Early Warning of Meteorological Drought in Morocco's Atlas Mountains: A Satellite- Augmented Spatiotemporal Deep Learning Approach

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. This study develops a convolutional Gated Recurrent Unit (ConvGRU) framework for spatiotemporal forecasting of the 6-month Standardized Precipitation Index (SPI-6) 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 achieves an RMSE of 0.714 (9.5% improvement over damped persistence; Diebold-Mariano p = 0.041), while this improvement diminishes and becomes non-significant at 2 and 3-month horizons, establishing a clear predictability boundary. The contribution of dynamic climate and vegetation inputs is fundamentally nonlinear: Ridge regression applied to the same inputs captured only 2.3% improvement. In the test-period April 2024 drought (92% area affected), the ConvGRU detected 81% of the drought extent one month in advance compared to 29% by damped persistence. ConvGRU performed best overall relative to ConvLSTM and pixel-wise LSTM baselines, particularly at 1- and 2-month lead times, confirming that spatial convolution is essential for drought detection at longer lead times. These findings demonstrate actionable 1-month early warning capability and suggest potential for longer-horizon forecasting in topographically complex mountain regions.