Early warning of meteorological drought in Morocco’s Atlas mountains: a multi-paradigm benchmark of satellite-augmented deep learning and tree-based models
Abstract
Meteorological drought monitoring in Morocco’s High Atlas and Anti-Atlas Mountains remains reactive, yet few predictive studies have combined satellite observations with spatiotemporal deep learning in mountainous terrain. We develop a Convolutional Gated Recurrent Unit (ConvGRU) framework to forecast the 6-month Standardized Precipitation Index (SPI-6) across the High Atlas and Anti-Atlas at 1- to 3-month lead times. The model ingests 18 input channels combining CHIRPS precipitation, ERA5-Land reanalysis fields, satellite-derived NDVI, and static terrain features over a 115 × 210-pixel domain at 5.5 km resolution, with a strict leakage-free temporal split. At a 1-month lead, the ConvGRU and a non-spatial gradient-boosted ensemble (XGBoost) achieved nearly equal forecast accuracy (RMSE = 0.696 and 0.698, respectively), with the ConvGRU improving on damped persistence by 11.8% (Diebold-Mariano p = 0.003) and ridge regression by 9.7% (RMSE = 0.771). The added architectural complexity of spatiotemporal deep learning did not translate into superior forecast skill in this data regime, with the non-spatial gradient-boosted ensemble matching the convolutional models across horizons. In a test-period evaluation of the severe April 2024 drought (92% of the area affected), both the spatiotemporal and non-spatial models recovered most of the drought extent one month in advance (ConvGRU 81%, XGBoost 94%), far exceeding the 29% recovered by damped persistence. As the forecast horizon extends to two and three months, predictive skill across all architectures decays rapidly toward the baseline, defining a clear practical limit on meteorological drought predictability in this mountainous region. Together, these results show that while an operational one-month early warning window is achievable in this topographically complex region, adding spatiotemporal network complexity does not extend the predictability window beyond one month.