Hybrid Deep Learning Framework with Satellite Imagery for Drought Prediction

Abstract Accurate and timely drought prediction in data-scarce, conflict-affected environments remains a critical yet largely unresolved challenge in hydrometeorological science. Therefore, this study proposes a Hybrid Deep Learning Framework (HDLF) integrating Convolutional Neural Network (CNN) spatial feature extraction, Bidirectional Long Short-Term Memory (BiLSTM) temporal encoding, and a multi-head self-attention mechanism, optimized via the Augmented Aquila–Harris Hawks Optimizer (A²-HHO), for multi-class drought severity prediction across three agro-climatic zones of Somalia. The model is trained on a 24-year (2000–2024) archive of six freely available satellite and reanalysis datasets, namely MODIS NDVI, CHIRPS precipitation, MODIS Land Surface Temperature, GRACE/GRACE-FO groundwater anomalies, ERA5 soil moisture, and ERA5 evapotranspiration with Standardized Precipitation Index (SPI) serving as the ground-truth drought severity label. The proposed framework achieved an overall accuracy of 92.7%, a macro-averaged F1-score of 90.3%, and a Cohen's Kappa coefficient of 0.908 on the independent test set (2021–2024), outperforming CNN-LSTM, standalone LSTM, Random Forest, and Support Vector Machine baselines by margins of 5.3 to 18.5 percentage points. Contribution analysis confirmed the incremental contributions of bidirectionality, attention, and A²-HHO optimization, with the latter alone yielding a 2.6 percentage point accuracy gain over manually tuned configurations. Temporal validation during the 2021–2023 Horn of Africa drought demonstrates reliable early warning capability up to three months in advance. This study provides a scalable and effective approach for satellite-based drought prediction in data-scarce environments.