Advanced Drought Prediction Using Hybrid Deep Learning Models: A Case Study of the High Atlas and Anti-Atlas Mountains

Abstract

Abstract. Morocco’s High Atlas and Anti-Atlas mountains have faced escalating drought severity in recent years, jeopardizing water security and rural livelihoods. Conventional drought monitoring often underperforms in these regions due to sparse meteorological stations and rugged terrain. This study develops a hybrid deep learning framework for operational SPI drought prediction at 5 km resolution, synthesizing remote sensing and climate variables (SPI, NDVI, soil moisture, precipitation, temperature) from 1990–2024. 128 engineered features—rolling statistics, seasonality, lag dependencies, and cross-variable interactions—enhance learning. We benchmark three recurrent neural network types (LSTM, Bi-LSTM, GRU), validated with held-out data (2021–2024). The GRU model achieved the highest predictive skill, reaching 91.89% accuracy within a ±0.2 SPI threshold and outperforming baselines (Random Forest, ARIMA). Our results demonstrate the value of advanced feature engineering and deep sequence learning for month-ahead drought early warning in semi-arid North Africa.