere’ reflectance values. 5.2.2. Geographic registration of imagery and vector data. So that all vector and raster data obtained throughout this project could be spatially compared and subsequently analysed it was reprojected to: Projection: Transverse Mercator; Datum: Geocentric Datum of Australia 1994 (GDA94). For the raster data, the initial image for each site was purchased as an orthorectified product. This is an image with high spatial integrity as it has been corrected for both topographic relief (provided by a digital elevation model: DEM) and vertical aspect. The orthorectified images were used as base layers, in which all subsequent imagery for each respective region was ‘warped’ to, using an image- to- image rectification process, refer to Appendix 1:Tutorial 2: ENVI: Georectification of satellite imagery using an orthorectified base layer and derivation of a GNDVI image. This process was suitable for most growing areas investigated due to their relatively flat topography. However, for more undulating regions such as Mulgrave, it is recommended that all images should be orthorectified. The registration accuracy of mill vector data, was generally high. However some exceptions indicated the need for a degree of quality assurance before further analysis. One issue identified was the compatibility of information when transferred from one GIS software to another. The Mills generally use the MapINFO, whilst the project team used ArcGIS. The obvious solution is for all parties to use the same software. However where this issue does occur a tutorial has been included in this report for the importing of MapINFO Tab files into ArcGIS as SHP files (Appendix1: Tutorial 3: ArcGIS: Conversion of Mapinfo (.TAB) files into ArcGIS (.SHP) files). 5.3. Extracting spectral data from imagery using mill vector data. SRDC Project DPI021 Final Report_without appendices.doc 15 For the development and then implementation of the generic yield prediction algorithms, a methodology was developed for the rapid extraction of spectral information from each image using the respective mill vector boundaries. Firstly, to ensure the spatial data was specific to cane, all boundaries outside of the image extent were removed, as well as those obscured by cloud or cloud shadows. An internal buffer was then applied to each crop boundary to ensure the image pixels used to extract spectral information did not include spectral data specific to headlands, roads, buildings etc. A tutorial has been included for undertaking these processing steps with the software ArcGIS (Appendix 1: Tutorial 4: ArcGIS: Buffering of polygons and removal of those effected by cloud before the extraction of spectral data). The freeware Starspan GUI was used to extract the average spectral information for the 4 band widths for each selected crop. These average values were exported in a .CSV file format along with the Mill attribute data corresponding to each crop, and interrogated within Microsoft Excel. A tutorial detailing the use of Starspan GUI is provided in the appendices of this report (Appendix 1: Tutorial 5: Starspan GUI: Extracting average spectral values and associated attribute information for multiple blocks). 5.4. Derivation of vegetation indices. Vegetation indices or band ratios are highly effective for identifying variations in plant vigour whilst also minimising errors associated with atmospheric attenuation, plant shading and interference from soil reflectance. The commonly used Normalised Difference Vegetation Index (NDVI) does address some atmospheric attenuation and shading. However it can saturate in large biomass crops such as sugar cane with a leaf area index (LAI) greater than 3 (Benvenuti and Weill 2010; Bégué et al. 2010; Xiao 2005; Xiao et al. 2004b; Xiao et al. 2004a; Huete et al. 2002; Huete et al. 1997). Therefore, following a review of published literature, a number of structural and pigment based indices were investigated to identify which one consistently produced a higher correlation with sugar cane yield (TCH) (Table 3). SRDC Project DPI021 Final Report_without appendices.doc 16 Table 3: Vegetation indices assessed for their correlation to sugar yield (TCH) Normalised Difference RNIR – RRed / RNIR + RRed Vegetation Index (NDVI) GreenNDVI RNIR - RGreen / RNIR + RGreen MidIRNDVI RMIR – RRed / RMIR – RRed Plant Cell Density (PCD) RNIR / RRed MidIRPCD RMIR / RRed NDVIPCD NDVI / RRed MidIRNDVIPCD MidIRNDVI / RRed Transformed chlorophyll -3*(RRed – RGreen) – 0.2*(RRed – RGreen) *(RRed / RNIR + Red)) absorption reflectance index (TCARI) Two-band Enhance 2.5*((RNIR - RRed )/(RNIR +(2.4* RRed)+1)) Vegetation Index (EVI_2) where for SPOT5: Green = wavelengths 0.5 - 0.59µm Red = wavelengths 0.61 – 0.68µm NIR =‘Near-Infrared’ wavelengths 0.78 – 0.89µm MIR = ‘Mid-Infrared’ wavelengths 1.58 – 1.75µm; and for IKONOS: Blue = wavelengths 0.45 - 0.516 µm Green = wavelengths 0.506 -0.595µm Red = wavelengths 0.632 – 0.698µm NIR =‘Near-Infrared’ wavelengths 0.757 – 0.853µm 5.5. Collection of field data for ground truthing imagery. To determine the relationship between the spectral reflectance characteristics of the sugar cane canopy and a corresponding measure of productivity in terms of yield (TCH) and CCS, field sampling was undertaken within a number of crops at strategic locations (Table 4). Table 4. Field sampling undertaken for the ground truthing of imagery. Sample Date Sample location 21 August 2009 Herbert (H2) with BPS001 6 July 2010 Burdekin (A Mann) 24 October 2010 Bundaberg (Relmay) 2 November 2010 Bundaberg (Bullseye) 27 September 2011 Bundaberg (Relmay) 18 July 2011 Bundaberg (Bullseye) 26 November 2011 Burdekin (Pozzebon) SRDC Project DPI021 Final Report_without appendices.doc 17 Each sample location was selected to represent a range of spectral values as defined by the 5 colour class classified NDVI image (Figure 3). At least three replicate samples were chosen to represent the high (Blue), medium (Green) and low (Red) NDVI zones. A tutorial for undertaking this process in ENVI is supplied in the appendices of this report (Appendix 1: Tutorial 6: ENVI: Producing a classified vegetation index map of a cane crop from a 4 band satellite image). Figure 3. Example of a classified NDVI image of a sugar cane crop with sampling coordinates representing a range of homogenous colour zones indicated. Field sampling coincided with the commercial harvesting of each respective crop (Figure 4a and c), so as to reduce the need of locating, cutting and then dragging samples through nearly impenetrable cane. The samples were located with a non- differential Garmin GPS (model: etrex LEGEND) with an accuracy +/- 6 m and marked with flagging tape (Figure 4b). A 5 m linear row of cane was measured with the number of stalks counted to provide an estimate of stalk density. Once exposed by the commercial harvester, the 5 m row was manually cut with a cane knife at ground level (Figure 4c) and then weighed on a trailer mounted load cell (Figure 4d). SRDC Project DPI021 Final Report_without appendices.doc 18 ab cd Figure 4. Photos of field sampling (a) burning of a Burdekin cane crop prior to harvest; (b) sampling lodged burnt cane; (c) coinciding the field sampling with commercial harvester and (d) cutting samples from a green Bundaberg crop, with samples then weighed on a portable trailer load cell. The weight of 5 m samples provided an estimate of total biomass whilst the additional weighing of a 20 stalks sub-sample, with and without its top leaf, provided an estimate of percent millable stalk. Using the known row spacing of the crop, the weight of the 5 m total biomass sample and percentage of millable stalk a measure of Tonnes of Cane per Hectare (TCH) was calculated. 6 stalks from each sample location were also retained and tested for Brix, Temp, Pol, Purity %, Fibre and CCS by the respective BSES station in each sampling region. To verify if soil health and structure were driving crop variability, 10 cm soil cores were also manually collected at depths of 0- 20 cm and 40 -60 cm from high and low NDVI sample sites. The samples were dried and sent to CSBP soil and plant analysis laboratory (http://www.csbp-fertilisers.com.au/nutrition-services/soil-and-plant-analysis-laboratory) for a ‘comprehensive’ soil analysis including Colwell P and K, KCl 40 S, organic carbon, Nitrate N, Ammonium N, EC, pH- water, pH-CaCl2, DTPA trace elements (Cu, ZN, Mn, Fe) and exchangeable cations (Ca, Mg, Na, K, Al), Boron, Acid P (BSES), and Chloride. From these sample results Estimated Cation Exchange Capacity (Est. CEC), Exchangeable Sodium Percentage (ESP) and Exchangeable Sodium to Potassium ratio (Na:K ratio) were calculated. Correlations were undertaken between all soil parameters with yield (TCH) and CCS to identify any likely drivers of reduced production. SRDC Project DPI021 Final Report_without appendices.doc 19 5.6. Developing image based yield maps from field sampled data. The spectral data corresponding to each sampling location was extracted using small areas of interest (average of 2 SPOT5 pixels 10 * 20 m; and the average of 9 IKONOS pixels 3.2 * 3.2 m, note the pan-sharpened images were not used for algorithm development). SPOT5 imagery warped to an orthorectified image had a RMS of less than 10 m, whilst for IKONOS an RMS less than 2 m was achieved. A tutorial for undertaking this process in ENVI is supplied in the appendices of this report (Appendix 1: Tutorial 7: ENVI: Extracting point source spectral information from imagery using regions of interest (ROI’s)). Once extracted, the average 4 band spectral data was entered into a Microsoft Excel spreadsheet along with the corresponding field measurements and soil chemistry results. To identify the best correlations between satellite imagery and crop yield (TCH), a number of vegetation indices (VI) (refer to section 5.4.) were examined, with the index that provided the highest correlation coefficient selected for further analysis. The average yield for the sampled crop was calculated by substituting the average VI value into the linear algorithm produced from the correlation. Total crop yield was then calculated by multiplying the average crop yield by the crop area, with the prediction accuracies validated against mill reports following harvest. Coinciding with the prediction of average and total yield, the algorithm developed for each sampled crop was applied to each corresponding image to convert the VI pixel values into yield. A tutorial for undertaking this process in ENVI is supplied in the appendices of this report (Appendix 1: Tutorial 8: ENVI: Converting VI pixel values into yield (TCH) using an exponential linear algorithm). By further classifying these converted images into a graduated scaling of yield via a density slice, a surrogate yield map was produced. 5.7. Development of a generic yield algorithm. Following the accuracies achieved in the prediction of sampled crops using field samples and site specific algorithms, the project team investigated the accuracies of a non cultivar, crop class and regionally specific yield prediction algorithm. The initial algorithm was developed from the correlation between the average GNDVI values of 112 Bundaberg cane blocks extracted from a TOA corrected SPOT5 image (acquired 10 May 2010), to average 2010 harvest yields for the corresponding blocks. This initial algorithm was applied to a 2008 SPOT5 image (acquired 31 March 2008) to predict the average and total yield of 39 cane blocks (600 ha) grown within the same Bundaberg region during the 2008 season. A ‘generic’ Bundaberg algorithm was subsequently developed from the combined 2008 and 2010 data sets. It was hypothesised that combined data set would make the algorithm less seasonally specific. Further statistical analysis of this algorithm in terms of predictive accuracy of yield was undertaken by Mary Barnes (CSIRO Mathematics Informatics and Statistics), where upper and lower 95% confidence intervals were calculated to show error on individual predictions. The standard 95% confidence intervals around the line of best fit were SRDC Project DPI021 Final Report_without appendices.doc 20 calculated on the log scale (i.e. log(TCH) a straight line relationship), with the resultant prediction intervals back transformed to an exponential relationship. 5.8. Regional predictions of average yield using a ‘generic’ algorithm. To further asses the accuracy of the ‘Bundaberg generic algorithm’, retrospective predictions of average regional yield were derived for the Bundaberg (3544 crops), Isis (2772 crops) and Burdekin (4573 crops) growing regions, following the 2010 harvest. Comprehensive vector layers defining crop boundaries, cultivar, class etc. provided from each respective mill, were used to extract spectral data from SPOT5 imagery captured acquired 10 May 2010 (Bundaberg and Isis) and 14 May 2010 (Burdekin). For the 2011 season, predictions of average yield were again made for the Bundaberg (3824 crops), Isis (4204 crops), Burdekin (4999 crops) and Herbert (8596 crops) regions with imagery captured on the 27 March 2011 (Bundaberg and Isis), 22 April 2011 (Burdekin) and 5 May 2011 (Herbert). For the 2012 season predictions were made for the Bundaberg (3217 crops), Isis (4000 crops), Burdekin (6921 crops) and Herbert (15463 crops) regions with imagery captured on the 1 April 2012 (Bundaberg and Isis), 16 May 2012 (Burdekin) and 4 April 2012 (Herbert). Additional estimates were made for the Condong Mill region (New South Wales) (2087 crop) by applying the ‘generic’ algorithm to a SPOT5 image captured on the 29 February 2012. Due to the poor predictive accuracy of the generic algorithm for the Burdekin region during both the 2010 and 2011 season, a Burdekin specific algorithm was derived from the correlation between the 2010 Burdekin crop GNDVI values and harvested yield. The new algorithm was evaluated during the 2011 and 2012 growing seasons. For the Mulgrave region, a scoping study was undertaken using an algorithm developed from the correlation of average block yield from 832 crops and GNDVI derived from an IKONOS image captured on the 26 May 2010. This algorithm was applied to 1324 crops captured by GeoEYE imagery on 4 May 2012. 5.9. Derivation and distribution of yield maps at the regional scale using yield prediction algorithms. Coinciding with the prediction of average regional yield, the project team evaluated the accuracies of producing yield maps for all crops defined in section 5.8. Using the process defined in section 5.6, the algorithms applicable for each region were applied to each subsetted crop, converting GNDVI values for each pixel into yield (TCH). The yield maps were classified into 8 yield classes via a density slice and distributed to industry as Google Earth files (KMZ). A tutorial detailing how to produce these files using ENVI is provided in the appendices of this report (Appendix 1: Tutorial 9: ENVI: Creating Google Earth KMZ files from Geotiffs). 5.10. Validation of image derived yield maps. The accuracy of prediction at the block level was validated against Mill data following the 2010, 2011 and 2012 harvest. This analysis was conducted by DAFF Qld biometricians and included 8 data sets (Bundaberg 2010, 2011, 2012; Isis 2010; Herbert 2011 and 2012 and SRDC Project DPI021 Final Report_without appendices.doc 21 Burdekin 2011 and 2012). Using the software GenStat, a simple regression was fitted between actual (y variable) and predicted yield (x variable) for each data set. For the initial analysis no data was excluded in spite of there being evidence of data with high leverage and or residuals. Following a t-test, the accuracy of prediction was identified by the closeness of the R- squared value to 100%, the intercept being not significantly different to zero, and the slope not significantly different to 1. As well as the validation of predicted yield at the whole crop level, predictive accuracies were also undertaken at the within the crop level, using point source field samples from two crops as defined in section 5.5. Additional analysis of all data sets in terms of seasonal conditions, growing year, cultivar and class will be undertaken in a follow on SRDC project (Developing remote sensing as an industry wide yield forecasting, nitrogen mapping and research aide). 5.11. Correlating image results against harvester derived yield maps (CSE022). In order to meet objective 3 (Section 4) of this project, classified VI images and image derived yield maps were correlated against those produced from harvester yield monitors (CSE022). At the time of this report, the 2011 yield data collected for the Bundaberg (Hubert6.7 ha) and Burdekin (Pozzebon- 26.8 ha) was available and subsequently analysed against the corresponding imagery for each site, IKONOS imagery captured 12 May 2011 (Burdekin) and 23 March 2011 (Bundaberg). The 4 band imagery was georectified, with pixel data within the boundary of CSE022 each site, extracted. Pixel values were then converted into GNDVI, and converted to text format. To remain consistent with the harvester derived yield maps the GNDVI images were resampled to a 2 metre grid by multiplying the 3.2 metre pixel resolution IKONOS images by a floating point 2 metre raster, with all values set to unity. The resultant 2 m GNDVI images were then smoothed using a moving average on a 5 x 5 array (= 10 m x 10 m) of pixels centred on the pixel of interest, this was achieved using the ‘focal statistics’ function (Spatial Analyst). The yield monitor data were kriged on to the same 2 m grid as the imagery using the software Vesper. The data was reduced to 3 second logging intervals with any points exceeding 3 standard deviations from the mean also removed prior to map interpolation. The data was then adjusted to match mill harvest data on a per- harvest event basis. At the Bundaberg site, the yield map was derived from a roller opening sensor, whilst at the Burdekin site data was provided by a Solinftec yield monitor. Note that the Solinftec yield monitor also relies on sensing of the roller opening, with data adjusted to yield (TCH) using a proprietary algorithm. Clustering of the various map layers was done using k-means clustering in JMP 8 (SAS, Cary, North Carolina, USA). SRDC Project DPI021 Final Report_without appendices.doc 22 6. Results The following section details the results of the project in terms of the 5 objectives. 6.1. Benchmarking and identifying the most feasible and suitable commercial imagery (i.e. spatial resolution, repeat time and economic feasibility) for identifying crop variability and thus directing targeted mid-season management within the Australian cane industry. As listed in Table 2 of this report, the project team obtained imagery from a wide range of active and passive sensors. With all passive multispectral platforms capable of producing VI images, the recommendation of the most suitable came down to cost, the most appropriate minimum purchase area, repeat capture time, appropriate spatial resolution, and overall ability to be easily manipulated in terms of georectification etc. Overall, imagery from the French owned SPOT5 satellite, supplied by Astrium (http://www.astrium-geo.com/), was identified to be the most suitable for a range of applications. A single SPOT5 scene (3600km2) encompassed the majority of cane crops within a particular growing region therefore eliminating the need for additional image processing such as mosaicing and colour balancing. The imagery was shown to be cost effective at $1AUS per km2 and due to its 2-3 day revisit interval, reliable in terms of providing cloud–free scenes. The spectral resolution of SPOT5 (green, red, near and mid infrared) allowed most accepted vegetation indices to be derived. At the block level, the classified GNDVI ‘zonal’ vigour maps produced from the 10 metre spatial resolution images were comparable to those produced by the higher resolution platforms. However, were unable to discern sub metre constraints such as weed infestations and damage resulting from grubs, soldier fly, rat and pig. The pan sharpened IKONOS product (0.8 metre) was shown to be effective in identifying such sub metre constraints. It is acknowledged that a number of commercially available satellites could have supplied similar high quality submetre imagery. However, at the time of this project, IKONOS imagery was available under a three * 50 km2 capture deal that equated to $22AUS per km2 as supplied by Geoimage (http://www.geoimage.com.au/geoimage/) and AAM (http://aamgroup.com/. The RapidEYE (RE) imagery held great promise as a optimal image source for the development of sugarcane applications due to its 5 m spatial resolution, 5 spectral bands (blue, green, red, red-edge and near infrared) and 5 satellites in the constellation meaning a high revisit rate, ideal for cloud prone areas. Unfortunately a number of issues limited the ability to fully assess the suitability of this imagery. Initially the minimum RE scene was $5000km2 scene and at AU$9850 (including the mosaicing of 14 tiles), which was cost prohibitive at the research level when compared to SPOT5. There were also issues in communication between the Australian distributors and German owners of the satellite. I understand that at the completion of this research, the minimum RE scene size has been reduced to 3500km2 and communications with the German owners has improved. As such this may be identified as an option for the future along with new generation platforms such as SPOT6 with 6 metre spatial resolution and Pleiades’. SRDC Project DPI021 Final Report_without appendices.doc 23 In regards to the active sensors, imagery from Terra SAR X and Lidar was obtained by other agencies for a number of the project sites. However, a lack of time to access and then correctly process the data meant that it could not be effectively evaluated. It is envisaged that radar sensors may play a future part in directing in season management due to their ability to provide data during the cloud prone early vegetative growth stage (February to March). Additional research is required to validate this. The ‘Raptor’ active, airborne optical reflectance sensor was considered to be a possible future source of crop vigour data for the cane industry due to its ability to operate under variable cloud cover and at night, and as such was evaluated through this project. 6.1.1. Evaluation of the Raptor sensor. Prior to this project, the Raptor sensor’s capacity to produce accurate measures of sugar cane vigour had not been assessed. As such a number of tests were undertaken to identify the impact that variables such as flying height (Figure 5) and direction (Figure 6) had on the accuracies of derived vigour maps. a bc Figure 5. a. Classified NDVI images identifying the spatial trends of crop vigour measured at three different flying heights (100, 135 and 180 ft AGL). b. correlation between NDVI SRDC Project DPI021 Final Report_without appendices.doc 24 measured at 100 ft AGL and 135 ft AGL. c. correlation between NDVI measured at 100 ft AGL and 180 ft AGL. The collection of NDVI data at the three different flying heights identified little difference between 100 and 135 ft above ground level (AGL), producing an R2 = 0.85, a slope close to 1 (0.995) and an intercept close to 0 (0.02). A lower correlation was identified between 100 ft AGL and 180 ft AGL, producing an R2 = 0.79, slope of 1.15 and intercept of 1.137. This result indicates that there is some flexibility in the flying height at which Raptor imagery is collected although erroneous results may occur if the height exceeds 135 ft AGL. This result is consistent with those identified by the University of New England in the capturing of Raptor imagery where the signal, particularly within the Red band decays beyond a sensorcanopy distance of approximately 60 metres. Given the row structure of cane, it was considered prudent to evaluate the impact of flying direction, namely across or along cane rows, on the derived maps. In Figure 6, it can be seen that similar spatial trends are present in the classified NDVI images from data collected from both flying directions (Figure 6 a), as well as a high correlation (R2 = 0.78), slope 0.98 and intercept of 0.005 produced when comparing the two (Figure 6 b). These results indicate that flying direction had little influence on data integrity, a result most likely attributed to the large footprint of the Raptor sensor and the fact that the sugar cane plants were close to full canopy at the timing of data collection, therefore reducing the visibility of inter row soil. a SRDC Project DPI021 Final Report_without appendices.doc 25 b Figure 6. a. Classified NDVI images of a cane crop derived from the Raptor sensor being flown across and along the cane rows. b. correlation between NDVI derived from data collected from the different flying directions. The Raptor data was initially collected at ~10m (on ground) transect spacing and a desk top study was undertaken to determine if the flight paths could be minimised without altering the spatial pattern of the resultant NDVI map (Figure 7). a. All flight paths Kriged NDVI layer b. Alternate flight paths 0.7 y = 0.9327x + 0.0306 0.6 R2 = 0.9526 Alternate (NDVI) 0.5 0.4 0.3 Kriged NDVI layer 0.2 0.2 0.3 0.4 0.5 All (NDVI) 0.6 correlation between Raptor NDVI all versus alternate transects. SRDC Project DPI021 Final Report_without appendices.doc 26 c. Alternating double. Kriged NDVI layer Double alternate (NDVI) 0.7 y = 0.9015x + 0.0454 0.6 R2 = 0.9263 0.5 0.4 0.3 0.2 0.2 0.3 0.4 0.5 All (NDVI) 0.6 correlation between Raptor NDVI all versus alternating double. 100m transect (NDVI) 0.7 y = 0.6756x + 0.1346 0.6 R2 = 0.7489 0.5 0.4 0.3 0.2 0.2 0.3 0.4 0.5 All (NDVI) 0.6 d. 100m transects Kriged NDVI layer correlation between Raptor NDVI all versus 100m transects. Figure 7. Assessing optimal frequency of flight paths for creating a representative NDVI layer from Raptor data. As seen in Figure 7, similar zonal paddock trends (kriged NDVI layers) were obtained from a range of flight transect configurations. Figure 7 b identifies very little variation between the original ~10 m transects to one using every second transect i.e. ~ 20 m producing a strong R2 = 0.96, slope of 0.93 and intercept of 0.034. A similar result was produced in Figure 7 c, using alternating transects of two retained and two removed (R2 = 0.93, slope of 0.90 and intercept of 0.05). Even with transects reduced to every ~100 m (Figure 7 d), the correlation remains relatively strong (R2 = 0.75). However, the slope (0.68) and intercept (0.134) indicate a greater separation from original ~ 10 m model. This result indicates that the ~ 10m flown transects may be excessive and that data collection may only be required at relatively wide transect intervals (i.e. 50 m) thereby saving flight time and associated costs. The temporal comparison of Raptor data captured across each of the three test crops identified a number of erroneous points that were believed to be associated with non cane spectral information or an ‘edge’ effect. With the on-ground footprint of Raptor being 7.5 m * 4.2 m (at 100 ft AGL) it was decided that an internal 15 m buffer be applied. The implementation of the buffer not only removed the spurious points (Figures 8), but for Block 3 (CSE022 Hubert site) resulted in a greater separation of data representing different cultivars with in the sub blocks. This clear segregation of data supports the potential future application of remote sensing for the rapid screening of varieties for plant breeding rights (PBR) auditing, if required. SRDC Project DPI021 Final Report_without appendices.doc 27 R2 = 48.1 R2 = 80 R2 = 74.2 a R2 = 80 R2 = 38.3 b R2 = 35.6 c Figure 8. Correlations between Raptor data captured on two occasions (February and March) over the three test crops, Block a (Bullseye), Block b (Hubert) and Block c (Relmay), before and after the implementation of an internal 15 m buffer (H-Buff). To validate the accuracy of Raptor derived NDVI maps, imagery captured over three Bundaberg crops on three occasions (23 Feb, 23 March and 2 May 2011) was compared against IKONOS NDVI images captured on two occasions (23 March 2011 and 30 March 2011) (Figure 9) as well as a SPOT5 image captured on the 27 March 2011 (Table 5). SRDC Project DPI021 Final Report_without appendices.doc 28 Differing cultivars a. IKONOS false colour (23 Mar. 11) NDVI (IKONOS) Raptor NDVI (23 Mar. 11) b. IKONOS false colour (30 Apr. 11) NDVI (IKONOS) Raptor NDVI (2 May 11) Differing cultivars c. IKONOS false colour (30 Apr. 11) NDVI (IKONOS) Raptor NDVI (2 May 11) Figure 9. IKONOS False colour and NDVI images of the three Bundaberg crops Relmay (a), Bullseye (b) and Hubert (c). Also NDVI images derived for from the Raptor sensor. The spatial NDVI patterns displayed in the IKONOS and Raptor images were visually comparable in Figure 9 a, b and c, with different cultivars grown within the sub blocks evident in Figure 9 c. The NDVI images derived from each platform were also statistically compared, following all data sets being interpolated to 10 metres to reduce the volume of data (Figure 10). SRDC Project DPI021 Final Report_without appendices.doc 29 a b SRDC Project DPI021 Final Report_without appendices.doc 30 c Figure 10. Correlation matrix ‘scatter plots’ comparing NDVI layers derived from Raptor and IKONOS imagery, collected on a number of occasions over three sites: Relmay (a), Bullseye (b) and Hubert (c). From the correlation scatter plots the NDVI layers derived for the Bullseye crop (Figure 10 b) produced the highest R2 values (0.71 to 0.88). The obvious separation in data points for the Hubert crop (Figure 10 c) was attributed to different varieties with in the sub blocks. The Relmay crop (Figure 10 a) also exhibited data separation as a result of multiple varieties but a slight malfunction with the Raptor sensor on the 23 February 2011 further confounded the correlations. To further validate the consistency between the sensors and capture dates, NDVI values sampled from 12 specific locations in the Relmay crop (Figure 11 a) and 15 for the Bullseye crop (Figure 11 b) were compared against corresponding values extracted from a SPOT5 image (27 March 2011) and to measured yield and CCS (Table 5). SRDC Project DPI021 Final Report_without appendices.doc 31 a. b. Figure 11. False colour IKONOS images of two Bundaberg crops with field sampling locations highlighted. 12 sample sites for Relmay (a) and 15 for the Bullseye crop (b). Table 5. Correlation matrix comparing NDVI values derived from imagery captured by the Raptor sensor, SPOT5 and IKONOS to yield and CCS for specific locations within two Bundaberg crops. Block 1 Raptor 23.02.11 Raptor 23.03.11 IKONOS 23.03.11 SPOT 27.03.11 IKONOS 30.04.11 Raptor 02.05.11 TCH CCS Raptor 23.02.11 1 Raptor 23.03.11 0.8738 1 IKONOS 23.03.11 0.5380 0.7588 1 SPOT 27.03.11 0.5498 0.8292 0.9242 1 IKONOS Raptor 30.04.11 02.05.11 0.4024 0.3436 0.6264 0.5913 0.9620 0.8501 0.8875 0.8645 1 0.9082 1 TCH CCS 0.6291 -0.5211 0.7845 -0.3755 0.8408 -0.4565 0.8460 -0.2630 0.7510 -0.4206 0.7474 -0.3440 1 -0.5614 1 Block 2 Raptor 23.02.11 Raptor 23.03.11 IKONOS 23.03.11 SPOT 27.03.11 IKONOS 30.04.11 Raptor 02.05.11 TCH CCS Raptor 23.02.11 1 Raptor 23.03.11 0.9845 1 IKONOS 23.03.11 0.9813 0.9826 1 SPOT 27.03.11 0.9716 0.9787 0.9828 1 IKONOS Raptor 30.04.11 02.05.11 0.9774 0.9457 0.9881 0.9816 0.9948 0.9545 0.9850 0.9564 1 0.9763 1 TCH CCS 0.9286 -0.7848 0.9429 -0.7628 0.9050 -0.7024 0.8927 -0.7286 0.9114 -0.7226 0.9217 -0.6687 1 -0.7522 1 The high correlations (R2) identified in Table 5, particularly for Block 2 (Bullseye), indicate that at specific sample locations the Raptor is producing a consistent spatial trend in NDVI values to those produced by the satellite platforms. The only exception being the lower R2 values identified from the 1st Raptor capture to that of the other platforms in Block 1 (Relmay). This indicates a systematic issue with the sensor on that day, with the Raptor performance observed to be degraded owing to suspected moisture infiltration into the sensor head. The high correlations produced between the Raptor data and yield (TCH) for both sites supports the possibility that yield maps could be derived from the Raptor sensor. Interestingly, for Block 2 (Bullseye) consistent negative correlations between NDVI derived from each sensor to CCS was also identified. This may indicate that CCS maps may also be derived from imagery VI values, although further research would be required. SRDC Project DPI021 Final Report_without appendices.doc 32 For the CSE022 Hubert site (Block 3) additional k-means clustering analysis was undertaken to compare the consistency of spatial trends in NDVI derived from both sensor types, across the capture dates. This involved the classification of the NDVI layers into 3 and 5 classes (Figure 12). The letters after the cluster rankings indicate whether the means are significantly different. There is no test of significance for the IKONOS data as it was not kriged and to undertake this on the raw data the large number of points would results in everything appearing significant. As with the visual comparisons (Figure 9), the trends identified from both sensors were not totally consistent for any given acquisition date, with the cluster analysis, again identifying the February Raptor image as differing from the other two collection dates. a. SRDC Project DPI021 Final Report_without appendices.doc 33 b. Figure 12. Comparison between classified NDVI layers derived from kriged Raptor data captured on three occasions and IKONOS imagery captured on two occasions. a. demonstrates the analysis of 5 zones, whilst b. includes 3 zones. Legends have been categorised into 20th percentiles. For the NDVI images, 1 indicates a higher value. From these results it can be seen that the Raptor sensor is providing NDVI maps consistent with those produced by IKONOS, and that there is little difference in NDVI maps derived from imagery captured between February and May. 6.2. Identifying the optimum time of image capture that will accurately depict mid- season crop variability whilst avoiding seasonal times most prone to cloud cover, across key Australian cane farming regions. For remote sensing to be successfully implemented as a tool for guiding management decisions, imagery has to be available at the appropriate phenological growth stage where the crop can respond to varied management and the management itself can be applied without crop damage. As such, this project attempted to identify that critical window for Australian sugar cane by attempting imagery capture throughout most of the growing season over the Bundaberg, Herbert and Burdekin regions. It was hypothesised that imagery captured early in the growing season i.e. January to March would not only identify variability in crop growth at the required vegetative growth stage, but provide it at a time where applications could still be applied non- destructively. Unfortunately the ability to capture imagery during this time frame was near impossible due to continual cloud cover, a result supported by previous research (Johnson and Kinsley-Henderson, 1997). With successful captures only occurring from March, the ability to implement alternative management strategies based on the imagery within that same season, was negated. This SRDC Project DPI021 Final Report_without appendices.doc 34 result indicated either the need for an active sensor such as Radar or the ‘Raptor’ that could provide earlier in the season i.e. January to March, or that alternate management decisions be applied post harvest for the benefit of the following ratoon crop. 6.2.1. Using spatial technologies to identify growth variability in cane crops, likely constraints to production and suggested remedial action. To demonstrate the capacity of remote sensing as an effective tool for identifying growth variability within a cane crop, 7 individual crops were imaged and intensively samples (refer to methodology section 5.5, Table 4). These included: Herbert site (H2): variety Q200 planted over 8.2 ha on 7 July 2008; plant-cane harvested on 21 August 2009.1.83 m spacing. a. b. Label CCS Purity% TCH c. R1 17.9 91.7 56.5 R2 18.9 95.3 46.9 R3 17.3 93.1 54.6 R4 17.6 93.7 56.4 G1 18.3 95.4 83.1 G2 18.4 94.5 73.2 G3 18.3 93.8 60.2 G4 18.0 97.6 68.3 B1 17.1 94.3 108.2 B2 17.5 94.4 110.8 Figure 13. a. False colour IKONOS image of H2 crop (2 August 2009). b. NDVI image of H2 site derived from IKONOS image, with field sampling sites indicated. c. Classified NDVI of H2 site. Table: field sampling results. From Figure 13, IKONOS imagery of the H2 site clearly identified a large degree of spatial variation of cane vigour, with regions of reduced vigour at the northern end of the crop and within a band extending across the crop at the southern end (red colour in Figure 13 c). Field samples taken at strategic locations within the crop (Figure 13 b) identified the variability in NDVI coincided with similar variations in cane production (TCH) but not to CCS (Figure 13 d.). To identify the likely driver of this reduced vigour the SRDC (BPS001) project team compared the spatial variation to that produced by a soil survey (EM- Veris) undertaken over the block in 2008 (Figure 14). SRDC Project DPI021 Final Report_without appendices.doc 35 Figure 14. Shallow and deep EM – Veris maps of the H2 site taken in 2008. As seen in Figure 14, similar spatial trends exist between both the shallow and deep EM maps with the NDVI image (Figure 14b), although low NDVI values occurred in regions of both high EM (red circles) and low EM values (blue circle) measured at both depths. Statistical analysis of these data supported this observation identifying poor correlations between NDVI and shallow electrical conductivity (EC) readings (R2 = 0.45) and NDVI with deep EC readings (R2 = 0.08). This result indicates that the spatial variability in crop performance was unlikely to be the result of factors driving soil conductivity alone. Further evaluation of the crop suggested that topography and soil type may have influenced hill height and ultimately plant establishment, particularly for a seam of heavy clay extending through at the southern end of the crop (blue circle). Further analysis and interpretation of this trial is provided in Coventry et al. (2010). Burdekin site (A. Mann): variety KQ228; Plant cane; 11.5 ha; cane harvested on 6 July 2010. a. b. SRDC Project DPI021 Final Report_without appendices.doc 36 Label CCS Purity% TCH c. R1 15.0 90.6 97.0 R2 14.3 90.8 96.4 R3 15.5 93.3 103.3 R4 16.3 92.7 82.9 G1 15.9 92.4 165.2 G2 16.7 93.9 172.8 G3 15.7 92.9 129.2 G4 * * 143.7 B1 14.5 91.0 126.7 B2 11.8 88.2 174.1 B3 14.0 88.3 177.1 B4 15.3 93.4 191.6 Figure 15. a. False colour IKONOS image of A.Mann crop (28 May 2010). b. NDVI image derived from IKONOS image with field sampling sites indicated. c. Classified NDVI image. Table: field sampling results. A high resolution IKONOS image of the Burdekin (A. Mann) crop identified a large zone of reduced NDVI (red region) at the northern end (Figure 15 c). Field sampling at strategic locations identified this low region to have a substantially lower average yield (95 TCH) than that measured at high vigour zones (167.4 TCH), and a slightly higher average CCS (Figure 15 d). Soil cores taken at same locations as the plant samples, indicated that the reduced production was the result of sandy soils with low water and nutrient holding capacity. At a depth of 40- 60 cm the poor zones (red) exhibited a low average CEC (6.37 meq/ 100g) compared to 21.5 meq/ 100g measured in the blue zones, as well as lower exchangeable nutrients (incl K) and trace elements. In an attempt to mitigate this issue mid – season, the grower was encouraged to irrigate with less water but at a greater frequency in an attempt to stop nutrient leaching. Following the harvest the grower applied 15t (1.25 t/ha) of activated silica (click icon- ActivatedSilica.pdf ), 160 kg/ ha of Nitrogen and 43 kg/ ha of Sulphur. The fertilizer was coated with Aqua Boost AG100, a polycrylamide granule, in an attempt to improve moisture retention. As seen in Figure 16, these post harvest applications resulted in little visible response to the 2011 ratoon crop. a. b. Figure 16. a. False colour IKONOS (28 May 2010) image of the A. Mann crop with sample locations and main region of reduced production identified. b. Repeat false colour image captured in 12 May 2011. SRDC Project DPI021 Final Report_without appendices.doc 37 Following this lack of response it was suggested that a deep (>20cm) application of clay, road base or mill mud be applied to the sandy areas in an attempt to increase its water and nutrient holding capacity. However, as this is a highly destructive form of remedial action it would have to occur at the end of the ratoon rotations and prior to replant. Unfortunately, the impact of this remedial action could not be assessed as the grower passed away during the 2011 season. Bundaberg site (Bullseye A9): Variety KQ200; Plant cane; Area 4.47 ha; harvested 3 Nov 2010; Row spacing 1.52 m. The block was land formed in 2008, with 2010 being the first re-plant. Predominantly bare soil- soil samples indicated heavy clay. a. b. High IR reflective – High growth region. c. Label R1 R2 R3 CCS 15.9 13.1 14.8 Purity% 93.2 87.9 93.1 TCH 23 4 14 R4 15.7 94.3 18 G1 16.4 94.0 G2 15.9 95.2 G3 14.5 92.7 G4 15.5 92.8 B1 16.8 94.3 49 43 67 47 82 B2 15.5 93.8 87 B3 15.8 93.1 98 Figure 17. a. False colour IKONOS image of Bullseye A9 crop (14 May 2010). b. NDVI image derived from IKONOS image with field sampling sites indicated. c. Classified NDVI image. Table: field sampling results. SRDC Project DPI021 Final Report_without appendices.doc 38 R2 R4 B1 G2 Figure 18. Photographs of cane growing at a number of the sampling locations taken prior to harvest. The Bundaberg A9 crop displayed a large degree of spatial variability in terms of NDVI, with the low vigour region (red) dominating the north eastern corner (Figure 17c). This region was identified to have extremely poor cane with an average plant density of 10 stalks per linear metre compared to 17 measured in the blue zones (Figure 18). The measured average yield of the low NDVI zones was 15 TCH compared to 89 TCH in the blue regions, with the later also providing a slightly higher CCS (Figure 17). Soil samples taken at depths of 0-20cm and 40- 60cm, identified low organic carbon, high salinity and sodicity to be the likely constraint to productivity. Using this information, the grower discontinued the second cane ratoon, as indicated by the bare soil in the 2011 image (Figure 19 b) and undertook remedial action on the soils. This included re-lasering, the application of 3 – 6 t/ha of gypsum based on the variability map, and a blanket application of 6 t/ha of chook manure. To further increase organic matter, a short fallow oats crop was grown and then ploughed in. For the 2011/ 2012 season, the block was re-planted with cane with only a slight improvement in growth, as inferred by an image captured on the 6 April 2012, identified along the north western edge of the crop (black circle in Figure 19 c). SRDC Project DPI021 Final Report_without appendices.doc 39 a. b. c. Figure 19. False colour images of Bullseye A9 block captured 14 May 2010, 23 March 2011 and 6 April 2012. The extremely low yield measured at the R2 site (4 TCH) (yellow circle Figure 19) was believed to be the result of a saline soak with soil samples at 40 -60cm exhibiting high levels of chloride (663 mg/Kg) and exchangeable sodium (2.63 meq/110g). The subsequent increase in the poor production area surrounding this point in 2012 (Figure 19c) is believed to be the result of increased rainfall from 2010 raising the groundwater table and possibly increasing the hydraulic pressure from the dam located to the north west of the crop. It has been suggested that the grower install table drains and investigate raised beds as methods to improve drainage off the crop. Bundaberg site (Relmay 3A): Variety Q208; Plant cane; Area 15.43 ha; harvested 25 October 2010; Row spacing 1.52 m. a. b. SRDC Project DPI021 Final Report_without appendices.doc 40 c. Label R1 R2 R3 R4 R5 G1 G2 G3 G4 B1 B2 B3 B4 CCS 14.3 14.7 15.6 12.4 15.9 15.1 15.7 15.5 15.5 15.0 15.5 15.0 14.0 Purity% 92.9 94.4 95.1 91.4 95.4 94.6 95.7 95.4 95.6 94.8 94.3 94.3 92.8 TCH 35 53 49 25 46 76 56 64 99 84 112 88 71 Figure 20. a. False colour SPOT5 image of Relmay 3A crop (10 May2010). b. NDVI image derived from SPOT5. c. Classified NDVI image with field sampling sites indicated. Table: field sampling results. Imagery of the Relmay 3A block (Figure 20) clearly identified a distinct segregation of crop vigour that was supported by similar differences in yield. Low NDVI red regions produced on average 41.6 TCH compared to 88.8 TCH in the blue high NDVI zones. There was little difference in the average CCS measured from each NDVI class, with the exception of a low value at R4 (12.4). Soil samples collected across the block identified a higher gravel content at the low NDVI regions R1 and R3, the result of prior remedial land forming where gravel was applied on sandy areas in an attempt to improve the water and nutrient holding capacity of the soil. Unfortunately as this was applied to the soil surface rather than at depth, this limited production rather than improved it. Sulphur was also identified to be at very low levels at 40- 60cm in the low NDVI regions (1.3 mg/kg) compared to (8.4 mg/kg) in the high NDVI regions. It was suggested that a test strip of higher sulphur be applied to determine if that alone would prompt a response, this unfortunately did not occur. This block provides a good example of how imagery acquired over a number of cropping seasons can provide growers with an understanding of the inherent spatial variability within their blocks. If the spatial orientation of both high and low crop regions remains unchanged across seasons and crop age (i.e. 2005, 2008 and 2010) (Figure 21) then well informed management decisions can be made prior to planting, including the use of variable rate technologies (VRT) or more suitable cultivars. If the zones are unstable from season to season (i.e. 2005 to 2007) then the impacts of climate, management or rotational effects should be considered and managed appropriately. SRDC Project DPI021 Final Report_without appendices.doc 41 a b, c d Figure 21. Classified NDVI images of cane grown within the Relmay 3A block during a. 2005 (Q188 2nd Ratoon); b. 2007 (Q205 April plant); c. 2008 (Q205 1st Ratoon) and d. 2010 (Q208 Spring replant). Bundaberg site (Relmay 45): Variety Q183; Spring fallow plant; Area 19.5 ha; Harvested 27 September 2011; Row spacing 1.52 m. The clustering of sample points (Figure 22 a) and cropped classified image (Figure 22 c) were the result of the northern end of the block being a different cultivar and the southern end being partially harvested prior to sample collection. a. b. c. Label R1 R2 R3 R4 R5 G1 G2 G3 B1 B2 B3 B4 CCS 15.3 16.1 15.7 15.1 15.9 15.4 14.6 15.6 15.6 14.8 15.1 16.1 Purity% 96.3 97.0 96.9 96.8 96.5 96.3 96.1 96.2 96.3 95.6 95.8 96.5 TCH 72 69 75 100 66 89 117 122 110 131 128 108 Figure 22. a. False colour IKONOS image of Relmay 45 crop (30 April 2011). b. NDVI image derived from IKONOS. c. Classified NDVI image with field sampling sites indicated. Table: field sampling results. SRDC Project DPI021 Final Report_without appendices.doc 42 As with the previous examples, the Relmay cv. Q183 crop displayed a large degree of NDVI variability that was supported by differences in average yield measured in the low NDVI regions (77 TCH) to that in the high NDVI regions (120 TCH). There was little difference in average CCS between the high (15.4) and low (15.6) regions. Soil samples collected at 0-20 cm and 40-60 cm indicated critically low levels of exchangeable magnesium at depth may be limiting yield (0.16 meq/100g: 0.2 is the critical level), and that low lying crop regions may have experienced water logging following the high rainfall in January 2011. The low reflectance location R1 was attributed to extensive rat damage (Figure 23). Figure 25. Cane stalks exhibiting extensive rat damage. Bundaberg site (Bullseye C2): Variety KQ228; Spring fallow plant; Area 18.7 ha; Harvested 25 July 2011; Row spacing 1.5 m. a. b. SRDC Project DPI021 Final Report_without appendices.doc 43 Label CCS Purity% TCH R1 15.6 96.0 43 R2 15.2 95.0 51 R3 12.2 81.9 42 R4 16.0 96.4 61 R5 * * 57 G1 15.7 95.5 86 G2 14.5 95.9 67 G3 15.3 96.5 66 G5 14.9 95.0 70 G6 15.6 96.3 77 B1 14.0 93.7 110 B2 13.7 94.1 98 B3 13.8 95.3 99 c. B4 14.1 94.9 96 B5 14.3 96.3 108 Figure 24. a. False colour IKONOS image of Bullseye C2 crop (30 April 2011). b. NDVI image derived from IKONOS. c. Classified NDVI image with field sampling sites indicated. Table: field sampling results. Imagery of the Bullseye C2 crop (cv. KQ228) identified a large region of low NDVI extending from the southern end to the north- western corner (Figure 24c). Field sampling undertaken on the 25 July 2011, identified a 50% yield deficit between samples collected in the high NDVI (102 TCH) to that from the low NDVI regions (51 TCH). A slightly lower average CCS was measured in the high NDVI regions (14.0) compared to the low (14.7). Soil samples identified moderately sodic soils and poor drainage to be the likely drivers of reduced production, with a low ESP (0.79: 0-20cm and 3.44: 40- 60cm) measured at the high growth areas compared to 4.58 (0-20cm) and 12.92 (40-60cm) in the poor growth areas. This moderate subsoil level can prevent water infiltration, reduce oxygen availability and cause root death. It can be assumed that the poor growth areas suffered severe stunting from excessive rainfall early in the 2011 season. Both chloride and sodium levels were also higher in the poorer performing crop regions. Suggested remedial action included the application of gypsum / organic matter to increase drainage and the re-lasering of beds following the cane rotation. Burdekin site (Pozzebon): Variety Q208; 1st ratoon cane; Area 12.4 ha; Harvested 26 November 2011; Row spacing 1.55 m. a. b. SRDC Project DPI021 Final Report_without appendices.doc 44 c. Label CCS Purity% TCH R1 15.3 86.9 29 R2 8.5 73.4 46 R3 13.8 85.4 30 R4 12.8 82.3 12 Y1 16.4 90.3 106 Y2 11.0 79.1 71 G1 14.5 84.8 103 G2 14.5 85.9 116 C1 16.2 89.6 134 C2 13.4 86.0 134 B1 13.2 82.4 85 B2 14.8 86.2 125 B3 13.9 88.3 123 B4 14.6 87.5 139 Figure 25. a. False colour IKONOS image of Burdekin Pozzebon crop (12 May 2011). b. NDVI image derived from IKONOS. c. Classified NDVI image with field sampling sites indicated. Table: field sampling results. The Pozzebon crop, again displayed a large region of low NDVI (Figure 25 a) that when sampled yielded on average 29 TCH compared to 118 TCH from the high NDVI zones. Soil samples taken at a number of the locations indicated that the low performing regions may be attributed to saline and/ or sodic soils with low levels of potassium and phosphorus. The agronomic group Farmacist (http://www.farmacist.com.au/) used the classified NDVI image (Figure 25 c) to develop a variable rate application of the liqiud Nitrogen fertilser Dundah, as well as a test strip (red strip) of 0 Nitrogen (Figure 26). A blanket application of 10- 15 kg/ ha of Sulphur was also applied. These applications were used in an attempt to verify the yield potential of cane in response to Nitrogen i.e. higher applications of N were applied to high NDVI zones, low rates applied to low NDVI zones. Figure 26. Variable rate fertiliser application derived from image information SRDC Project DPI021 Final Report_without appendices.doc 45 a. b. 0 N strip Figure 27. False colour IKONOS images captured 12 May 2011 (a) and 11 April 2012 (b), before and after the variable rate application of Dundah and the 0 N strip. As seen in Figure 27 b, the variable N application produced little visual change in the 2012 crop, with only a minimal reduction in cane vigour observed within the 0 nitrogen strip. This result indicates that the spatial variability of the crop was not primarily driven by N deficiency but rather other constraints. Although this example did not definitively identify the driver of reduced production, it does demonstrate how imagery cane be used to derive variable rate applications, and then subsequently used to monitor crop responses. 6.3. Assess the utility of this imagery for explaining the yield variability measured through the CSE022 “A coordinated approach to Precision Agriculture RDE for the Australian Sugar Industry’ project. Accurate in-season predictions of regional yield are of vital importance for formulating harvesting, milling and forward selling decisions, whilst at a block scale, they provide growers with an understanding of both in-crop variability and total production. Currently, annual cane production estimates are made by visual yield assessments. Although this method can produce accuracies of up to 95% (Pitt pers. comm. 2011) it can be influenced by variable climatic conditions such as those experienced in 2010. As such, geographic information systems (GIS) and remote sensing (RS) may offer an additional tool for validating these predictions as well as potentially provide a more accurate seasonally sensitive method of prediction. The results presented within this section support this hypothesis. As mentioned in section 5.6, a number of vegetation indices were investigated to identify that which produced the highest consistent correlation with yield (TCH), with GNDVI identified to the best suited. This result is consistent with other research that has identified SRDC Project DPI021 Final Report_without appendices.doc 46 the Green visible band to be sensitive to chlorophyll content, but yet, less likely to saturate under high LAI (Gitelson et al 2002). Absorbance by the red spectral band has been identified to saturate at an LAI greater than 3, whilst research conducted by Wang et al (2007) identified the Green band to be sensitive to LAI beyond 5 or 6. 6.3.1. CSEO22 sites. The following two examples were crops extensively sampled by SRDC Project CSE022, and as such only a brief overview of findings for the 2011 season are provided. Burdekin site (CSE022 Pozzebon): Area 26.8 ha; Row spacing 1.55 m. a. b. Multiple varieties Figure 28. False colour IKONOS images captured 12 May 2011 (a) and derived GNDVI image (b) As seen in Figure 28 a, the CSE022 Burdekin site included multiple varieties including Q183, Q208 and Tellus as well as a number of classes ranging from 3rd ratoon through to 6th ratoon. This variability within sub blocks resulted in an extended harvest period over 6 weeks. The GNDVI image (Figure 28 b) exhibited a large region of reduced vigour along the western edge, a trend that was also apparent in the yield map derived from the yield monitor and the high resolution EM38 and VERIS electromagnetic soil surveys (Bramley et al., 2012) conducted after the 2011 harvest (Figure 29). SRDC Project DPI021 Final Report_without appendices.doc 47 Figure 29. Comparisons of a classified GNDVI layer from an IKONOS image captured 12 May 2011; elevation and electro-conductivity (ECa) maps as well as a 3 zone classified map derived from the cluster analysis of the three layers. The three zone layer derived from the cluster analysis (Figure 29) clearly demonstrates the consistency of spatial trends indentified by the ECa, GNDVI and yield layers. This result indicates that soil variation is the likely driver of crop variability, with topography, particularly poor drainage within low lying sodic areas contributing to reduced production. In order to determine whether imagery could accurately explain yield variability, a direct correlation was undertaken between imagery GNDVI values and corresponding yields (TCH) for 50 specific locations extracted from the crop (Figure 30). The linear algorithm produced from this correlation was then used to convert each GNDVI pixel value into yield, allowing a surrogate yield map to be produced (Figure 31). 300 y = 1064x - 290.97 R2 = 0.594 250 Cane Yield (TCH) 200 150 100 50 0.35 0.4 0.45 0.5 GNDVI 0.55 Figure 30. Correlation between yield (TCH) and GNDVI values from an IKONOS image (12 May 2011) extracted from 50 locations within the Pozzebon crop. SRDC Project DPI021 Final Report_without appendices.doc 48 Figure 31. Classified yield maps produced from the Solfintec harvest monitor (left image) and derived from an IKONOS GNDVI image (middle image), with a difference map comparing the two provided on the right. In order to compare the image-based yield map to that produced from the Solfintec yield monitor, a difference map was produced by subtracting one map from the other (Figure 31). The difference map identified much error between the predicted and actual yield values with large regions of over (blue) and under (red) prediction. In some case this prediction error was found to 62.5 TCH. The yield layer derived from the Solfintec yield monitor was identified to have a confidence interval of ~55 TCH, which would have contributed to this large error. A correlation matrix undertaken on the same 50 points did provide a more encouraging comparison of the two derived yield maps, producing and R2 of 0.59 (Figure 32). SRDC Project DPI021 Final Report_without appendices.doc 49 Figure 32. Correlation matrix developed between mill adjusted yield values from the Solfintec monitor to those derived from a GNDVI image. Bundaberg site (CSE022 Hubert): Variety Q232; 1st ratoon cane; Area 6.7 ha. a. b. Figure 33. False colour IKONOS image of Hubert site captured 23 March 2011 (a) and derived GNDVI image (b). The false colour and subsequent GNDVI image of the Hubert (CSE022) site for the 2011 season identified reduced crop vigour at the south- western end of the crop (Figure 33). This region of low vigour was again evident in the following April image capture, as well as within a yield map provided by from a roller opening yield sensor (Figure 34). SRDC Project DPI021 Final Report_without appendices.doc 50 Figure 34. Comparisons of a number of spatial layers including classified GNDVI maps from IKONOS imagery captured 23 March and 30 April 2011; a classified yield map derived from CSE022 yield monitor and result of cluster analysis A cluster analysis confirmed the consistency between the two GNDVI images (top right map in Figure 34), whilst the additional cluster analysis with yield monitor derived yield map identified a similar spatial pattern (bottom left map in Figure 34). For both image/yield cluster analyses, the mean zone yields were found to be significantly different with the 95% confidence interval for the yield map being ~ 11 t/ha. The results indicated that a potential yield difference of ~15 t/ha existed between the two sub-blocks. A correlation matrix undertaken between the two imagery derived yield maps and the monitor derived map achieved higher r values than that from the Pozzebon crop (Figure 35), a result most likely attributed to higher confidence interval ~12 t/ha of the ‘roller opener’ yield monitor. SRDC Project DPI021 Final Report_without appendices.doc 51 Figure 35. Correlation matrix between the harvest monitor derived yield map and the yield maps from two GNDVI IKONOS images acquired 23 March and 30 April 2011. From the similarities observed between the classified zonal maps, it can be said that the image derived yield maps can be useful in characterising patterns of spatial variation in yield, and could therefore be a viable delineation tool for management zones. However, whether the imagery is able to predict yield with sufficient accuracy at the sub-block level requires further research and validation. 6.3.2. Additional yield validation sites. The following crops sampled for yield, coincide with those presented in section 6.2.1. Herbert H2 IKONOS 2 August 2009. 120 y = 406.96x - 42.28 R2 = 0.76 100 Cane Yield (TCH) 80 60 40 20 0.2 0.25 0.3 0.35 GNDVI 0.4 Predicted average yield= 84.7 TCH act. Ave yld = 72.4 TCH SRDC Project DPI021 Final Report_without appendices.doc 52 Mann IKONOS 28 May 2010 210 y = 22.78e4.18x 190 R2 = 0.76 170 C ane Y ield (TC H ) 150 130 110 90 70 50 0.30 0.35 0.40 0.45 GNDVI 0.50 Predicted average yield= 147.2 TCH act. Ave yld = 132.5 TCH C ane Yield (TC H ) Bullseye A9 IKONOS 14 May 2010 100 y = 240.35x - 13.24 80 R2 = 0.94 60 40 20 0 0.00 0.10 0.20 0.30 0.40 0.50 GNDVI Predicted average Yield= 51.3 TCH Actual Ave yield was 40.1 TCH Cane Yield (TCH) Relmay 3A SPOT5 14 May 2010 160 y = 399.14x - 148.61 140 R2 = 0.52 120 100 80 60 40 20 0 0.45 0.5 0.55 0.6 0.65 0.7 GNDVI Predicted average Yield= 85.7 TCH Actual Ave yield was 93.4 TCH SRDC Project DPI021 Final Report_without appendices.doc 53 Cane Yield (TCH) Relmay 45 IKONOS 23 March 2011 140 y = 882.02x - 410.01 R2 = 0.72 120 100 80 60 40 0.52 0.54 0.56 0.58 0.60 0.62 GNDVI Predicted average Yield= 98.3 TCH Actual Ave yield was 100 TCH Bullseye C2 IKONOS 30 April 2011 120 y = 330.18x - 91.97 100 R2 = 0.85 Cane Yield (TCH ) 80 60 40 20 0.40 0.45 0.50 0.55 GNDVI 0.60 Predicted average Yield= 80.4 TCH Actual Ave yield was 88.7 TCH Pozzebon 13 IKONOS 12 May2011 160 y = 827.65x - 314.59 140 R2 = 0.80 120 C ane Yield (TC H ) 100 80 60 40 20 0 0.40 0.45 0.50 GNDVI 0.55 Predicted average Yield= 93.2 TCH Actual Ave yield was 88.6 TCH Figure 36. Correlation between GNDVI and crop yield (TCH) for 7 crops hand sampled during 2009, 2010 and 2011. Derivation of yield maps from each respective correlation algorithm. SRDC Project DPI021 Final Report_without appendices.doc 54 The examples provided in Figure 36, demonstrate how surrogate yield maps can be derived from imagery and coordinated field sampling. At block level, predictions of average yield derived from the average GNDVI values ranged from 28% over prediction (Bullseye A9) to 9.4% under (Bullseye C2) (Figure 36). These inaccuracies may have resulted from the 5 metre linear sampling area not providing a true representative measure of actual yield found within each of the zones, or alternatively the result of errors with the consigned mill data. In an attempt to evaluate the accuracy of the yield prediction algorithms at the within crop level, an estimate of yield was recalculated from the GNDVI value for each sampled location, using the respective crop algorithm. As seen in Figure 37, a slight under prediction of higher yield values and over prediction of low values occurred. Figure 37. Comparison of actual cane yield (TCH) measured at each sample location (7 crops in Figure 36) to predicted yields derived from each respective crop algorithm. 6.3.3. Derivation of a generic algorithm. The previous section identified imagery to be a potentially useful tool for generating surrogate yield maps. However, the need for labour and time intensive field measurements to calibrate the imagery was considered to be a major restriction to future commercial adoption. In an attempt to remove field sampling, the project team developed and evaluated a ‘generic’ non cultivar, non class specific algorithm. The preliminary algorithm was derived from the average GNDVI values of 112 blocks (SPOT5 image captured 10 May 2010) with corresponding average 2010 cane yields (R2= 0.61). This included multiple varieties and crop classes but excluded stand over crops. The predictive ability of this algorithm was evaluated over 600 ha of cane (39 crops) grown within the Bundaberg region during the 2008 season. Using a SPOT5 image captured 31 March 2008, a predicted average yield of 66.5 TCH was achieved, 3.8% under the average actual delivered yield (69 TCH). This close estimation was SRDC Project DPI021 Final Report_without appendices.doc 55 encouraging and resulted in a new algorithm being developed from both the 2008 and 2010 data (Figure 38). Although the inclusion of the 39 2008 data points did not increase the correlation coefficient of the overall equation, it did provide another season of data. It was hoped that this temporal addition would increase the algorithms ability to compensate for seasonal variability. 250 2008/ 2010 data Lower 5% Prediction Interval 200 Upper 95% Prediction Interval 150 100 y = 3.15e5.70x R2 = 0.59 TCH 50 0 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 GNDVI Figure 38. Correlation between GNDVI and Yield (TCH) measured during the 2008 and 2010 growing seasons (n=151). 95% prediction intervals were calculated to show error on individual predictions. The resultant algorithm (Equation 1) was evaluated over a number of point source locations within sampled crops as well as for the prediction of average yield for the Bundaberg, Isis, Burdekin and Herbert growing regions. Equation 1: y = 3.1528 * е (5.6973 * x) Where y = predicted average yield (TCH) x = average GNDVI value extracted from TOA SPOT5 image. 6.3.4. Validation of the generic algorithm at the regional level. For remote sensing to be considered as a useful and adoptable tool for making predictions of regional yield it has to achieve accuracies as good as or better than those achieved from the current visual assessments i.e. 95%. In order to test this; the generic algorithm was evaluated over the three main growing regions, Bundaberg/ Isis (Figure 39), Herbert (Figure 41) and Burdekin (Figure 43) during the 2010, 2011 and 2012 growing seasons. Additional regions of NSW (Condong) (Figure 45) and Mulgrave (Figure 46) were also investigated during the 2012 season. SRDC Project DPI021 Final Report_without appendices.doc 56 Bundaberg/ Isis: Figure 39. SPOT5 image of the Bundaberg/ Isis growing region (3600km2) overlayed with crop boundary vector file (black lines). The evaluation of the generic algorithm over the Bundaberg and Isis growing regions during 2010- 2012, produced encouraging results in the prediction of average yield, as shown in Table 6. At the block level the accuracies were quite poor, ranging from predictions that were 89% under estimated to 494% over estimated (Isis 2012). Note 100% signifies an accurate prediction. Slight inaccuracies with Mill data as well as discrepancies with yield consignments to actual blocks also contributed to this degree of variability. Table 6 . Predicted versus actual average yield for the Bundaberg and Isis growing regions. Range of Gro wing Number of Pred. ave. Act. ave. Prediction at block Harvest year Region crops Yield (TCH) Yield (TCH) level (% of actual) 2010 Bundaberg 354 4 7 9.7 81.8 36 to 364 2011 Bundaberg 382 4 8 0.1 73.3 16 to 436 2012 Bundaberg 321 7 8 8.0 88.9 13 to 410 2010 2011 2012 Isis Isis Isis 277 2 420 5 400 0 8 4.0 9 8.4 9 2.5 84.0 83.3 96.0 40 to 638 39 to 594 11 to 595 SRDC Project DPI021 Final Report_without appendices.doc 57 The obvious exception to these predictions is the 2011 season, where severe rainfall events occurring at the end of 2010, significantly reduced crop yield and therefore contributed to the over predictions (Table 6). This yield deficit can be clearly seen in Figure 40 a, with the separation of the exponential trend line representing the generic algorithm (dotted line) to that derived from 2011 crop data (green line). The high similarity between the generic trend line to that produced by the 2010 (blue line) and the 2012 (red line) data explains the close estimates achieved for those seasons. For the Isis region (Figure 40 b) a similar pattern to that of the Bundaberg region exists for the 2010 and 2011 trend lines. However, the 2012 (red line) displays a very different slope and intercept. The 4% under estimation of yield achieved for the 2012 season was due to the average GNDVI value (0.5931) for the 2012 crops (n = 4000) coinciding with the intercept between the 2012 and generic algorithm trend lines. The reason for the different 2012 trend line is unknown. It is hypothesised to be possible a seasonal effect or the result of inconsistencies with the mill vector file and consignment data. 2010 ave. GNDVI: 0.567 2011 ave. GNDVI: 0.5697 2012 ave. GNDVI: 0.585 a. 2010 ave. GNDVI: 0.5762 2011 ave. GNDVI: 0.6039 2012 ave. GNDVI: 0.5931 b. Figure 40. Exponential trend lines produced between GNDVI and yield for the generic algorithm (dashed line) as well as all crops imaged within the Bundaberg (a) and Isis (b) region during the 2010 (blue line), 2011 (green line) and 2012 (red line) growing seasons. SRDC Project DPI021 Final Report_without appendices.doc 58 Herbert: Figure 41. SPOT5 image of the Herbert growing region (3600km2) overlayed with crop boundary vector file (black lines). The evaluation of the generic algorithm over the Herbert growing region during 2011 and 2012 also produced encouraging results (Table 7). Note no image was captured during 2010 due to continual cloud cover. Again predictions at the block level were also highly variable. Table 7. Predicted versus actual average yield for the Herbert growing region. Range of Gro wing Number of Pred. ave. Act. ave. Prediction at block Harvest year Region crops Yield (TCH) Yield (TCH) level (% of actual) 2011 Herbert 859 6 5 1.4 55.0 16 to 350 2011 Herbert 6481 (no SO*) 5 6.9 53.5 29 to 350 2012 Herbert 1 5463 7 5.0 72.0 17 to 514 * SO refers to stand over crops that were not harvested in 2010 due to severe weather events. The influence of stand over (SO) crops was clearly identified in the prediction accuracies achieved for the 2011 growing season, with an under prediction occurring when SO crop were included and an over prediction once removed. These SO crops contributed up to 33% of the all crops imaged for this season. On average, the SO crops displayed a lower GNDVI value (SO = 0.4898; non SO = 0.5077) due mainly to being severely lodged, it did however yield slightly higher as a result of an extra season of growth. With crops denoted as SO removed for the 2012 season, an over prediction of 6.4% was achieved. A comparison of the generic algorithm trend line to those produced from the 2011 and 2012 data sets (Figure 42) SRDC Project DPI021 Final Report_without appendices.doc 59 demonstrates some separation at the higher GNDVI values and therefore explanation for the over predictions. It is therefore suggested that a Herbert specific algorithm be investigated in the future. 2011 ave. GNDVI: 0.5077 (no stand over) 2012 ave. GNDVI: 0.5563 Figure 42. Exponential trend line produced between GNDVI and yield for the generic algorithm (dashed line) as well as all crops imaged within the Herbert region during the 2011 (blue line) and 2012 (green line) growing seasons. Burdekin Figure 43. SPOT5 image of the Burdekin growing region (3600km2) overlayed with crop boundary vector file (black lines). SRDC Project DPI021 Final Report_without appendices.doc 60 The prediction estimates for average yield produced by the generic algorithm for the Burdekin region was extremely poor for both the 2010 and 2011 season (Table 8). Table 8 . Predicted versus actual average yield for the Burdekin growing region, generic algorithm. Range of Gro wing Number of Pred. ave. Act. ave. Prediction at block Harvest year Region crops Yield (TCH) Yield (TCH) level (% of actual) 2010 Burdekin 4573 91.9 129.6 22 to 411 2011 Burdekin 4999 83.2 120.0 17 to 360 The under predictions of 29% (2010) and 31% (2011) were attributed to the different seasonal and growing conditions experienced within this region, when compared to Bundaberg/ Isis. As such a Burdekin specific generic algorithm (equation 2) was developed from the 2010 data and then evaluated over the 2011 and 2012 seasons. This resulted in vastly improved prediction estimates (Table 9). Equation 2: y = 12.691 * е (3.8928 * x) Where y = predicted average yield (TCH) x = average GNDVI value extracted from TOA SPOT5 image. Table 9. Predicted versus actual average yield for the Burdekin growing region, Burdekin algorithm. Range of Gro wing Number of Pred. ave. Act. ave. Prediction at block Harvest year Region crops Yield (TCH) Yield (TCH) level (% of actual) 2011 Burdekin 499 9 118.8 1 20.0 29 to 468 2012 Burdekin 692 1 110.0 1 05.0 16 to 448 As seen in Figure 44, differing trend lines were produced from the correlations between GNDVI and yield for the three growing seasons, a result most likely attributed to seasonal variability. Considering this, it was fortunate that the prediction estimates for 2011 and 2012 were so high, i.e. 1% under prediction (2011) and 4.8% over prediction (2012). For 2011 this was the result of the average GNDVI value (0.5744) closely aligning to the 2010 and 2011 trend line intercept; while for 2012, the average GNDVI value (0.5543) corresponded to a point where both the 2010 and 2012 trend lines exhibited minor separation i.e. 5 TCH. SRDC Project DPI021 Final Report_without appendices.doc 61 2010 ave. GNDVI: 0.5919 2011 ave. GNDVI: 0.5744 2012 ave. GNDVI: 0.5543 Figure 44. Exponential trend line produced between GNDVI and yield for the Burdekin generic algorithm (blue line) as well as all crops imaged within the Burdekin region during the 2011 (green line) and 2012 (red line) growing seasons. Condong (NSW) Figure 45. SPOT5 image of the Condong growing region (3600km2) overlayed with crop boundary vector file (black lines). Following interest from the Condong (NSW) growing region, an estimate of average yield for the 2012 season was derived using the Bundaberg algorithm. The prediction produced for the 1 year cane was again encouraging (Table 10). SRDC Project DPI021 Final Report_without appendices.doc 62 Table 10 . Predicted versus actual average yield for the Condong growing region. Range of Gro wing Number of Pred. ave. Act. ave. Prediction at block Harvest year Region crops Yield (TCH) Yield (TCH) level (% of actual) 2012 Condong 208 7 6 7.6 70.0 23 to 468 As this prediction was only applied for one season, the high estimate achieved i.e. 3.4% under prediction (Table 10), may be an anomaly and as such additional research is required to further validate the algorithm. It is hypothesised that a NSW specific algorithm would be required to account for different climate within this region when compared to Bundaberg. An additional algorithm would also be required to account for the cane grown within this region over two years. Mulgrave An additional scoping study assessing yield prediction over the Mulgrave growing region was also undertaken during the 2012 season. Without access to a SPOT5 image the algorithm (equation 3) was derived from GNDVI values extracted from an IKONOS image captured 26 May 2010 and corresponding yield of 833 crops (coloured polygons in Figure 46 b). The prediction was applied to GNDVI values extracted from a GeoEYE image captured 4 May 2012 (1324 crops). 2010 target region used to develop algorithm from IKONOS image a. b. SRDC Project DPI021 Final Report_without appendices.doc 63 Figure 46. a. Area within the Mulgrave growing region where yield was predicted for the 2012 growing season. b. target area used to derive the 2010 yield prediction algorithm. Equation 3: y = 15.641 * е (3.4775 * x) Where y = predicted average yield (TCH) x = average GNDVI value extracted from TOA SPOT5 image. The predicted average yield for the 1324 crops closely aligned with that of the whole Mulgrave growing region for 2012 (Table 11). Table 11 . Predicted versus actual average yield for the Mulgrave growing region Range of Gro wing Number of Pred. ave. Act. ave. Prediction at block Harvest year Region crops Yield (TCH) Yield (TCH) level (% of actual) 2012 M ul gra ve 132 4 8 6.1 84.4 33 to 388 This result, although encouraging requires additional validation over a number of growing seasons. It does however demonstrate that imagery other than SPOT5 can be used to derive yield prediction algorithms. Summary of regional yield prediction The estimates of regional average yield produced by this project were found to be constantly high for all growing regions investigated over the majority of growing seasons. The results are very encouraging and support remote sensing as an effective additional tool for validating seasonal estimates provided by current methods. The fact that these algorithms are non cultivar and non crop class specific, as well as in some cases regionally and seasonally insensitive reduce the complexity of analysis in the event that these protocols are adopted by industry. Between 2009- 2012, fifty-three different varieties were planted in the Herbert, twenty-six in Bundaberg and nineteen in the Burdekin growing region. If other variables such as the segregation of regions into smaller climate driven micro regions or crop class were also accounted for then the number of algorithms required would be substantial. The regional separation of trend lines identified across some growing seasons, particularly for the Burdekin and Isis regions, does pose a concern for the future accuracy of this approach. It is therefore suggested that the implementation of an agro- meteorological model be investigated for the purpose of normalising seasonal variability. Similarly, the use of the generic algorithms was identified to have major limitations when used for the prediction of yield at the crop level, with large under and over predictions identified at all sites. This indicates that if predictions are required at this scale then it is likely that algorithms for each cultivar and possibly crop class would be required. SRDC Project DPI021 Final Report_without appendices.doc 64 6.3.5. Validation of the generic algorithm at the block and within block level. Yield maps derived from SPOT5 imagery using the generic algorithm (Figure 47 a and c) displayed similar spatial trends to those produced by the field calibrated IKONOS maps (Figure 47 b and d). However, some inconsistencies were identified particularly at the high and low range, a result consistent to that identified in Figure 37. a. b. c. d. Figure 47. Classified yield map derived from SPOT5 image using the generic algorithm (a and c) and corresponding IKONOS yield map derived from field calibration samples (b and d). A test at the within crop level identified a similar inability of the generic algorithm to accurately predict high and low yield in terms of TCH. In Figure 48, the slope (0.771) and intercept (10.369) indicate the difference between actual yield measured at a number of locations within two Bundaberg crops and predicted yield using the SPOT5 algorithm (grey markers; Figure 48). When compared to in field measurements from two crops in the Burdekin region (red makers; Figure 48), the generic algorithm was found to consistently under predict yield. This result supported the need for a Burdekin specific algorithm that could account for the substantially different growing conditions within that region. SRDC Project DPI021 Final Report_without appendices.doc 65 Figure 48. Comparison of actual cane yield (TCH) measured at sample locations within 2 Bundaberg and 2 Burdekin crops, to predicted yields derived from the SPOT5 generic algorithm. A number of small growing areas within the Bundaberg region exhibited very high predictions of yield at the block level. Relmay farms for example, produced an average prediction estimate of 97%; S.D. 10% for 61 blocks, similar to that for Bullseye with also 97%; S.D. 17% (49 crops) and Cayley 87%; S.D. 11% (31 crops). A similar test in the Burdekin and Herbert region produced very poor prediction estimates (data not shown). To further test the accuracies of yield prediction at the bock level for each of the regions a simple linear regression was fitted between actual and predicted observations for each of the data sets (Bundy 2010, 2011 and 2012; Isis 2010; Burdekin 2011 and 2012 and Herbert 2011 and 2012) (Table 12). The y variables were actual yield data provided by each mill whilst the predicted values were the x variables. If the predictions were identified to be 100% accurate then the regression line would have an intercept of zero and a slope of one, with the Rsquare value to be 100%. Table 12. Results of linear regressions run between actual and predicted block yields for a number of growing regions and seasons. Location Year R-Square Intercept Slope Bundaberg 2010 52.7 -1.6 1.0 Bundaberg 2011 38.9 7.9 0.8 Bundaberg 2012 49.8 5.9 1.0 Isis 2010 34.1 12.1 0.9 Herbert 2011 7.3 31.0 0.5 Herbert 2012 23.0 18.9 0.7 Burdekin (all data) 2011 4.2 69.9 0.4 Burdekin (without SO) 2011 32.0 -29.0 1.1 Burdekin (SO only) 2011 22.0 3.4 1.4 Burdekin 2012 14.0 25.7 0.7 SRDC Project DPI021 Final Report_without appendices.doc 66 From this analysis, the Bundaberg 2010 data set provided the strongest prediction estimates at the block level, with an intercept found to be non-significantly different to zero following a t-test, with a slope close to 1 (-1.6; Table 11). With the initial generic algorithm being predominantly derived from this 2010 Bundaberg data set, this result is not surprising. This also explains why high prediction accuracies were also obtained for the Bundaberg 2011 and 2012 data sets. Significant variation still existed about the regression line. The predictions from the other data sets were poor, with the intercepts and slopes of these regression lines being far from the ideal. For Burdekin 2011 the removal of standover crops (without SO) greatly improved the prediction although the intercept was still significantly different to zero. Interestingly the analysis of the standover crops alone produced encouraging results, indicating high estimates of yield may be achieved for stand over crops as long as they are analysed independently. Note for this analysis, no data was omitted despite many data points having large residuals or high leverage. These results indicate that although the generic algorithm has the ability to identify yield trends at the crop level; its ability to accurately predict actual tonnes of cane per hectare (TCH) was poor. It is suggested that for this to be achieved, specific algorithms would need to be developed for differing cultivars, crop class and growing region, as well as be normalised for seasonal variability. An example of how this may occur is provided in Figure 49, where the refining of data for the Burdekin 2011 season improves the correlation coefficient from R2 = 0.052 when all data is used, to R2 = 0.205 with the removal of stand over, R2 = 0.291 with the removal of all crop classes except for plant cane and lastly R2 = 0.421 with the inclusion of one variety cv. Q208. TCH TCH 200 All data y = 16.659e2.2368x 160 R2 = 0.052 120 80 40 0 0.2 0.3 0.4 0.5 GNDVI 0.6 200 No Standover y = 2.4232e5.873x 160 R2 = 0.2052 120 80 40 0 0.2 0.3 0.4 0.5 0.6 0.7 GNDVI SRDC Project DPI021 Final Report_without appendices.doc 67 200 Plant Cane Only y = 1.6901e6.8089x 160 R2 = 0.2921 120 200 Plant Cane cv. Q208 y = 0.5895e8.6808x 160 R2 = 0.4206 120 TCH TCH 80 80 40 40 0 0.25 0.35 0.45 0.55 0.65 GNDVI 0 0.4 0.5 GNDVI 0.6 Figure 49. Burdekin 2011 example demonstrating how the refining of data in terms of stand over, crop class and variety can improve the correlation between GNDVI and cane yield (TCH). By having access to the comprehensive Mill vector data layers, the ability to derive and then implement yield predictions based on crop class and cultivar is highly achievable. However further research is required to develop the appropriate processing steps and delivery protocols. 6.3.6. Production of classified yield maps using the generic algorithms. Through this project, a process was established for the rapid production of classified yield maps by applying the generic algorithms to each block at the pixel level. Although the accuracies of these maps varied to some degree as discussed in section 6.3.5., the ability to access these maps and identify yield trends at the block, farm and regional level, within a growing season, offers some benefit. As seen in Figure 50, the majority of crops display yield variability, with the overall yield trend across the farm only becomes apparent when all blocks are displayed. In this case there is an obvious reduction in yield towards the south eastern corner of the farm, a result of the large dam (indicated), being filled to capacity during the late 2010 heavy rainfall event. This is believed to have increased the hydraulic pressure of the sub surface aquifers, and forced the surrounding water table to the surface producing anaerobic soils and incidences of salinity. With this information, the grower could easily identify the blocks impacted, and by having access to similar maps derived over a number of seasons, be able to determine if the area affected increases or decreases with time. In this example, the predictions of average block yields were identified to be highly accurate, so there is some confidence that accurate estimates of lost productivity can be made, thus justifying the need to apply remedial action or not. SRDC Project DPI021 Final Report_without appendices.doc 68 DAM Figure 50. Classified yield maps derived from the generic algorithm overlayed on to a SPOT5 false colour image. The ability to view these classified yield maps at the regional level also provides some indication of sub-regional trends which can assist in harvest management or support localised efforts to increase productivity. In the example below (Figure 51 a) it can be seen that there are a number of high yielding (highlighted in red) and low yielding (highlighted in black) sub regions within the Herbert growing area. These are likely to be attributed to soil type variability and topography. These same trends were also apparent in the classified image derived from actual yield values post 2012 harvest, derived by the Herbert resource information Centre (HRIC) (Figure 51b). This result gives some support to the accuracy of trends derived from the imagery based yield prediction algorithm; although again there is under prediction of higher yield values and over prediction of low values occurred. SRDC Project DPI021 Final Report_without appendices.doc 69 a. b. Figure 51 a. Sub regional trends of yield production as indicated by classified yield maps for the Herbert (a) and Bundaberg (b) regions. SRDC Project DPI021 Final Report_without appendices.doc 70 In Figure 52, the variation between the high and low yielding sub-regions in this Bundaberg example was the result of severe flooding following the 2010 heavy rainfall event. The flood water removed top soil and nutrient from the northern side of the river which resulted in reduced yield during the 2011 season (highlighted in white). c. Figure 52. Sub regional trends of yield production as indicated by classified yield maps for the North Bundaberg region. This information if generated annually can provide a strong indication of how yield trends within sub regions respond to variable seasonal conditions. This can guide where more tolerant varieties for example should be planted, when they should be planted and where crops grown in some sub-regions should be possibly discontinued due to continual poor performance. Note: The overlayed yield maps for all regions investigated as well as the associated GoogleEarth files are not included within this report due to confidentiality agreements with the respective Mills. 6.4. Implement optimal image processing and delivery protocols for the rapid distribution of classified imagery to agronomists, growers etc. This project successfully developed a number of processing protocols that enabled all project objectives to be achieved. The most novel of which included the derivation and then application of the regional yield prediction algorithms and classified yield maps en mass. The project identified a number of softwares suitable for achieving these objectives including ENVI, ArcGIS, Starspan GUI, and GoogleEarth with the later identified to be highly effective for the distribution of spatial information. The appendices of this report include a number of Tutorials that best demonstrate the processes developed, these include: SRDC Project DPI021 Final Report_without appendices.doc 71 • Tutorial 1: ENVI: Converting ‘At Sensor’ Digital Numbers to ‘Top of Atmosphere’ reflectance values • Tutorial 2: ENVI: Georectification of satellite imagery using an orthorectified base layer and derivation of a GNDVI image. • Tutorial 3: ArcGIS: Conversion of Mapinfo (.TAB) files into ArcGIS (.SHP) files: • Tutorial 4: ArcGIS: Buffering of polygons and removal of those affected by cloud before the extraction of spectral data. • Tutorial 5: Starspan GUI: Extracting average spectral values and associated attribute information for multiple blocks. • Tutorial 6: ENVI: Producing a classified vegetation index map of a cane crop from a 4 band satellite image. • Tutorial 7: ENVI: Extracting point source spectral information from imagery using regions of interest (ROI’s). • Tutorial 8: ENVI: Converting VI pixel values into yield (TCH) using an exponential linear algorithm. • Tutorial 9: ENVI: Creating Google Earth KMZ files from Geotiffs. 6.5. Provide recommendations to participating growers, consultants and industry representatives on the potential cost / benefit of implementing RS technologies into current agronomic management practices The strongest indication of the cost/ benefit of the outcomes generated by this research were the increasing level of industry support and involvement experienced throughout the life of the project. From an initial target objective of three farms in the Herbert, Burdekin and Bundaberg growing regions, the project ended up collaborating and producing outcomes for 6 growing regions including the generation of over 33,000 yield maps in 2012. The development of the accurate regional yield prediction algorithms was seen by industry as offering the greatest benefit-cost. In 2010, severe weather events caused a major discrepancy between the in season yield estimations made for each growing region to that delivered after harvest. This resulted in a sugar deficit of millions of tonnes being available to fill forward selling obligations and ultimately cost the industry, including the growers, a substantial amount of money. As a result the development of an additional tool for predicting yield, particularly one that can identify the direct impacts of within season weather anomalies, is of great benefit. As mentioned the cost for each SPOT5 scene required for each region is ~ $AUD3800 plus processing costs, which in comparison to potential financial penalties passed on from incorrect predictions, is minimal. At the crop level, the ability to identify low performing zones can allow an estimation of lost productivity in monetary terms to be made. The example below (Figure 53) demonstrates how a simple estimate of lost production can made from an image derived yield map. In this example the low NDVI regions yielded on average 51 TCH, the medium 73 TCH and the high 102 TCH (based on field sampling). By multiplying the area of each of three yield classes by the average corresponding yield and then summing the results, the total crop SRDC Project DPI021 Final Report_without appendices.doc 72 yield can be estimated at 1535 tonnes of cane from 18.6 ha. If the entire crop yielded at 102 TCH (average yield measured in the high NDVI zones) then the total yield would have equated to 1898 tonnes. A simple subtraction of actual total yield from optimal total yield identifies a yield deficit of 364 tonnes, or 19.6 TCH. Expressed in monetary terms this would equate to $725 per hectare (at $37 tonne of cane). With the low and medium yielding area extending over 10.2 ha the total monetary loss could be estimated at $7,400 from less than optimum productivity. Figure 53. 3 zone yield map derived from IKONOS imagery with field sampling zones indicated by yellow markers. By identifying the nature of the limiting factor, in this case poor drainage and sodic soils, a decision can be made on the benefit-cost of applying remedial action such as the application of gypsum / organic matter to increase drainage as well as the re-lasering of beds prior to replant. To further improve production costs, the 3 zone yield map can be used to direct the variable rate application of gypsum, rather than a blanket application. As well as soil related issues the high resolution IKONOS imagery was effective in identifying additional constraints to production such as weed infestation (Figure 54a) and damage from rat, cane grub (Figure 54b), soldier fly (Figure 54c) and pig. Although these are just visual observations they do enable a grower to see what is occurring within a cane field and therefore guide within crop assessments. Again, by understanding where a constraint is occurring as well as the area affected, the grower can implement a modified management strategy such as targeted herbicide and insecticide applications. SRDC Project DPI021 Final Report_without appendices.doc 73 Figure 54. Occurrences of weed infestation (a), cane grub (b) and soldier fly (c) damage identified by IKONOS imagery. Cane grubs in particular were identified to be an economically important pest and as such an additional SRDC funded project (BSS342) is being undertaken to determine if an automated image based detection and warning systems can be developed. This will incorporate both textural and spectral discrimination of imagery in an attempt to define damage specific to cane grub. Lastly, high resolution imagery may offer some benefit to industry as a screening tool for replicated trials. Temporal imagery may indicate more homogenous locations for the placement of trials whilst imagery captured in- season can be used to identify the effects of inherent block constraints on individual replicates (Figure 55). With a surrogate measure of these external constraints such as NDVI, biometricians can add some weighting to the analysis so as to better account for non replicate specific influences. Furthermore, if trials are sampled for specific in season measure of performance (i.e. biomass, SPAD, foliar N etc) then these measures can be used to calibrate the image allowing for a trait specific map to be derived. This then could be used as a rapid screening device for the trial or possibly allow the correlations to be extrapolated over the surrounding growing area as a predictive tool. Further research will be undertaken to investigate this possibility. NDVI- brighter the plot the higher the plant vigour Low vigour area transcending across blocks- inherent to block maybe related to soil or drainage Figure 55. High resolution imagery of a replicated cane trial with the effects of an inherent block constraint i.e. soil type, drainage influencing a number of replicates. SRDC Project DPI021 Final Report_without appendices.doc 74 7. Conclusion This project successfully achieved its 5 main objectives producing applications that were directed by industry and therefore considered both relevant and adoptable. SPOT5 imagery was identified to be the best suited for coverage at the regional level due to cost, tile size, repeat capture time and spatial resolution. IKONOS was identified to provide the most feasible high resolution imagery when purchased under a multi capture deal, as well as highly effective for identifying a range of biotic and abiotic sub metre constraints. The Raptor sensor was demonstrated as a possible future provider of NDVI maps, particularly under cloud prone environments. This project identified the effectiveness of imagery in accurately identifying within season growth variability, particularly when expressed through a vegetation index. These maps were highly effective for directing in crop sampling, with the information gained providing a useful tool for supporting modified management strategies. Although the timing of image capture, based on delays from continued cloud cover, meant the implementation of revised strategies could not occur until post harvest, due to the phenological stage and size of the crop. The imagery captured between March and May was found to be highly correlated to final harvested yield and thus enabled surrogate yield maps to be derived from calibrated GNDVI maps using in field yield measurements. This result was consistent with previous research that also identified the stabilisation period of cane, between vegetative growth and senescence to be the best correlated to final yield. The Bundaberg and Burdekin non cultivar, non crop class specific algorithms were found to be highly accurate in predicting average yield at the regional level. This result supports the adoption of imagery as an additional tool to support existing methods, for guiding harvesting scheduling and forward selling decisions. At the crop level, only predictions within the Bundaberg region were found to be accurate. Further research will investigate if the development of crop and class specific algorithms, normalised for seasonal variability, will improve predictions accuracies of actual yield at the block level. Analysis protocols and methodologies developed by this project will greatly assist subsequent adopters of these technologies both within Australian and overseas, particularly with the inclusion of tutorials with this final report. The identification of the freeware software Starspan GUI for the rapid extraction of attribute and spectral data as well as GoogleEarth for the delivery of derived image products were both novel outputs of this project. It is hoped that further research will allow the development of a ‘turn key’ analysis software that will assist industry in the rapid prediction of cane yield as well as the distribution of in season yield maps. Finally, this project received strong interest and collaboration from all facets of the Australian cane industry, a result that strongly indicates an understanding of the potential benefit that these spectral technologies offer. This support led to additional funding by SRDC Project DPI021 Final Report_without appendices.doc 75 SRDC to further refine the yield prediction algorithms, particularly at the crop level, as well as investigate remote sensing as tool for the non destructive screening of breeding trials, the mapping of foliar nitrogen and as a tool for the detection of cane grub. 8. Acknowledgements This project received over whelming support from the Australian cane industry including access to highly sensitive GIS information, essential data that allowed many of the project outcomes to be achieved. As such the project team would like to acknowledge the assistance of Bundaberg Sugar Ltd, Isis Central Sugar Mill Co. Ltd, Maryborough Sugar Factory; Mulgrave Central Mill, Mackay Sugar Mill; as well as Sucrogen and the Herbert Resource Information Centre (HRIC). The project team would also like to acknowledge the contributions made to the project by a number of agronomic services including Farmacist, Herbert Cane Productivity Services Ltd and Burdekin Productivity Services Ltd; Research bodies including James Cook University (JCU), University of Southern Queensland (USQ) and University of Queensland (UQ); Government departments DAFFQ, DSITIA, CSIRO, Reef Catchments; Industry bodies such as BSES; and growers in particular Relmay Farming, Bullseye Precision Farming, Denis Pozzebon, Brian Tabone, Jay Hubert and the late Alan Mann. Lastly the project team acknowledge SRDC whom in collaboration with DAFF Qld, CSIRO and University of New England (UNE) provided the funding to undertake this research. 9. References Abdel-Rahman, EM and Ahmed, FB (2008). The application of remote sensing techniques to sugarcane (Saccharum spp Hybrid) production: a review of the literature. International Journal of Remote Sensing 29(13), 3753- 3767. Almeida, TIR, Filho, CR De Souza and RossettoR (2006). ASTER and Landsat ETM+ images applied to sugarcane yield forecast. International Journal of Remote Sensing 27. pp 4057-4069. Bégué, A, Lebourgeois, V, Bappel, E, Todoroff, P, Pellegrino, A, Baillarin, F, and Siegmund, B (2010). Spatio-temporal variability of sugarcane fields and recommendations for yield forecasting using NDVI. International Journal of Remote Sensing. 31(20).5391-5407. Bégué, A, Todoroff, P and Pater, J (2008). Multi-time scale analysis of sugarcane within-field variability: improved crop diagnosis using satellite time series? Precision Agriculture 9.161171. SRDC Project DPI021 Final Report_without appendices.doc 76 Benvenuti, F and Weill, M (2010). Relationship between multi-spectral data and sugarcane crop yield. Proceedings of the 19th world Congress of Soil Science Soil Solutions for a Changing World 1-6 August 2010 Brisbane Australia.33-36. Bramley RGV, Gobbett DL, Panitz JH, Webster AJ, McDonnell P. 2012. Soil sensing at high spatial resolution – broadening the options available to the sugar industry. Proceedings of the Australian Society of Sugar Cane Technologists, 34th Conference, Cairns. Electronic format. 8 pp. Coventry, R.J., Hughes, J.R., Reid, A.E., McDonnell, P. (2010). Stability of spatial patterns defined by electrical conductivity mapping of soils within sugarcane paddocks. Proceedings of the Australian Society of Sugar Cane Technologists, 32, 397409. Davis, R, Bartels, R and Schmidt, E (2007). Precision Agriculture Technologies- Relevance and application to sugarcane production. In SRDC Technical Report 3/2007: Precision agricultural options for the Australian sugarcane industry Eds R Bruce.60-117. De Lai, R, Packer, G, Sefton, M, Kerkwyk, R and Wood, AW (2011). The Herbert Information Portal: Delivering real-time spatial information to The Herbert River sugar community. Proceedings of The Australian Society of Sugar Cane Technologists 33. Fernandes, JL, Rocha, JV and Lamparelli, RAC (2011). Sugarcane yield estimates using time series analysis of spot vegetation images. Science in Agriculture 68(2).139-146. Gitelson A A, Stark R, Grits U, Rundquist D, Kaufman Y, Derry D. Vegetation and soil lines in visible spectral space: A concept and technique for remote estimation of vegetation fraction. Int J Remote Sens, 2002, 23(13): 2537-2562. Huete, AR, Liu, HQ, Batchily, K and Leeuwen, W (1997). A Comparison of vegetation indices over a global set of TM images for EOS-MODIS. Remote Sensing of Environment 59.440-451. Huete, A, Didan, K, Miura, T, Rodriguez, EP, Gao, X and Ferreira, LG (2002). Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sensing of Environment 83.195-213. Lee-Lovick, G and Kirchner, L (1991). Limitations of Landsat TM data in monitoring growth and predicting yields in sugarcane. Proceedings of Australian Society of Sugar Cane Technologists 13.124-129. Kishna Rao, PV, Venkateswara Rao, V, Venkataratnam, L (2002). Remote Sensing: A technology for assessment of sugarcane crop acreage and yield. Sugar Tech 4 (3&4). 97-101. SRDC Project DPI021 Final Report_without appendices.doc 77 Markley, J Ashburner, B and Beech, M (2008). The development of a spatial recording and reporting system for productivity service providers. Proceedings of Australian Society of Sugar Cane Technologists 30.10-16. Markley, J, Raines, A and Crossley, R (2003). The development and integration of remote sensing GIS and data processing tools for effective harvest management. Proceedings of Australian Society of Sugar Cane Technologists 25.2003. Minasny, B., A.B. McBratney, and B.M. Whelan. 2005. VESPER version 1.62. Available at http://www.usyd.edu.au/su/agri/acpa (verified 21 Noonan, MJ (1999). Classification of fallow and yields using Landsat TM data in the sugarcane lands of the Herbert River Catchment. Herbert Resource Information Centre Qld Website link: http://wwwhricorgau/home/JournalPublicationsaspx. Pitt A (Pers comm). Grower Services Superintendent Bundaberg Sugar Pty Ltd. Robson, A, Abbott, C, Lamb, D and Bramley, R (2011). Paddock and regional scale yield prediction of cane using satellite imagery Poster Abstract. Proceedings of the Australian Society of Sugar Cane Technologists. 33rd Conference Mackay Qld AUS 4 – 6th May 2011. Robson, A, Abbott, C, Lamb, D and Bramley, R (2010). Remote Sensing of Sugarcane; answering some FAQ’s. Australian Sugarcane 2011. p6-8. Rudorff, BFT and Batista, GT (1990). Yield estimation of sugarcane based on agrometerological- spectral models. Remote Sensing of Environment. 33.183-192. Rueda, CA, Greenberg, JA and Ustin, SL (2005). StarSpan: A Tool for Fast Selective Pixel Extraction from Remotely Sensed Data Center for Spatial Technologies and Remote Sensing (CSTARS). University of California at Davis Davis CA (Starspan GUI website link: https://projectsatlascagov/frs/downloadphp/581/install-starspan-win32-020jar) Simões, MDS, Rocha, JV and Lamparelli, RAC (2005). Spectral variables growth analysis and yield of sugarcane. Science in Agriculture 62.199-207. Simões, MDS, Rocha, JV and Lamparelli, RAC (2009). Orbital spectral variables growth analysis and sugarcane yield. Science in Agriculture 66(4).451-461. SPOT Image (2008) SPOT Image Homepage 28th April 2008. Website link http://wwwspotimagecom. Wang, Fu-min, Huang, Jing-feng, Tang, Yan-lin and Wang, Xiu-zhen (2007). New vegetation index and its application in estimating leaf area index of Rice. Rice Science 14(3).195-203. SRDC Project DPI021 Final Report_without appendices.doc 78 10. Appendix 1: Tutorials Tutorial 1: ENVI: Converting ‘At Sensor’ Digital Numbers to ‘Top of Atmosphere’ reflectance values On the main toolbar open: -Basic Tools- Band Math In the Band Math window type the expression: In this example (IKONOS image) the text will appear like this: (3.14159265358979323846*1.0307122576*((10000*float(b1))/(728*71.3)))/(cos(0.851987187)*1930.9) Where: π = 3.14159265358979323846 (double precision) d2 = 1.0307122576 = ((10000*float(b1))/(728*71.3)) = 0.851987187 = 1930.9 A new window will open asking to link B1 to an appropriate band.Choose the suitable Band to transform. Choose an appropriate file name identifying the band in the output. In the next step, several of these will be layer stacked together to form one file. Repeat for each band. To create one data file from the transformation outputs use the ‘Layer Stacking’ tool. On the main toolbar open- Basic Tools - Layer Stacking Choose the appropriate layers for the new data file using the ‘Import File’ button. The order in which the individual files are selected, denotes their subsequent order in the final layer stacked image. Choose a filename and projection – press OK to finish the process. Theory Converting ‘At Sensor’ digital numbers to ‘Top of Atmosphere’ reflectance values. Equation 1: SRDC Project DPI021 Final Report_Appendices.doc 2 Where: = Top of Atmosphere Reflectance (Unitless Planetary Reflectance) = Earth-Sun distance Factor (ratio of the actual distance to the mean distance) = Solar Zenith angle in degrees (converted to radians) = Spectral radiance at the sensor’s aperture = Mean solar exoatmospheric irradiances Finding = Earth-Sun distance factor (ratio of the actual distance to the mean distance) This value is a ratio of the distance of the Earth to the Sun on the individual day of image capture to the mean distance of the Earth to the Sun for every day in the year. Earth-Sun distance (d) in astronomical units for Day of the Year (DOY)+ Day of the Month (DOM) for Non-Leap Years. For leap years add 1 DOY to dates after 28 February. (January 1 = Julian Day 1) DOY 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 Earth-Sun distance (d) in astronomical units for Day of the Year (DOY) d DOY d DOY d DOY d DOY d DOY 0.98331 62 0.99133 123 1.00806 184 1.01670 245 1.00898 306 0.98330 63 0.99158 124 1.00831 185 1.01670 246 1.00874 307 0.98330 64 0.99183 125 1.00856 186 1.01670 247 1.00850 308 0.98330 65 0.99208 126 1.00880 187 1.01670 248 1.00825 309 0.98330 66 0.99234 127 1.00904 188 1.01669 249 1.00800 310 0.98332 67 0.99260 128 1.00928 189 1.01668 250 1.00775 311 0.98333 68 0.99286 129 1.00952 190 1.01666 251 1.00750 312 0.98335 69 0.99312 130 1.00975 191 1.01664 252 1.00724 313 0.98338 70 0.99339 131 1.00998 192 1.01661 253 1.00698 314 0.98341 71 0.99365 132 1.01020 193 1.01658 254 1.00672 315 0.98345 72 0.99392 133 1.01043 194 1.01655 255 1.00646 316 0.98349 73 0.99419 134 1.01065 195 1.01650 256 1.00620 317 0.98354 74 0.99446 135 1.01087 196 1.01646 257 1.00593 318 0.98359 75 0.99474 136 1.01108 197 1.01641 258 1.00566 319 0.98365 76 0.99501 137 1.01129 198 1.01635 259 1.00539 320 0.98371 77 0.99529 138 1.01150 199 1.01629 260 1.00512 321 0.98378 78 0.99556 139 1.01170 200 1.01623 261 1.00485 322 0.98385 79 0.99584 140 1.01191 201 1.01616 262 1.00457 323 0.98393 80 0.99612 141 1.01210 202 1.01609 263 1.00430 324 0.98401 81 0.99640 142 1.01230 203 1.01601 264 1.00402 325 0.98410 82 0.99669 143 1.01249 204 1.01592 265 1.00374 326 0.98419 83 0.99697 144 1.01267 205 1.01584 266 1.00346 327 0.98428 84 0.99725 145 1.01286 206 1.01575 267 1.00318 328 0.98439 85 0.99754 146 1.01304 207 1.01565 268 1.00290 329 0.98449 86 0.99782 147 1.01321 208 1.01555 269 1.00262 330 0.98460 87 0.99811 148 1.01338 209 1.01544 270 1.00234 331 0.98472 88 0.99840 149 1.01355 210 1.01533 271 1.00205 332 0.98484 89 0.99868 150 1.01371 211 1.01522 272 1.00177 333 0.98496 90 0.99897 151 1.01387 212 1.01510 273 1.00148 334 0.98509 91 0.99926 152 1.01403 213 1.01497 274 1.00119 335 0.98523 92 0.99954 153 1.01418 214 1.01485 275 1.00091 336 0.98536 93 0.99983 154 1.01433 215 1.01471 276 1.00062 337 0.98551 94 1.00012 155 1.01447 216 1.01458 277 1.00033 338 0.98565 95 1.00041 156 1.01461 217 1.01444 278 1.00005 339 0.98580 96 1.00069 157 1.01475 218 1.01429 279 0.99976 340 0.98596 97 1.00098 158 1.01488 219 1.01414 280 0.99947 341 0.98612 98 1.00127 159 1.01500 220 1.01399 281 0.99918 342 0.98628 99 1.00155 160 1.01513 221 1.01383 282 0.99890 343 0.98645 100 1.00184 161 1.01524 222 1.01367 283 0.99861 344 0.98662 101 1.00212 162 1.01536 223 1.01351 284 0.99832 345 0.98680 102 1.00240 163 1.01547 224 1.01334 285 0.99804 346 0.98698 103 1.00269 164 1.01557 225 1.01317 286 0.99775 347 0.98717 104 1.00297 165 1.01567 226 1.01299 287 0.99747 348 0.98735 105 1.00325 166 1.01577 227 1.01281 288 0.99718 349 0.98755 106 1.00353 167 1.01586 228 1.01263 289 0.99690 350 0.98774 107 1.00381 168 1.01595 229 1.01244 290 0.99662 351 0.98794 108 1.00409 169 1.01603 230 1.01225 291 0.99634 352 0.98814 109 1.00437 170 1.01610 231 1.01205 292 0.99605 353 0.98835 110 1.00464 171 1.01618 232 1.01186 293 0.99577 354 0.98856 111 1.00492 172 1.01625 233 1.01165 294 0.99550 355 0.98877 112 1.00519 173 1.01631 234 1.01145 295 0.99522 356 0.98899 113 1.00546 174 1.01637 235 1.01124 296 0.99494 357 0.98921 114 1.00573 175 1.01642 236 1.01103 297 0.99467 358 0.98944 115 1.00600 176 1.01647 237 1.01081 298 0.99440 359 0.98966 116 1.00626 177 1.01652 238 1.01060 299 0.99412 360 0.98989 117 1.00653 178 1.01656 239 1.01037 300 0.99385 361 0.99012 118 1.00679 179 1.01659 240 1.01015 301 0.99359 362 0.99036 119 1.00705 180 1.01662 241 1.00992 302 0.99332 363 0.99060 120 1.00731 181 1.01665 242 1.00969 303 0.99306 364 0.99084 121 1.00756 182 1.01667 243 1.00946 304 0.99279 365 0.99108 122 1.00781 183 1.01668 244 1.00922 305 0.99253 366 d 0.99228 0.99202 0.99177 0.99152 0.99127 0.99102 0.99078 0.99054 0.99030 0.99007 0.98983 0.98961 0.98938 0.98916 0.98894 0.98872 0.98851 0.98830 0.98809 0.98789 0.98769 0.98750 0.98731 0.98712 0.98694 0.98676 0.98658 0.98641 0.98624 0.98608 0.98592 0.98577 0.98562 0.98547 0.98533 0.98519 0.98506 0.98493 0.98481 0.98469 0.98457 0.98446 0.98436 0.98426 0.98416 0.98407 0.98399 0.98391 0.98383 0.98376 0.98370 0.98363 0.98358 0.98353 0.98348 0.98344 0.98340 0.98337 0.98335 0.98333 0.98331 http://landsathandbook.gsfc.nasa.gov/handbook/handbook_htmls/chapter11/chapter11.html SRDC Project DPI021 Final Report_Appendices.doc 3 Finding = Solar Zenith angle in degrees (converted to radians) This is the angle between the sun and the zenith or the vertical direction above any particular location. Visual representation of solar zenith and solar elevation angle. Equation 2: Solar Zenith angle = 90º – solar altitude (solar elevation) This value is in degrees and therefore needs to be converted into radians before being entered into the formula. Equation 3: Angle in Radians = (Angle in degrees * π / 180) The solar elevation can be located in the image metadata as shown below. Metadata from SPOT5, GeoEYE and Digital Globe defining sun angle elevation. SRDC Project DPI021 Final Report_Appendices.doc 4 If not available in the metadata file, it can be calculated using: the date, the time of day and the longitude and latitude of the centre pixel of the image. An on-line calculator can be found at: http://www.unitconversion.org/angle/degrees-to-radians-conversion.html) Finding = Mean solar exo-atmospheric irradiances (Esun λ) This is the solar irradiance for each spectral band measured at the top of the atmosphere before it is influenced by sun angles, atmospheric scattering etc. Each spectral band encompasses a range of wavelength irradiance values derived from detector sensitivity, signal to measurement transfer etc. To compensate for this, the wavelength irradiance values within a spectral band are weighted by a response function at each wavelength, summed and then averaged. A list of Esun λ values for a range of platforms are supplied below. Sensor Esun λ values for a range of satellite platforms and band widths. Finding = Spectral radiance at the sensor’s aperture (commonly referred to as ‘At-Sensor Radiance’) – calculated from the Digital Numbers of the raw image Each platform sensor has its own equation for calculating , as shown below: IKONOS Equation 4: Where: (Radiance) = SRDC Project DPI021 Final Report_Appendices.doc 5 CalCoef λ = Radiometric calibration coefficient [DN/(mW/cm2-sr)]. Bandwidth λ = Bandwidth of spectral band λ (nm). Radiance = equivalent irradiance at the input of the sensor [W/m2/µm/sr]. IKONOS radiometric calibration coefficients GeoEye-1 Equation 5: (Radiance) = (DN*Absolute calibration gain) +Absolute calibration Offset Where: Radiance = equivalent irradiance at the input of the sensor [W/m2/µm/sr]. Each image has its own unique ‘Gain’ and ‘Offset’ values and can be found in the po_#####_metadata.txt file of the image. Note that the Offset values were 0 for this image (refer below). GeoEye-1 radiance offset and gains. SPOT: Equation 6: (Radiance) = (DN/Absolute calibration gain) +Absolute calibration Offset Where: Radiance = equivalent irradiance at the input of the sensor [W/m2/µm/sr]. SRDC Project DPI021 Final Report_Appendices.doc 6 Each image has its own unique ‘Gain’ and ‘Offset’ values and can be found in the ‘metadata.dim’ file or associated PDF files (refer below). Note that the ‘Offset’ values were 0 for this image. SPOT radiance absolute calibration offset and gains. Technical Sheet References for further reading: o IKONOS: Taylor (2005) “IKONOS Planetary reflectance and mean solar exoatmospheric irradiance. o Quickbird: Krause (2005) “Radiometric Use of Quickbird Imagery” o Worldview 2: Updike and Comp (2010) “Radiometric Use of WorldView-2 Imagery” . o SPOT: “Technical Information”: http://www.spotimage.com/web/en/584-faq.php o Landsat 7: Landsat 7 Science Data Users Handbook: Chapter 11: http://landsathandbook.gsfc.nasa.gov/handbook/handbook_htmls/chapter11/chapter1 1.html o RapidEye: RapidEye (2011) “RapidEye Standard Image Product Specifications” : http://www.rapideye.de/upload/documents/PDF/RE_Product_Specifications_ENG.p d SRDC Project DPI021 Final Report_Appendices.doc 7 Tutorial 2: ENVI: Georectification of satellite imagery using an orthorectified base layer and derivation of a GNDVI image. Selecting ground control points (GCPs). Open the orthorectified base image as well as the new image to be warped using the main menu toolbar, select File- Open image. This will list both images in the “Available band List” window. Load the images into two different displays. To select GCPs select Map- Registration- Select GCPs: Image to Image from the ENVI toolbar From the ‘image to image registration’ window select the orthorectified image as the base image ‘Display 1’ and the image to be warped as the Warp image ‘Display 2’. Select OK to open the ‘Ground Control Points Selection’ window. SRDC Project DPI021 Final Report_Appendices.doc 8 Using the cross hair in the ‘Zoom’ displays identify pixels that represent the same target within each image. Press “Add Point” to record the GCPs. Try to avoid targets that may have changed over time such as the edge of a water bodies or crop boundaries. Road intersections or corners of roof structures are favourable. Once four GCPs have been selected, ENVI will provide a Root Mean Square (RMS) error so that the accuracy of the existing GCPs as well as those subsequently added can be determined. The ability to predict the location of where the next GCP will be in the Warp Image will also be activated. Select a target in the Base Image, press ‘Predict’, and the Warp Image will display a corresponding location. The cross hairs can then be moved to the required pixel. The list of GCPs can be saved and then restored for later use by using the ‘Ground Control Points Selection’ window. It is recommended that at least 80 GCP’s are selected to warp an image, and that RMS error is as close to 0 as possible, particularly when warping high resolution images such as IKONOS (3.2m). SRDC Project DPI021 Final Report_Appendices.doc 9 Image warping. Once GCPs have been entered, select ‘Options’- ‘Warp File’ from the ‘Ground Control Points Selection’ window (above). Select the image too be warped (generally this will the image that is opened), Select ‘Registration Parameters’ window- output name- ‘OK’. The accuracy of the georectification can be assessed by overlaying the orthorecitifed image or an accurate vector layer i.e. crop boundaries, using software such as ArcGIS, to see how they align. Some misalignment between vector data and imagery Alignment after georectification SRDC Project DPI021 Final Report_Appendices.doc 10 Producing a GNDVI image. Under the ‘Transform’ option in the main ENVI title bar select NDVI (above), this will open the ‘NDVI Calculation Input File’ window (below). Select the name of the image to be transformed- then ‘OK’. This will open the ‘NDVI Calculation Parameters’ window- select the ‘Input File Type’; the corresponding bands for Green (instead of Red) and NIR in ‘NDVI Bands’ and ‘Output data type’ as shown. Select an output file location and name (eg. QB09_17march09_Kumbia_GNDVI) this identifies the satellite used, acquisition date, location and vegetation index used. On the completion of the index transformation the new GNDVI file will automatically appear in the ‘Available Band List’ window. SRDC Project DPI021 Final Report_Appendices.doc 11 Tutorial 3: ArcGIS: Conversion of Mapinfo (.TAB) files into ArcGIS (.SHP) files. This section is relevant to ArcGIS users whom need to open mill data created in MapInfo as a .TAB file. The ArcGIS user can either purchase a copy of MapInfo, or download the free spatial software package “FW Tools” (http://fwtools.maptools.org/). In the C:\Program Files\FWTools directory create a new folder named 'Mapinfo2Shp' and unzip the MapInfo file; This includes four extensions: .DAT, .ID, .MAP and .TAB Open the ‘FWTools Shell’ command window. Copy and paste the following command into the window, inserting the required file names (indicated in red: the first file is the output shape file, while the second is the original TAB file) and press ‘Enter’ ogr2ogr -f "ESRI Shapefile" Mapinfo2Shp\FILENAME.shp Mapinfo2Shp\FILENAME.tab The ESRI shape file set will appear in the selected directory and available for input into an Arc project. (Note: ensure there are no spaces in either the TAB or Shp file name as this will prevent the conversion. Although effective, the output .SHP file from FW Tools can sometimes display a projection error where files are misaligned by approximately 200m. If this error does occur, the projection of the newly created shape file needs to be defined and then reprojected. Import the SHP file into ArcMAP- within the Arctoolbox select ‘Data management tools’‘Projections and Transformations’- ‘Define Projection’, select the .SHP file. Select the coordinate system icon in the ‘Define Projection’ window- ‘Select’ in the ‘Spatial Reference Properties’ window- then in the ‘Browse for Coordinate System’- select ‘Projected Coordinate Systems’- ‘National Grids’- ‘Australia’- ‘AGD 1984 AMG Zone (and the appropriate zone i.e. 55 or 56, refer below)’- name the output file and then ‘OK’. SRDC Project DPI021 Final Report_Appendices.doc 12 Grid zones of Australia (http://www.environment.gov.au/ssd/publications/ir/pubs/ir473.pdf) With the projection of the shape file now defined it needs to be reprojected into a common coordinate system. In ‘Arctoolbox’, select ‘Data management tools’- ‘Projections and Transformations’- ‘Feature’- ‘Project’, input shape file, define the name and location of the output file, and select output coordinate system. ‘Select’ in the ‘Spatial Reference Properties’ window- then in the ‘Browse for Coordinate System’- select ‘Projected Coordinate Systems’‘National Grids’- ‘Australia’- GDA 1994 MGA and appropriate zone. The out put file will now align with the georectified imagery. Tutorial 4: ArcGIS: Buffering of polygons and removal of those affected by cloud before the extraction of spectral data. Within ArcGIS open both the georectified image (.TIF format) and mill vector file (.SHP). As seen below some crop boundaries extend beyond the image coverage area and therefore they require removal prior to additional analysis. This is achieved by creating a new shape that denotes the extent of the image. Then after drawing convert the graphic into a shapefile. SRDC Project DPI021 Final Report_Appendices.doc 13 Create a new file name that denotes the removal of crop boundaries outside the extent of the image. For the removal of clouds repeat the process defined above, with a polygon defining the area of cloud cover rather than the extent of the satellite image. Buffering of crop boundaries. Even with accurately georeferenced imagery and vector layers, some crop boundaries can include spectral information that is not specific to cane such as head lands, roads etc. To eliminate this ‘spectral contamination’ an internal buffer is applied to the boundary of each crop. For SPOT 5 imagery an internal buffer of 20m (2 pixels) was applied whilst for higher resolution imagery a 10m buffer was used. SRDC Project DPI021 Final Report_Appendices.doc 14 Crop boundary includes non cane related pixels Imagery overlayed with a crop boundary vector layer (from Mill data) Crop boundaries buffered by 10m (IKONOS imagery) The ArcGIS software The internal buffering was undertaken achieved by A negative effect of buffering process particularly for the SPOT 5 imagery is the loss of spectral information from small crops or small sub-blocks of differing verities. In some cases the buffering removes all pixel information, ultimately negating that block from all further analysis. SRDC Project DPI021 Final Report_Appendices.doc 15 Tutorial 5: Starspan GUI: Extracting average spectral values and associated attribute information for multiple blocks. In order to identify the correlation between average GNDVI value and average block yields (as supplied in mill GIS layers) a rapid method for extracting imagery data for every block and aligning it with mill attribute data was required. The freeware program Starspan GUI was identified to be highly effective for completing this task. Datasets required: - Ortho-rectified remotely sensed imagery that includes separate spectral bands that allow the calculation of GNDVI, or alternatively a layer stacked multispectral image that includes a derived GNDVI layer. - Accurate vector file outlining crop boundaries as well as associated attribute information such as crop variety, class, yield, CCS etc. Download StarSpan from: https://projects.atlas.ca.gov/frs/download.php/581/install-starspan-win32-0.2.0.jar Within the StarSpan ‘Inputs’ screen add the ‘Vector file’ (mill GIS layer) and ‘Raster file’ (image). To extract GNDVI values: Within the ‘Options’ screen- Change pixel proportion to 0.9 With in the ‘Commands’ screen - select ‘avg’ as the output statistic and name the output file a name and directory. Select ‘Execute’ to complete task. SRDC Project DPI021 Final Report_Appendices.doc 16 The output file that includes all attribute information from the vector file and spectral data extracted for each corresponding block will be in a .CSV that can opened in Microsoft Excel. SRDC Project DPI021 Final Report_Appendices.doc 17 Tutorial 6: ENVI: Producing a classified vegetation index map of a cane crop from a 4 band satellite image. With the required multispectral image opened in ENVI, from the main toolbar select ‘Basic Tools’- ‘Region of Interest’- ‘ROI Tool’ and then outline the required crop (below). From the ‘ROI Tool’ - ‘File’ menu select ‘Subset Data via ROIs’. This will activate the ‘Select input File to Subset via ROI’ window. Select all four spectral bands from the ‘File Spectral Subset’ window, then ‘OK’. From the ‘Spatial Subset via ROI Parameters’ window (below left) select the ROI, ‘Yes’ to ‘Mask pixels outside of ROI’, ‘0.0000’ as the ‘Mask Background Value’, name the output file and then ‘OK’. Open the output file within a new display (below right). To apply a vegetation index to the 4 band satellite image of the crop, select ‘Transform’‘NDVI’ from the main ENVI Toolbar (below left). Select the input image from the ‘NDVI Calculation Input File’ window and ‘OK’ to activate the ‘NDVI Calculation Parameters’ window (below right). Select the sensor type, in this example ‘SPOT’ and the appropriate bands. Note: although this method is for NDVI, a GNDVI index can be applied by selecting SRDC Project DPI021 Final Report_Appendices.doc 18 the corresponding band (1) for the ‘Green’ bandwidth rather than (2) for the ‘Red’. Select ‘Output data Type’ as ‘Floating Point’ and then ‘OK’ Open the indexed image as a new display (below left). From the main ENVI Toolbar select ‘Classification’- ‘Unsupervised’ – ‘Isodata’ (Below right). This will open the ‘Classification Input File’, select the image to be classified as well as the mask band, which can be any band from the original image, select ‘OK’. From the ‘ISODATA Parameters’ window (Below) select the ‘Number of Classes’ to ‘Min’ 5 and ‘Max’ 5, ‘Maximum Iterations’ to 100 and name the output file, select ‘OK’. SRDC Project DPI021 Final Report_Appendices.doc 19 Open the classified image as a new display. Using the classified image display toolbar, select ‘Tools’- ‘Color Mapping’ – which will open a ‘Class Color Mapping’ window. Using this window change the colour scheme to that of the SPAA standard (as below), then File save changes. The classified image can then be saved as a TIF file and imported into any GIS software. SRDC Project DPI021 Final Report_Appendices.doc 20 Tutorial 7: ENVI: Extracting point source spectral information from imagery using regions of interest (ROI’s). With the required multispectral image opened in ENVI, from the main toolbar select ‘Basic Tools’- ‘Region of Interest’- ‘ROI Tool’. Use the ‘cursor location/ value’ window (above right) from the main image tool bar- ‘Tools’ option, to locate the sampling point coordinates within the image. These points can also be directly overlayed on to image if stored as a shape file. Once the points have been located, select the ‘Zoom’ option in the ‘ROI Tool’ window (below left) and then in the zoom window (below right) manually draw the ROI around the sampling coordinate. We suggest a 3*3 pixel array for IKONOS and 2*1 pixels for SPOT5. After each ROI is drawn select ‘New Region’ in the ‘ROI Tool’ window. ‘Zoom’ option ‘New Region’ option ‘Stats’ option Once all ROI’s have been drawn, the mean spectral values for each ROI can be viewed by selecting the ‘Stats’ option in the ‘ROI Tool’ window. Record these values in Excel for further analysis. SRDC Project DPI021 Final Report_Appendices.doc 21 Tutorial 8: ENVI: Converting VI pixel values into yield (TCH) using an exponential linear algorithm. For this example a GNDVI image developed through the process described in Tutorial 6 will be used. From the main ENVI Toolbar open the indexed single band image as a new display (below left). From the ‘Basic Tools’ option select ‘Band Math’, which will open a ‘Band Math window’ (below right). For this example, enter the exponential equation developed from the correlation between TCH and GNDVI. Note the use of ‘float’ in the equation, this ensures the 8- bit GNDVI values are seen as a floating point value and therefore are not rounded up to the nearest whole number. Select ‘OK’. Select the GNDVI band from the ‘Variables to Band Pairings’ window name the output yield file and select ‘OK’ (Below left). To identify if the conversion has worked correctly open the derived yield map and double click the cursor on the main image, the values shown as ‘Data’ should correspond with expected yield values. SRDC Project DPI021 Final Report_Appendices.doc 22 Tutorial 9: ENVI: Creating Google Earth KMZ files from Geotiffs. With the required 3 band Geotiff opened in ENVI, select ‘Spectral’- ‘SPEAR Tools’- ‘Google Earth Bridge’ from the main ENVI menu bar (refer below). Within the Google Earth bridge window select ‘Add Files’ and locate the required files from the drop down list and select ‘Next’. A warning (as below) may be displayed. Select ‘OK’ as this does not influence the conversion process. Google Earth allows images to be displayed in a time series but for this exercise it is unnecessary. SRDC Project DPI021 Final Report_Appendices.doc 23 From the following window (as below), select ‘Thumbnails’ from the ‘Output Type’; ‘PNG’ as the ‘Output Format’; ‘1MB’ as the ‘Image Size’; and ‘R3’ ‘G2’ and ‘B1’ as ‘Color’. Then select ‘Next’. From the following window, select the box for ‘Do not export any vectors’, then ‘Next’. From the subsequent window select ‘Select Output File’ , name the KML file and save it to the desired directory. Select ‘Open in Google Earth when done’ box if you want to view the file immediately after it is created. Click ‘Next’ to create the image. Once opened in Google Earth, right mouse click on the ‘Thumbnails’ legend option of the image (refer to image below) and select ‘Save place As’. The ‘Save file’ window that opens will allow the KML file to be saved as a KMZ. SRDC Project DPI021 Final Report_Appendices.doc 24 SRDC Project DPI021 Final Report_Appendices.doc 25 11. Appendix 2: Media/ Publications 2012 ASSCT presentation. AG 24. DEVELOPING SUGAR CANE YIELD PREDICTION ALGORITHMS FROM SATELLITE IMAGERY By ANDREW ROBSON1, CHRIS ABBOTT1, DAVID LAMB2, ROB BRAMLEY3 1Department of Employment Economic Development and Innovation, Kingaroy. 2University of New England, Armidale 3CSIRO, Adelaide, SA andrew.robson@deedi.qld.gov.au KEYWORDS: Yield forecasting, satellite imagery, SPOT5, GNDVI Abstract The research presented in this paper discusses the accuracies of remote sensing and GIS as yield prediction tools at both a regional and crop scale over three Australian cane growing regions; Bundaberg, Burdekin and the Herbert. For the Burdekin region, the prediction of total tonnes of cane per hectare (TCH) produced from 4999 crops during the 2011 season was 99% using an algorithm derived from 2010 imagery (green normalised difference vegetation index) and average yield (TCH) data extracted from 4573 crops. Similar accuracies were produced for the Bundaberg region during 2010 (95.5% from 3544 blocks) and 2011 (91.5% for 3824 crops) using a Bundaberg specific algorithm derived from 2008/2010 imagery and yield data. The Bundaberg algorithm was also accurate in predicting yield at specific in-crop locations (91.5% accuracy; SE = 0.028). 1. Introduction Accurate in-season predictions of regional yield are of vital importance for formulating harvesting, milling and forward selling decisions, whilst at a block scale, they provide growers with an understanding of both in-crop variability and total production. Currently, annual cane production estimates are made by quantifying the area of cane grown within a region by visual in-season yield assessments. Although this method can produce accuracies of up to 95% (Pitt pers. comm. 2011) it can be influenced by variable climatic conditions such as those experienced in 2010. As such geographic information systems (GIS) and remote sensing (RS) may offer an additional tool for validating these predictions as well as potentially provide a more accurate seasonally sensitive method of prediction. 1.1 GIS and Remote Sensing in the Sugar Industry Geographic information systems (GIS) have been widely adopted by the Australian sugar industry as an essential tool for the recording and managing spatial data (Davis et al. 2007). One such system developed for the Mackay and Burdekin region has greatly increased the integration of mill and productivity datasets, thus enabling greater efficiencies in data retrieval and analysis of client information (Markley et al. 2008). Similarly, the development of a whole-of-community GIS system by the Herbert River sugar district has created the capacity to record real-time cane harvester operations via GPS, enabling improvements in the coordination and planning of the cane harvest, efficient reporting of harvest performance and the identification of consignment errors. This SRDC Project DPI021 Final Report_Appendices.doc 26 information has also been used to improve rail transport infrastructure safety and efficiency (De Lai et al. 2011). Globally, satellite imagery has been identified as an effective tool for predicting sugar cane yield (Fernandes et al. 2011; Benvenuti and Weill 2010; Bégué et al. 2010; Simões et al. 2009; Abdel-Rahman and Ahmed 2008; Bégué et al. 2008; Almeida et al 2006; Simões et al. 2005; Krishna Rao et al. 2002; and Rudorff and Batista 1990), although such research has been limited in Australia (Noonan. 1999; Markley et al. 2003; Robson et al. 2011; Robson et al. 2010; Lee-Lovick and Kirchner 1991). For the last decade, Mackay Sugar Ltd has been the predominant adopter of satellite imagery as a commercial yield forecasting tool for the Mackay region, utilising yield prediction algorithms derived from SPOT imagery (Markley et al. 2003). The research presented in this paper investigates the development and validation of similar algorithms over three additional Australian growing regions including Bundaberg, Burdekin and Herbert. 1.2 Yield Predictions using Remote Sensing Techniques The amount of electro-magnetic radiation (EMR) reflected from a sugarcane canopy is positively correlated to the leaf area index (LAI), which in turn may correspond to the amount of biomass within the crop, and therefore yield (Bégué et al. 2010). However, this relationship can be influenced by variations in canopy architecture, foliar chemistry, agronomic parameters and sensor and atmospheric conditions (Abdel-Rahman and Ahmed 2008). More specifically, variety, crop class (plant or ratoon), date of crop planting or ratooning, duration of harvest period and environmental variability are all factors that have been shown to influence the accuracies of yield prediction algorithms developed from remotely sensed imagery (Zhou et al 2003; Singels et al. 2005; InmanBamber 1994). In an attempt to remove influences such as spectral interference or ‘noise’, previous researchers have investigated a number of vegetation indices. The most commonly used Normalised Difference Vegetation Index (NDVI), does address some measurement errors associated with atmospheric attenuation and shading, however it can saturate in large biomass crops such as sugar cane with a LAI greater than 3 (Benvenuti and Weill 2010; Bégué et al. 2010; Xiao 2005; Xiao et al. 2004b; Xiao et al. 2004a; Huete et al. 2002; Huete et al. 1997). To reduce the effects of saturation, a number of additional indices have been employed including the Green Normalised Difference Vegetation Index (GNDVI) (Gitelson et al. 1996; Benvenuti and Weill 2010). Timing of image capture has also been identified to be an important consideration when predicting cane yield, especially when compared to the growth phase of the crop. Sugar cane undergoes three distinct growth phases including germination or establishment and tillering, vegetative development or stalk growth and stabilisation, senescence or maturation (Bégué et al. 2010; Simões et al. 2005; Fernandes et al. 2011; Krischna Rao et al. 2002). During the vegetative growth stage NDVI can increase from 0.15 to 0.7, before remaining relatively stable (if unstressed) during the maturation phase, until harvest (Bégué et al. 2010). Almeida et al. (2006) identified this time period to be 3-6 months prior to harvest, whilst Simões et al. (2005) suggested 240 days after planting or ratooning. As well as a stabilisation period of NDVI, a ‘synchronisation’ of NDVI was also observed across various plant and ratoon ages due to climatic factors such as rain and temperature. This synchronisation and stabilisation of NDVI is important as it indicates that there is likely to be an extended window of image capture where variability in the canopies spectral response as well as differences across crops is minimalised (Bégué et al. 2010; Almeida et al. 2006; Krischna Rao et al. 2002; Rudorff and Batista 1990). SRDC Project DPI021 Final Report_Appendices.doc 27 2. Methodology 2.1 Study Districts Research was conducted in three climatically distinct Queensland cane growing regions of the Herbert (2107mm of rainfall annually), the Burdekin (1005mm) and Bundaberg (930mm) during the 2010 and 2011 growing seasons. 2.2 Satellite Imagery and Spatial Data During the 2010 and 2011 cane growing seasons, full scene (3600 km2) SPOT 5 satellite images were captured over the Herbert (2 June 2011); the Bundaberg region (10 May 2010 and 27 March 2011); and over the Burdekin region (14 May 2010 and 22 April 2011). The spectral resolution of SPOT5 imagery is Green (0.5-0.59µm), Red (0.61-0.68µm), Near Infrared (0.78-0.89µm) and Shortwave Infrared (1.58-1.75µm), with a spatial resolution of 10 metre pixels. All SPOT5 imagery used for this research was corrected for top of atmosphere reflectance (TOA) (SPOT Image, 2008) and orthorectified to a corrected base layer. Block boundary GIS vector layers detailing attribute tables of agronomic data including; variety, class, total area harvested and tonnes cane harvested were sourced from either milling or productivity services within each region. 2.3 Extraction of Spectral Information For all cane blocks within the extent of each SPOT 5 image (Figure 1) spectral information was extracted using the open source software Starspan GUI (Rueda et al. 2005). A 20 m metre buffer was applied to each paddock boundary to ensure the extracted information did not include non canespecific pixels. Spectral and agronomic information including mill data was exported to a single text file to enable additional analysis. a. b. c. SRDC Project DPI021 Final Report_Appendices.doc 28 d. Figure 1: SPOT5 images captured over each growing region (a). Burdekin, (b). Herbert and (c). Bundaberg. (d). closer view of agronomic information provided within the GIS attribute table. 2.4 Vegetation Indices To identify the best correlations between satellite imagery and crop yield (TCH), a number of vegetation indices were examined including the Normalised Difference Vegetation Index (NDVI), Green Normalised Difference Vegetation Index (GNDVI) (equation 1), The Soil Adjusted Vegetation Index (SAVI) and the Two-band Enhance Vegetation Index (EVI_2). These indices were calculated for every cane block defined by a GIS paddock boundary within each image capture area. Using harvested tonnes of cane per hectare (TCH) supplied by the respective mills, the index that provided the highest correlation coefficient were identified. For all regions the GIS attribute data was used to separate the spectral information on the basis of variety, crop class (plant or ratoons) and age of crop, in an attempt to improve the correlations. GNDVI = (ΡNIR – PGREEN) / ( PNIR + PGREEN) (1) Where PGREEN, and ΡNIR are the TOA reflectance values measured in the green and near infrared spectral bands. Additionally, predictions of average yield were made for 3544 (2010) and 3,824 (2011) cane crops within the Bundaberg region using an algorithm derived from the linear relationship between 2008 and 2010 crop yield and corresponding SPOT5 data (Robson et al. 2011) (equation 2). The accuracy of this algorithm was also evaluated against point source locations within a single crop and validated within field measurements. Sampling coincided with the commercial harvest of the crop and consisted of 5m linear cane rows hand cut at replicated locations representing high, medium and low GNDVI values, located with a non- differential GPS unit. GNDVI yield prediction y = 3.1528 *EXP(5.6973 * x) (2) algorithm (Bundaberg) Where y = predicted average yield (TCH) and x = average GNDVI value extracted from TOA SPOT5 image. (n= 150 crops) A similar prediction was also undertaken for 4999 cane crops grown within the Burdekin region (2011 season) using an algorithm derived from the correlation between 2010 Burdekin crop yields and corresponding 2010 imagery (equation 3). GNDVI yield prediction y = 12.691 *EXP(3.8928 * x) (3) algorithm (Burdekin) Where y = predicted average yield (TCH) and x = average GNDVI value extracted from TOA SPOT5 image. (n= 4573 crops) 3. Results SRDC Project DPI021 Final Report_Appendices.doc 29 The initial aim of this research was to develop a generic image-based yield algorithm for all Queensland growing regions that was non-specific to variety, growth stage, and even seasonal variability. However, it was quickly identified that one algorithm would be insufficient due to the large range of varieties planted as well as variation in growing and climate conditions across each region. As such, each growing region was evaluated separately. 3.1 Bundaberg The correlation between TCH and spectral data extracted for 3824 cane crops grown within the Bundaberg region during 2011 (including 26 varieties with nine ratoon stages, plant, replant and standover classes) was promising with all vegetation indices producing correlation coefficients above 0.6, with GNDVI producing the highest (r = 0.63) (Table 1). This correlation was further improved by segregating the data into plant and ratoon classes. Table 1: Correlation coefficients (r) identified between TCH and individual spectral bands/vegetation indices for the Bundaberg district, 2011 growing season. Band/VI All Blocks Plant Cane 1st Ratoon Bundaberg District 2nd Ratoon 3rd Ratoon Variety Q208 Plant 1st Rat Variety KQ228 Plant 1st Rat Green 0.20 0.23 0.17 0.15 0.16 0.51 0.50 0.41 0.50 Red 0.42 0.40 0.44 0.47 0.48 0.58 0.55 0.45 0.58 NIR 0.58 0.70 0.62 0.60 0.57 0.64 0.60 0.74 0.63 SWIR 0.37 0.28 0.37 0.38 0.38 0.45 0.44 0.44 0.47 NDVI 0.61 0.67 0.68 0.69 0.66 0.68 0.65 0.66 0.66 GNDVI 0.63 0.71 0.71 0.70 0.66 0.68 0.68 0.72 0.70 SAVI 0.61 0.71 0.66 0.65 0.62 0.68 0.64 0.73 0.66 EVI_2 0.61 0.72 0.66 0.65 0.62 0.68 0.64 0.74 0.66 The stability of correlation across varieties for both GNDVI and NDVI is important as it indicates that a ‘generic’ algorithm which is not cultivar specific may be possible for the Bundaberg region, a finding that supports initial results presented by Robson et al. (2011). To further investigate the consistency of GNDVI values across varying classes, variety and seasons, the 2011 data (n= 3824) were overlayed with similar data used to develop the 2008/ 2010 algorithm (n= 150)(Figure 2). From Figure 2 it can be seen that although there is variance around the line of best fit, the overall trend between GNDVI and TCH is relatively consistent across the two data sets. 250 y = 4.0566e5.0316x 200 2011 2008&2010 150 Expon. (2011) TCH 100 50 0 0.3 0.4 0.5 0.6 GNDVI 0.7 Figure 2. Correlation between GNDVI (SPOT 5) and TCH from Bundaberg cane blocks during the 2008/2010 (black points) and 2011 (grey points) seasons. SRDC Project DPI021 Final Report_Appendices.doc 30 The calculation and then subsequent substitution of average GNDVI value (0.567) from 3544 crops grown during 2010 into the 2008/ 2010 algorithm produced an estimated average yield of 78.1 TCH, highly comparable to the actual milled yield of 81.8 TCH (95.5% accurate). For the 2011 harvest season, an average yield of 80.1 TCH was predicted following the substitution of average GNDVI value (0.57) sourced from 3824 crops into 2008/2010 algorithm. This prediction was within 9% of actual milled harvest yield of 73.3 TCH (91.5%). The accuracy of overall prediction, and the fact that the data was not segregated into variety or growth stage, indicates that this technology has the potential to predict regional cane yield within the Bundaberg growing region, a result that differs from previous findings by Lee-Lovick and Kirchner (1991). To coincide with regional forecasting, the development of such an algorithm offers the potential for predicting individual crop yield as well as the derivation of surrogate yield maps, prior to harvest. To test this, the accuracy of the GNDVI yield algorithm was also evaluated over point source locations within an individual Bundaberg cane crop (area 18.7 ha, var. KQ228) harvested 25 July 2011 (Figure 3). This analysis identified a strong relationship between predicted yield from a SPOT5 image captured on the 27 March 2011, and final yield measured on the 25 July 2011. 120 y = 0.915x S.E. = 0.028 90 60 b. 30 40 60 80 100 120 a. Predicted TCH Measured TCH c. Figure 3. (a). False colour image of Bundaberg cane crop (area 18.7 ha, var. KQ228) harvested 25 July 2011, with yellow markers indicating field sampling locations. (b). measured verse predicted cane yield at the locations identified in (a). (c). Classified yield map generated by applying the 2008/2010 GNDVI yield algorithm to the SPOT5 (27 March 2011) pixel values. The generation of a classified yield map (Figure 3c) and subsequent accurate prediction of total crop yield from the average crop GNDVI value (predicted of 92 TCH, actual delivered yield of 88.7 TCH) further supports the potential of this technology for producing in-season yield variability maps. 3.2 Burdekin For the Burdekin region, correlation coefficients produced between TCH and SPOT5 derived vegetation indices (captured 14 May 2010) for all 4573 crops were relatively consistent ranging from r=0.39 (NDVI) to r=0.44 (SAVI and EVI_2) (Table 2). This correlation remained relatively SRDC Project DPI021 Final Report_Appendices.doc 31 unchanged when data was segregated into the different cultivars Q208 and KQ228, indicating that a yield prediction algorithm for this region may not be required to be cultivar specific. Table 2: Correlation coefficients (r) identified between TCH and individual spectral bands/vegetation indices for the Burdekin district. Band/VI All Blocks Plant Cane 1st Ratoon Burdekin District 2nd Ratoon 3rd Ratoon Variety Q208 Plant 1st Rat Variety KQ228 Plant 1st Rat Green 0.12 0.08 0.11 0.10 0.18 0.20 0.16 0.02 0.04 Red 0.18 0.19 0.15 0.17 0.11 0.27 0.23 0.13 0.12 NIR 0.41 0.40 0.39 0.28 0.20 0.36 0.36 0.50 0.42 SWIR NDVI GNDVI 0.19 0.18 0.12 0.11 0.06 0.28 0.10 0.07 0.06 0.39 0.42 0.35 0.29 0.21 0.39 0.35 0.41 0.31 0.43 0.44 0.40 0.29 0.21 0.41 0.35 0.45 0.35 SAVI 0.44 0.44 0.41 0.31 0.23 0.39 0.37 0.50 0.40 EVI_2 0.44 0.43 0.41 0.31 0.23 0.39 0.37 0.50 0.40 Unlike the Bundaberg analysis however, there was a noticeable drop in the correlation coefficients with ratoon age, especially 2nd and 3rd ratoon (Table 2). This variation indicates that an algorithm that is not crop class specific may be inaccurate, especially when predicting point source yield within individual crops such as that displayed in Figure 3. At a regional level the predicted average crop yield of 4999 crops grown during the 2011 season using the 2010 algorithm (equation 3) was 99% (actual average yield of 120 TCH, predicted 118.8 TCH). Although highly accurate, the result is not considered robust, due to the large spread of data (r2 = 0.07) produced particularly with standover crops (grey markers in Figure 4). This predictive accuracy will however be further validated during the 2011/2012 season. 350 300 250 2011 with standover 2011 without standover Expon. (2011 with standover) y = 32.325e2.1824x R2 = 0.0696 200 TCH 150 100 50 0 0.4 0.5 0.6 GNDVI 0.7 Figure 4. Correlation between GNDVI (SPOT 5) and TCH from 2011 Burdekin cane blocks with black points indicating non-standover crops whilst grey points indicating standover. 3.2 Herbert The initial correlation between TCH and spectral data (SPOT 5 captured 2 June 2011) for 8596 cane crops grown in the Herbert region (including 53 varieties, multiple ratoon stages, plant, replant and standover) was poor (Table 3). This result is believed to be attributed to severe climatic conditions experienced towards the end of 2010 and start of 2011. The Herbert region had around 25% of the 2011 crop as ‘stand over’ i.e. not harvested from 2010, with the remainder exhibiting SRDC Project DPI021 Final Report_Appendices.doc 32 various degrees of flood related damage. The removal of standover blocks did improve the coefficients. The highest regression coefficients were identified by segregating the data into crop class and then variety, for example KQ228 plant crop r=0.65 (GNDVI). Table 3: Correlation coefficients (r) identified between TCH and individual spectral bands/vegetation indices for the Herbert district. Band/VI All Standover Blocks removed Plant Cane Herbert District 1st 2nd 3rd Ratoon Ratoon Ratoon Variety Q208 Variety KQ228 Plant 1st Rat Plant 1st Rat Variety Q200 Plant 1st Rat Green 0.01 0.03 0.01 0.15 0.15 0.35 0.03 0.13 0.01 0.31 0.06 0.02 Red 0.10 0.09 0.10 0.05 0.05 0.16 0.17 0.17 0.07 0.20 0.17 0.09 NIR 0.23 0.46 0.54 0.45 0.40 0.46 0.65 0.55 0.59 0.49 0.50 0.39 SWIR 0.38 0.42 0.37 0.35 0.05 0.46 0.53 0.48 0.30 0.29 0.35 0.33 NDVI 0.22 0.40 0.47 0.34 0.31 0.37 0.60 0.16 0.47 0.36 0.50 0.33 GNDVI 0.23 0.45 0.54 0.42 0.42 0.44 0.65 0.55 0.58 0.45 0.50 0.37 SAVI 0.23 0.45 0.54 0.42 0.37 0.46 0.65 0.56 0.57 0.45 0.51 0.37 EVI_2 0.24 0.46 0.54 0.42 0.37 0.46 0.65 0.56 0.58 0.46 0.51 0.37 These results indicate that for the accurate prediction of yield within the Herbert region a number of algorithms representing different growth stages and even varieties may be required. This hypothesis requires further validation over subsequent growing seasons, particularly seasons that are not influenced by extreme climatic conditions. Discussion The undertaking of this research over the three distinct growing regions was highly beneficial considering the array of success identified. Results from the Bundaberg region, and to a lesser extent the Burdekin, indicated that a ‘generic’ yield prediction algorithm may be developed and then used to accurately predict regional production and even within crop yield variability. Although improved correlations were produced following the segregation of data into different groups such as crop class (Burdekin) and variety (Herbert) some consideration has to be made on the number of algorithms developed. In regards to variety, fifty-three were planted in the Herbert, twenty-six in Bundaberg and nineteen in the Burdekin in the years encompassed by this study. If other variables such as the segregation of regions into smaller climate driven micro regions or crop class are also accounted for then the number of algorithms required would grow substantially. One method to address this may be to develop algorithms for only the dominant varieties. For example, only three varieties (of nineteen) in the Burdekin accounted for 83% of the total number of planted blocks. Alternatively, varieties could be categorised into groups based on their spectral signatures. The use of multiple algorithms may increase the flexibility of the predictive models for the season upon which it is applied, allowing it to better compensate for changing percentages of varieties and classes throughout a district and the addition of new varieties. In the past, the adoption of remote sensing as a yield prediction tool by the Australian Sugar industry has been severely hampered by a number of limitations including: a lack of yield data from the mills due to privacy issues, an extended harvesting period resulting in a patchwork of different varieties, growth and ratoon stages in close proximity and, seasonal or climatic variability, constant cloud clover, insufficient computational demands for image processing, a shortage of skilled analysts and concerns regarding the benefit-cost of adopting the technology. Irrespective of these concerns the research presented in this paper identified satellite imagery and associated GIS data as useful tools for supporting current methods of yield forecasting, with the potential of improving both regional and incrop yield predictions in the future following further validation. Acknowledgements. SRDC Project DPI021 Final Report_Appendices.doc 33 The authors would like to acknowledge SRDC for providing funding for this research as well as those growers and industry partners whom have collaborated, particularly Bundaberg Sugar Pty Ltd, Herbert Resource Information Centre, Sucrogen and Farmacist Pty Ltd. References: Abdel-Rahman, EM and Ahmed, FB (2008). The application of remote sensing techniques to sugarcane (Saccharum spp Hybrid) production: a review of the literature. International Journal of Remote Sensing 29(13), 3753- 3767. Almeida, TIR, Filho, CR De Souza and RossettoR (2006). ASTER and Landsat ETM+ images applied to sugarcane yield forecast. International Journal of Remote Sensing 27. pp 4057-4069. Bégué, A, Lebourgeois, V, Bappel, E, Todoroff, P, Pellegrino, A, Baillarin, F, and Siegmund, B (2010). Spatio-temporal variability of sugarcane fields and recommendations for yield forecasting using NDVI. International Journal of Remote Sensing. 31(20).5391-5407. Bégué, A, Todoroff, P and Pater, J (2008). Multi-time scale analysis of sugarcane within-field variability: improved crop diagnosis using satellite time series? Precision Agriculture 9.161-171. Benvenuti, F and Weill, M (2010). Relationship between multi-spectral data and sugarcane crop yield. Proceedings of the 19th world Congress of Soil Science Soil Solutions for a Changing World 1-6 August 2010 Brisbane Australia.33-36. Davis, R, Bartels, R and Schmidt, E (2007). Precision Agriculture Technologies- Relevance and application to sugarcane production. In SRDC Technical Report 3/2007: Precision agricultural options for the Australian sugarcane industry Eds R Bruce.60-117. De Lai, R, Packer, G, Sefton, M, Kerkwyk, R and Wood, AW (2011). The Herbert Information Portal: Delivering real-time spatial information to The Herbert River sugar community. Proceedings of The Australian Society of Sugar Cane Technologists 33. Fernandes, JL, Rocha, JV and Lamparelli, RAC (2011). Sugarcane yield estimates using time series analysis of spot vegetation images. Science in Agriculture 68(2).139-146. Huete, AR, Liu, HQ, Batchily, K and Leeuwen, W (1997). A Comparison of vegetation indices over a global set of TM images for EOS-MODIS. Remote Sensing of Environment 59.440-451. Huete, A, Didan, K, Miura, T, Rodriguez, EP, Gao, X and Ferreira, LG (2002). Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sensing of Environment 83.195-213. Inman-Bamber, NG (1994). Temperature and seasonal effects on canopy development and light interception of sugarcane. Field Crops Research 36.41-51. Kishna Rao, PV, Venkateswara Rao, V, Venkataratnam, L (2002). Remote Sensing: A technology for assessment of sugarcane crop acreage and yield. Sugar Tech 4 (3&4). 97-101. Lee-Lovick, G and Kirchner, L (1991). Limitations of Landsat TM data in monitoring growth and predicting yields in sugarcane. Proceedings of Australian Society of Sugar Cane Technologists 13.124-129. SRDC Project DPI021 Final Report_Appendices.doc 34 Markley, J Ashburner, B and Beech, M (2008). The development of a spatial recording and reporting system for productivity service providers. Proceedings of Australian Society of Sugar Cane Technologists 30.10-16. Markley, J, Raines, A and Crossley, R (2003). The development and integration of remote sensing GIS and data processing tools for effective harvest management. Proceedings of Australian Society of Sugar Cane Technologists 25.2003. Noonan, MJ (1999). Classification of fallow and yields using Landsat TM data in the sugarcane lands of the Herbert River Catchment. Herbert Resource Information Centre Qld Website link: http://wwwhricorgau/home/JournalPublicationsaspx. Pitt A (Pers comm). Grower Services Superintendent Bundaberg Sugar Pty Ltd. Robson, A, Abbott, C, Lamb, D and Bramley, R (2011). Paddock and regional scale yield prediction of cane using satellite imagery Poster Abstract. Proceedings of the Australian Society of Sugar Cane Technologists. 33rd Conference Mackay Qld AUS 4 – 6th May 2011. Robson, A, Abbott, C, Lamb, D and Bramley, R (2010). Remote Sensing of Sugarcane; answering some FAQ’s. Australian Sugarcane 2011. p6-8. Rudorff, BFT and Batista, GT (1990). Yield estimation of sugarcane based on agrometerologicalspectral models. Remote Sensing of Environment. 33.183-192. Rueda, CA, Greenberg, JA and Ustin, SL (2005). StarSpan: A Tool for Fast Selective Pixel Extraction from Remotely Sensed Data Center for Spatial Technologies and Remote Sensing (CSTARS). University of California at Davis Davis CA (Starspan GUI website link: https://projectsatlascagov/frs/downloadphp/581/install-starspan-win32-020jar) Simões, MDS, Rocha, JV and Lamparelli, RAC (2005). Spectral variables growth analysis and yield of sugarcane. Science in Agriculture 62.199-207. Simões, MDS, Rocha, JV and Lamparelli, RAC (2009). Orbital spectral variables growth analysis and sugarcane yield. Science in Agriculture 66(4).451-461. Singels, A, Smit, MA, Redshaw, KA and Donaldson, RA (2005). The effect of crop start date crop class and cultivar on sugarcane development and radiation interception. Field Crops Research 92.249260. SPOT Image (2008) SPOT Image Homepage 28th April 2008.Website link http://wwwspotimagecom. Xiao, X, Hollinger, D, Aber, J, Goltz, M, Davidson, E, Zhang, Q and Moore, B (2004a). Satellitebased modelling of gross primary production in an evergreen needleleaf forest. Remote Sensing of Environment 89.519-534. Xiao, X, Zhang, Q, Braswell, B, Urbanski, S, Boles, S, Wofsy, S, Moore, B and Ojima, D (2004b). Modelling gross primary production of temperate deciduous broadleaf forest using satellite images and climate data. Remote Sensing of Environment 91.256-270. Xiao, X, Zhang, Q, Saleska, Hutyra, L, De Camargo, P, Wofsy, S, Frolking, S, Boles, S, Keller, M and Moore, B (2005). Satellite-based modelling of gross primary production in a seasonally moist tropical evergreen forest. Remote Sensing of Environment 94.105-122. SRDC Project DPI021 Final Report_Appendices.doc 35 Zhou, MM, Singels, A and Savage, MJ (2003). Physiological parameters for modelling difference in canopy development between sugarcane cultivars. In Proceedings of the South African Sugar Technologist Association 7.610-621. Article in Australian Sugarcane. February- March 2011. Remote Sensing of Sugarcane; answering some FAQ’s. Andrew Robson1, Chris Abbott1, David Lamb2 and Rob Bramley3. 1Department of Employment Economic Development and Innovation. Kingaroy, QLD. 4610. 2University of New England. Armidale NSW. 2351. 3CSIRO, Adelaide. SA. 5064. Remote sensing technologies have the potential to drastically improve the monitoring of spatial variability within Australian Sugarcane crops aiding in the better implementation of inputs and the optimisation of crop yields. Applications such as the detection of crop damage from pest, disease and poor nutrition, as well as the prediction of yield and its stability over time, have been investigated within many agricultural systems including sugarcane. However, grower adoption in the sugar industry remains low. This is likely attributed to a poor understanding of what technologies are available, who provides them, what they cost and the cost/benefit of implementation, as well as an overall shortage of knowledge and skills in the interpretation and then dissemination of the data to end users. This article addresses some of these issues in an attempt to inform consultants and growers of the potential benefits of adopting these technologies, particularly with the ongoing pressures of maintaining economical and environmental sustainability. What is Measured? For the assessment of agricultural crops, remote sensing platforms measure the amount of solar electro-magnetic radiation (EMR) reflected and transmitted by a plant canopy. The measurement of EMR within the Near-Infrared (NIR) region (700 – 1300nm) provides an indication of a plant’s internal canopy structure as predominantly influenced by leaf water content and morphology and therefore is specific to plant stress or desiccation. The Red region (600-700nm) is specific to chlorophyll concentration. Ratios of NIR and Red reflectance such as NDVI (Normalised Difference Vegetation Index = NIR-Red/NIR+Red) or Plant Cell Density (PCD= NIR/Red) when mapped across a crop are commonly used to emphasise spatial variability in plant structure and condition. The greater the NDVI or PCD value at a given location the more vigorous the plant and generally the higher the associated crop yield. What is the best Spatial Resolution? Currently available commercial satellites offer a wide range of spatial resolutions, defined by the size of the on ground image picture element or pixel. These range from 0.5m2 to 1km2 with the optimum resolution determined by the required application i.e. plant, farm or landscape scale. Using too low a spatial resolution may limit the ability to define specific crop features such as disease or pest ‘hot spots’ or even crop boundaries as each pixel displays multiple features smoothing out actual variability. Conversely, too high a spatial resolution may complicate the definition of larger management zones due to a ‘salt and pepper’ effect where each pixel is providing a measure of independent features such as plant canopy, trash, soil and shadow. In general, high resolution imagery is better suited for measuring localised plant stress such as that from cane grubs or nematode damage, weed and disease, as well as overall spatial variability within smaller crops. Mid resolution imagery i.e. 10m to 20m, could be considered to be more suitable for identifying variability trends across whole crops, farms and catchments including those arising from soil variability, topography or prior history. An example of a single cane block captured at two different spatial resolutions is included below. SRDC Project DPI021 Final Report_Appendices.doc 36 False colour image of plant cane as identified by 0.8m IKONOS imagery (left) and 10m SPOT5 imagery (right). The brighter the red colour, the higher the infrared reflectance and the more vigorous the crop. Commercially available satellite options and their associated cost: Selection of the most cost-effective imagery ultimately depends on the intended application, with spatial resolution as well as the minimum capture area having the greatest influence. In general, higher spatial resolution images are more expensive per hectare than lower resolution images, but generally have a lot smaller minimum required capture area i.e. 47km2 compared to 5000km2. If an individual grower is purchasing an image directly from a provider, they will generally have to pay for an area greater than their farm, resulting in the cost for useable imagery on a hectare basis increasing. This can be minimised by including a number of neighbouring growers to share the cost, have a consultant provide the imagery as part of their agronomic service or use an image on-seller who can purchase whole scenes and then on-sell each property as required. The later may result in a slight increase in imagery cost. However, the resultant product is likely to be correctly processed for geographic accuracy and have vegetation indices such as NDVI applied. Also worth noting is the image revisit time, where the higher the frequency the more chance that imagery will be successfully captured in regions with continued cloud cover. Other platforms such as LIDAR, Radar and aerial imagery are available but are not covered in this article. A list of commercially available satellite imagery platforms and associated costs is provided below. Commercially available satellite imagery options for assessing within-field variability. SRDC Project DPI021 Final Report_Appendices.doc 37 Some commercial providers and on-sellers of satellite imagery: AAM: www.aamgroup.com Geoimage: www.geoimage.com.au SPOT imaging Services: www.spotimage.com.au. Sinclair Knight and Merz: www.skmconsulting.com Precision Agriculture.com: www.precisionagriculture.com.au CTF solutions: www.ctfsolutions.com.au Terranean mapping technologies: www.terranean.com.au Agrecon: http://www.agrecon.com. What applications can imagery be used for?: Growers generally have a good understanding of inherent spatial variability within their cane blocks. Satellite imagery can improve this awareness by indicating the exact location and area affected by a cropping constraint, as well as identifying those events such as pest incursion or lodging that do not persist from season to season. From the following example, a large degree of variability within a plant cane crop can be seen, with high vigour or PCD shown as Blue and low vigour or PCD as Red (like NDVI, PCD gives an indication of the size and health of the plant canopy). From this map, GPS guided agronomic and yield assessments can be made to determine the nature of the constraints as well as their impact on productivity. In this example yield, commercial cane sugar (CCS) and soil samples were collected to coincide with commercial harvest. Classified PCD IKONOS 0.8m image of plant cane crop with sample locations. (Area =11.5ha) The low PCD regions within this crop yielded 90 tonnes of cane per hectare (TCH) compared to 170TCH in the high areas. The relationship between the point source measurement of cane yield and the corresponding image information can allow a surrogate yield map to be developed as well as provide an estimate of total yield. SRDC Project DPI021 Final Report_Appendices.doc 38 Correlation between PCD and TCH Surrogate yield map derived from PCD/ TCH correlation The development of a generic yield algorithm that can accurately predict total yield as well as yield variability without the need for crop sampling would be of great benefit to the Australian sugar industry. This would not only assist growers with management decisions but, as Mackay Sugar Ltd has demonstrated, can also allow more accurate regional decisions regarding forward selling, handling and storage to be made prior to harvest. Cost-benefit of informed decisions regarding remedial action: By identifying the area of reduced productivity as well as the resulting yield deficit, an estimation of lost productivity in monetary terms can be made. From the example above the low vigour regions yielded 10 tons of harvestable sugar per hectare less than the high yielding areas. Expressed in monetary terms this would equate to $4,500 (at $450t) per hectare. With the low yielding area extending over 4ha this would equate to $18,000 of less than optimum productivity. By identifying the nature of the limiting factor, in this case sandy subsoil with reduced water holding capacity, low EC, exchangeable nutrients and trace elements, a decision can be made on the cost/ benefit of applying remedial action such as the deep application of mill mud or clay. Understanding your blocks inherent variability: Imagery acquired over a number of cropping seasons can allow growers to understand the inherent spatial variability within their blocks. If the spatial orientation of both high and low crop regions remains unchanged across seasons and crop age (i.e.2005, 2008 and 2010 below) then well informed decisions can be made on the management of these blocks prior to planting including the use of variable rate technologies (VRT). If the zones are unstable from season to season (i.e. 2005 to 2007) then the impacts of climate, management or rotational effects should be considered and managed appropriately. 2005 Q188 2R 2007 Q205 AP 2008 Q205 1R 2010 Q208 SR SRDC Project DPI021 Final Report_Appendices.doc 39 Conclusion Pioneers of remote sensing technologies such as independent precision agricultural services and some industry groups have long understood the benefits of satellite imagery. However, for industry wide adoption it is imperative that all facets obtain some understanding of what technologies are available and what possible applications and cost/benefits they can provide. To increase overall adoption there is an obvious need to address the limitations of imagery accessibility, availability and minimum area; A web based framework linked to commercial image provider could possibly provide a solution. This method of accessing imagery, aligned with integrated farm management software would enable growers to form management decisions based on a culmination of spatial information including yield and soil maps, elevation etc. This would aide better data compatibility, recording and interrogation, resulting in the improved management of crop inputs and ultimately increased productivity. Although this is not a new concept, the availability of new commercial platforms and a greater awareness of what the technology can offer may improve adoptability and enable Australian cane growers to maintain economic and environmental sustainability. Acknowledgements. The authors would like to acknowledge SRDC for providing funding for this research as well as those growers and industry partners whom have collaborated. Also DERM for providing access to imagery over a number of the target sites. For more information please contact Andrew Robson (andrew.robson@deedi.qld.gov.au) or Chris Abbott (chris.abbott@deedi.qld.gov.au). 2011 ASSCT Poster presentation. PADDOCK AND REGIONAL SCALE YIELD PREDICTION OF CANE USING SATELLITE IMAGERY Andrew Robson1, Chris Abbott1, David Lamb2 and Rob Bramley3. 1Department of Employment Economic Development and Innovation. Kingaroy, QLD. 4610. 2Precision Agriculture Research Group, University of New England. Armidale NSW. 2351. 3CSIRO Ecosystem Sciences, Adelaide. SA. 5064. KEYWORDS: Yield prediction, NDVI. Abstract The pre-harvest forecasting of regional cane production within any given season is of great importance to the Australian cane industry. If inaccurate, significant financial penalties can be incurred by marketers with roll on effects for mills and growers. The use of remote sensing to predict yield is not a new concept, with progressive mills such as Mackay Sugar using SPOT satellite imagery for a number of years. However other regions are yet to implement this technology. Current research, funded by SRDC, has developed a preliminary algorithm for the Bundaberg region that has demonstrated accurate yield predictions for both large cropping regions and within individual crops. The algorithm was developed from the correlation between NDVI (normalised difference vegetation index) values derived from a SPOT 5 image captured on the 10th May 2010, with 2010 cane yields measured from whole blocks and from point source locations within individual crops (R2= 0.61; n= 112). These data included 12 varieties and 15 planting stages. To assess the robustness of the algorithm, it was applied to 2008 season imagery captured on the 31st March 2008. For 600ha of cane, a yield of 39,707 tonnes of harvested cane or 66.5 tonnes cane per hectare (TCH) was predicted which was 3.8% under the actual delivered yield (41,255 t at 69 TCH). The development of a subsequent SRDC Project DPI021 Final Report_Appendices.doc 40 algorithm using both 2008/2010 data (R2= 0.6; n= 151) did not improve the accuracy of the prediction, indicating that the relationship between yield and NDVI for the Bundaberg region may be consistent across seasons; this requires further validation using expanded datasets. The production of potential yield maps prior to harvest is also of great benefit to cane growers. To test whether such maps could be developed from the regional yield prediction algorithm, the predicted yields of point source locations (area 200m2) within two crops from the 2010 season were validated against measured hand cut samples. When compared to a one to one relationship between actual and predicted yield, the predicted yields showed a tendency to over-predict in low yielding areas and under-predict in high yielding areas. Again, further refinement and validation of the algorithm in the 2011 season is expected improve the prediction accuracy. SRDC Project DPI021 Final Report_Appendices.doc 41 SRDC Project DPI021 Final Report_Appendices.doc 42 2010 ASSCT Poster presentation. USING SPATIAL MAPPING LAYERS TO UNDERSTAND VARIABILITY IN PRECISION AGRICULTURAL SYSTEMS FOR SUGARCANE PRODUCTION. By JR HUGHES1, RJ COVENTRY2, A ROBSON3 1Department of Employment, Economic Development, and Innovation, Mackay 2Soil Horizons Pty Ltd, Townsville 3Department of Employment, Economic Development, and Innovation, Kingaroy John.Hughes@deedi.qld.gov.au KEYWORDS: Soil Variability, EC Mapping, Crop Yield, Yield Monitoring, NDVI, Sitespecific Management Abstract Precision agriculture (PA) has been identified as an effective tool for identifying and then managing crop production across a wide range of farming systems, globally. The implementation of such technologies within the Australian Sugar cane industry also holds much potential however it is imperative that a strong cohesion between sound agronomy and PA technologies is first achieved. Intensive yield observations across five study sites in the Mackay, Burdekin, and Herbert districts identified that to manage within-paddock variability, improved strategies must consider the multifaceted interactions of variables including nutritional issues, seasonal conditions, management practices, and biological factors such as plant disease and pest damage. It therefore follows that there is also a need to combine a number of existing PA tools, quantified with corresponding field samples, to ensure a more accurate and robust diagnosis of crop production is achieved. The following example from the Herbert cane growing region demonstrates how the interaction of spatial data layers (satellite imagery, EC mapping, yield monitoring) can be used effectively to identify the spatial variability of crop production, including the use of strategic soil and yield sampling and for the prediction of lost production resulting from underperforming regions. The infrared reflectance images of plant cane crops, derived from high resolution satellite images captured just before harvest, identified in-season crop variability that related well with the expected variability driven by contrasting soil properties portrayed through soil EC mapping. A linear algorithm developed between Normalised Difference Vegetation Index (NDVI) values with strategically located manually harvested yield samples, was shown to be a reliable predictor of spatial variability as well as total crop yield when compared to final harvest weights and results obtained from a yield monitor on a commercial sugarcane harvester. These results indicate how decisions based on multiple mapping layers are likely to underpin new farm management strategies in the further development of a precision agricultural framework for the sugar industry. SRDC Project DPI021 Final Report_Appendices.doc 43 SRDC Project DPI021 Final Report_Appendices.doc 44 Article in the Burdekin ‘The Advocate’ Newspaper promoting the presentation at Burdekin Productivity Services Annual General Meeting 16th August 2011. SRDC promotion of DPI021 involvement in the Herbert Resource Information Centre Spatial Community in Action Conference. Ingham (18-19 August 2011) SRDC Project DPI021 Final Report_Appendices.doc 45 Hand out for BSES cane talks Developing Remote Sensing Applications in Cane D. """ oo! " Em ploy,", "\ Eooo",", o D. yo~pm"'t ' M ~"" ' on K ~ .. " y "",,"''"' "...~,--- --c-: ~. - " ,, SRDC Project DPI021 Final Report_Appendices.doc 46 The SRDC sugar cane project (DPI021) featured in a Department of Innovation, Industry, Science, and Research (DIISR): Space Policy Unit (SPU) handout promoting Earth Observation R&D being conducted in Australia (click icon- ).EOR&DSurvey Overview and Guidance Jul11.pdf SRDC Project DPI021 Final Report_Appendices.doc 47 Article in the Australian Canegrower, December 2010: 16. “Cane monitoring made easy with new sensor” Ca ne monitoring made easy with new sensor ---_ _.. ...--"""""""-oI_ .......~_ ....... _ " ...01 . ..,~._"... --'''_'-_-. _- -----_-".. ...,..--_"'-~-.."....-.,.'" ---"_'----- --""-...-..-......-" .... .. ...-..... ...~"""_bod~""""- -_-__-"-"-'.__-...., -. .. -. . -_,___,._._.,.-- --_---._. ---.., _-_..-,.....-..-,.'_"'-- .. .... ..--.. -.-....-"~.."..'"".".""._.,.."....-..,-.._-............. -_--_- -- ,__-__---","' -_-_-.__._--.. . ,,_ ...._....... .. .. -... ~-.,......,-- --_---_-_- -"'_-_."__--__" ..- -----_ --""_ ..... .-..... _..- ... .. ..... ...____ ...._M ~ ~_....-....... ........ "-",",, " - « --::':-~~ ---.~- '; .~-. ... ~..­ ........'...;.<.>..;fI...,..........~b.io.oi.<._,..",' -...,.,..;m"",,", """n...!• _ hruprtia> ofio>occtY _ ' CTtttrlo: ¥' ' ' ' _.!oiIioI of_l imItI'idon .. I>< . ........ io do< Il0l_ of"" pruj«I "' .. ",ofln oS ......... 101........ ......,_ of ·" ..__ . _.1·~1-'"-r.... ~.... __ .... _1....,J ..... _ 1><..,.._~, "'It . \>0)'" 0<"Ii000 '" """""" ~"""" (i.<. ill "" >0000 " .._ _ ""Iodio& lleld doys II todo """'" ~ -..ioI ~ io OobIry.;do pubIicII- · Dntt p<-.:.b io « · f"NI """""",of .~ _ _... ioc..... rx«Ibaoollt_!pis. - "'_1'I""idod.· t'iooJ ...;". ____ """Jl.Iiuj• ReportinI 01'_ .. _ _ iooblq JioJd do», """,It 1 _ ..,;l ~.-rioI ia _ , wll< ~Io. ~.OOO MiIe>Ooo>o I ",11<-0'. .,., e rN, rIo: • . OMondri< ..- 1be _ ~. . .0. .1__ v'" i-m"pr'os"'io"l,Jppm-I.O:T:IlKio>o.............,.._..... .nP-...'."_"'~ _ .r~" I>< cU:~ lor ado bIock.,;,u, olllbm: Iocotioos . . _ill. . . .".'......•....,._;.l."..".."..",.o,..-i.l...".."..",.'~.,.;-"...;............".'.''.'.'.M"j...,.~................... ,of1l>o -- _"II_"""""_._.,1>< _ _ ..·Ability.._iWoJ~ .. IIioMdM"''''''''''oI, • ........ ~ -.,iII """ .... ",,",'ioo be",fjs of........_ _ 0ptI e - . (o...::..T«Io) _ of ....m.. ~, rn.90l FOooi Rtfoort At....._ , ,,.-.., '+-. off..J _ ~.... _ . ,~ ~I(lx.:. SRDC Project DPI021 Final Report_Appendices.doc 66 Xiao, X, Hollinger, D, Aber, J, Goltz, M, Davidson, E, Zhang, Q and Moore, B (2004a). Satellite-based modelling of gross primary production in an evergreen needleleaf forest. Remote Sensing of Environment 89.519-534. Xiao, X, Zhang, Q, Braswell, B, Urbanski, S, Boles, S, Wofsy, S, Moore, B and Ojima, D (2004b). Modelling gross primary production of temperate deciduous broadleaf forest using satellite images and climate data. Remote Sensing of Environment 91.256-270. Xiao, X, Zhang, Q, Saleska, Hutyra, L, De Camargo, P, Wofsy, S, Frolking, S, Boles, S, Keller, M and Moore, B (2005). Satellite-based modelling of gross primary production in a seasonally moist tropical evergreen forest. Remote Sensing of Environment 94.105-122. SRDC Project DPI021 Final Report_without appendices.doc 79