wcm.01.2026.230.242

Agricultural drought poses an escalating threat to crop production and on-farm water management across the semi-arid Mediterranean basin, yet Al-driven spatially explicit prediction frameworks remain scarce for North African drylands. This study develops an explainable artificial intelligence (XAI) framework for agricultural drought prediction across five semi-arid governorates of central Tunisia over the 2001-2022 period. A multi-source dataset of 154,704 pixel-month observations (586 pixels at 0.05° spatial resolution) was assembled by integrating 15 predictor variables spanning meteorological, topographic, edaphic, and socioeconomic domains relevant to crop stress assessment. The drought target variable (Standardized Soil Moisture Index, SSMI) was derived exclusively from GLEAM v4.2a, while soil moisture predictors were drawn from the independent NASA POWER MERRA-2 atmospheric reanalysis to ensure methodological rigor and preclude mathematical circularity. Six machine learning models were evaluated (XGBoost, LightGBM, CatBoost, RF, BPNN, and LSTM) using a strict temporal split (training: 2001-2014; testing: 2015-2020; validation: 2021-2022) focused on an unprecedented multi-year drought episode. Results show that BPNN achieved the highest predictive performance on the test set ( R ^ 2 = 0.86 , SDI = 0.626), whereas XGBoost demonstrated superior generalization during the extreme 2021-2022 validation period (R ^ 2 = 0.696, SDI = 0.448) establishing it as the most robust architecture for drought-stress prediction under extreme climate conditions. TreeSHAP interpretability analysis identifies MERRA-2 soil moisture as the dominant predictor (26.0%), followed by temperature (14.2%), sand content (10.0%), and precipitation (8.6%), revealing that edaphic conditions strongly modulate drought severity and crop water stress in the Tunisian interior. These findings provide an operational smart decision-support tool for drought early-warning and precision irrigation planning in semi-arid agricultural regions.