Predicting Crop Pest and Disease Outbreaks for Jordbruksverket

Agricultural productivity and food security are significantly challenged by crop diseases and pests. This study explores the development of interpretable Machine Learning models to predict the severity of two major crop diseases: Bladfläcksvampar (leaf spot fungi) and Mjöldagg (powdery mildew), affecting Swedish cereal crops. Historical pest, crop, and weather data obtained from Jordbruksverket and SMHI were utilized to compare single-target and multi-target modeling method employing regression models such as Random Forest, XGBoost, HistGradientBoosting, and Multi-Layer Perceptron, with multi-target models implemented via a MultiOutputRegressor wrapper. The analysis reveals that, single-target models generally achieved higher predictive accuracy, particularly with tree-based algorithms, due to their ability to use target-specific preprocessing, especially smoothed out-of-fold encoding for high-cardinality categorical variables. XGBoost showed slight gains in multi-target mode, indicating limited inter-target synergy. Interpretability analysis using SHAP and LIME highlighted that explanation quality depended more on data preprocessing, specifically encoding strategy, than modeling method. When preprocessing was consistent, both modelling methods produced similar, agronomically meaningful attributions. While MultiOutputRegressor simplified multi-target implementation, it increased training time and required per-target application of explanation tools. A user-friendly graphical user interface was also developed to support end-users such as farmers and policymakers. The study concludes that model success in multi-crop pest or disease prediction depends more on preprocessing design, algorithm selection, and alignment with interpretability goals, than the choice between single- or multi-target modeling alone, particularly in the context of agriculture-specific data and the use of MultiOutputRegressor.