Explainable Machine Learning to Map the Impact of Weather and Soil on Wheat Yield and Revenue Across the Eastern Australian Grain Belt

Understanding the causes of spatio-temporal variation in crop yield across large areas is an important step forward in closing yield gaps and producing more food and fibre for the growing global population. While there has been much focus on using data-driven models to predict crop yield, there is also an opportunity to use these empirical models to understand which factors are driving variation in crop yield, and to quantify the contribution. This study uses a large database of 625 rainfed wheat yield maps from 14 different seasons (2007-2020) across the eastern grain belt of Australia. XGBoost models were used with predictor variables including digital maps of important soil attributes (pH, sodicity, and available water capacity), and weather indices (rainfall, frost, growing degree days). Together, the model and predictor variables could predict field-scale yield well, with a Lin’s Concordance Correlation Coefficient (LCCC) of 0.75 when using 10-fold cross-validation. SHapley Additive exPlanation (SHAP) values, a form of interpretive machine learning (IML), were then used to assess the impact of the variables on yield using the calibration dataset. SHAP values for each predictor variable were also mapped onto a grid of the study area for the 2020 season. These maps showed the impact of each predictor variable on wheat yield (t ha-1) and revenue ($ ha-1) in interpretable units. Weather variables had the largest impact, but results showed that soil variables impact yield differently depending on seasonal conditions, despite their largely temporally stable nature. Overall, the results show that data-driven models and IML are valuable in understanding the nature and degree of impact of important weather and soil variables on yield across large areas. This could have important implications on guiding governmental and industry policies and local farm management, as well as determining the magnitude and economic impact of constraints to crop production