Modeling Agricultural Drought Severity using Explainable Machine Learning Models Enhanced with Bayesian Optimization

Drought is one of the most harmful and economically damaging natural disasters which poses a threat to food systems around the world, with agricultural drought being the biggest problem faced by crops. It causes massive yield losses, price volatility, and instability of agrarian economies, where according to estimates economic damage exceeds USD 29 billion per year [1]. The use of drought indices is well-known as the primary tool used to describe drought onset in terms of timing, severity, and duration due to the conversion of multidimensional hydrometeorological information into one single value. However, despite being useful, drought indices suffer from some shortcomings while being applied individually; namely, the lack of temperature-related water stress, soil moisture content, stress response of vegetation, and land-atmosphere interactions affecting agricultural drought [2]. The rapid advancement of machine learning has transformed the methodological landscape for drought prediction and severity classification, offering data-driven architectures capable of learning complex, nonlinear relationships among hydrometeorological predictors without requiring explicit process parameterization. This study develops and evaluates an explainable machine learning framework for forecasting moderate and severe agricultural drought and for predicting 6-month Standardized Precipitation-Evapotranspiration Index (SPEI-6) 1-month ahead (2-month and 3-month are still in progress), using Northern Nigeria's Sudano-Sahelian belt as study area of approximately 27,000 0.05° grid cells. Four modelling approaches were carried out: Logistic regression as baseline, Bayesian-optimized XGBoost classifier and regressor, and probability-averaged ensemble with over 80 engineered features per configuration, derived from raw data and reanalysis data. SHapley Additive exPlanations (SHAP) values were computed for the best performing model to move beyond aggregate performance metrics to interpretable and defensible model behaviour. The results of the experiments are presented in the tables below. The Bayesian-optimized XGBoost classifier achieved the best accuracy and AUC-ROC for the operational season window (May to October), while the ensemble model traded precision for markedly higher recall (0.693) and balanced accuracy (0.802), a property directly relevant to early-warning systems that prioritise not missing drought onset. The XGBoost regressor explained 51.7% of the variance in continuous SPEI-6 values (RMSE = 0.670, Spearman rho = 0.776) but retained a systematic negative bias (-0.29), underestimating the most extreme wet and dry anomalies. SHAP analysis was applied to the Bayesian-optimized XGBoost classifier which identified the current-month SPEI-6 anomaly as the dominant driver of predicted drought probability by a wide margin, followed by its one-month lag, water-balance anomaly, and independent rainfall corroboration - confirming that the model's reasoning is physically grounded, while also revealing a degree of circularity from using SPEI-6 as both predictor and target. These findings so far demonstrate that systematic hyperparameter tuning, probability calibration, and post-hoc explainability can jointly deliver a transparent, accurate, and operationally interpretable drought early-warning tool.