Interpretable Machine Learning Model for Spring Wheat Yield Forecasting

The research was conducted to construct an interpretable machine learning model (explainable artificial intelligence) for spring wheat yield forecasting. The data of a long-term field experiment (2001–2024) carried out in the forest-steppe of the Altai Priobye region were used. The scheme of the experiment provided for the analysis of predecessors’ role, methods of basic cultivation of leached chernozem, as well as levels of application of mineral fertilizers and chemical plant protection products in the formation of spring soft wheat yield. Extreme gradient boosting (XGBoost) was used to construct the model, and SHapley Additive exPlanations (SHAP) were used for its interpretation, which allowed estimating the contribution of each feature. The constructed XGBoost model demonstrated high prediction accuracy (R² = 0.95, MAE = 0.13 t/ha, RMSE = 0.17 t/ha), and integration with SHAP-analysis revealed the most significant features (5…6 out of 18) determining yield in the forest-steppe of the Altai Priobie. The greatest contribution to high predicted yield in the experiment was made by sufficient moisture for the crop year (596.5 mm; 1.19 t/ha), the use of fallow as a predecessor (0.58 t/ha) and the use of nitrogen-phosphorus fertilizers (0.21 t/ha). Low predicted yields are due to lack of moisture during the crop year (317 mm; –0.77 t/ha) and from May to October (246 mm; –0.24 t/ha), as well as high values of the sum of positive temperatures (2527.5 °C; –0.13 t/ha), low precipitation during wheat growing season (175 mm; –0.10 t/ha) and lack of application of plant protection products (–0.10 t/ha). The model expands the possibilities of applying machine learning, allowing us to obtain more reliable and informative results.

XGBoost, SHAP, explicable artificial intelligence, spring wheat yield, multifactorial field experiments, XGBoost, SHAP