An explainable deep learning framework for sub-seasonal to seasonal agricultural drought prediction over India

Agricultural drought (AGD) prediction is critical for agricultural water conservation in India, where ∼60% of the population depends on farming. This study presents an explainable data-driven framework to predict AGD at sub-seasonal-to-seasonal (S2S) scales (1–3-month lead) using a Stacked Long Short-Term Memory (SLSTM) model at 0.25° grid resolution across croplands in India. The framework integrates local antecedent land-atmosphere variables and non-local climate drivers, represented by Sea Surface Temperature (SST) anomalies, with Soil Moisture Percentile (SMP) as the drought indicator. A novel extreme-aware custom loss function combining weighted mean squared error and a time-shift penalty is used to train SLSTM model specifically to capture drought events. The model is trained from June 1981–May 2015 and tested from June 2015–May 2022. SLSTM model outperformed benchmark deep learning models (XGBoost and LSTM) and the SEAS5 dynamical forecast system in predicting major AGD events, including the 2015–16 event. At a 1-month lead time, SLSTM model demonstrated strong performance, with percentage correct and probability of detection between 0.6–1 and false alarm rate between 0–0.3, particularly in Northern, Southern, and Eastern India. Although performance declined with increasing lead time, the model remains conservative while minimizing overprediction. Predicted drought area correlates reasonably well with that of observations across all three lead times. We further analyzed shifts in grid-wise predictor importance between local and non-local features in response to changes in lead time at seasonal and climatological scales. This research supports the development of early warning systems for AGD risk management in India.