Optimising productivity and variety recommendations through analysis of mill data : Final report 2016/32

Abstract

The annual productivity of the Australia sugarcane industry fluctuates significantly across most sugarcane growing regions. Although some of this regional variation can be explained by extreme weather events or disease incursions it is important to identify those causes which can be controlled in order to increase profitability for industry. Development of innovative tools to analyse and summarise mill data within a region can be used to identify those farm production units performing below potential and the factors associated with this. An increased understanding of these factors will broaden the adoption of improved farming practices by working with local industry to enable more appropriate selection of varieties to match field conditions, addressing impediments to farm productivity and Nutrient Management Planning.

The project developed an Industry Productivity Data Analysis and Reporting Tool (IP-DART) which assisted advisory groups, mills and research organisations to identify factors that can improve productivity and extension advice to growers and millers in the Tully, Burdekin and Herbert mill areas.

A ratoonability index was developed that combines the economic return of first ratoon crops and the rate of decline of economic returns between first and third ratoon to estimate the number of years until the crop reaches a defined economic threshold for profitability. This ratoonability index had a high correlation with the percentage of area ploughed out before 3R for the major varieties in the Herbert.

Machine learning techniques were used to explore what variables are most correlated with varieties that become commercially successful in Queensland. The results demonstrated that these approaches can be applied to practical questions on sugarcane productivity. However, linking the very large and distributed data from experimental clones in the breeding program with the small number of commercially dominant varieties resulted in a number of confounding factors influencing trait predictions. Large changes such as disease incursions also introduce bias into the data with the demise of Q124 resulting in orange rust resistance ranking very highly as a key success indicator. In practice orange rust resistance is a maintenance trait offering little in terms of future gains. There was also 100% accuracy from the confusion matrix which is a concern as the model may fail to generalize to future datasets. The trait hierarchy predicted for commercial success is not considered informative, but the application of the technique has demonstrated the importance of the underlying datasets. The methods could be further refined using data from a specific region as varieties are released on a regional basis. Use of climatic factors and a larger more balanced dataset could also be explored.