A Deep Learning Approach to Flood Prediction and Early Warning Using Multi-Source Environmental Data: Evidence from Zimbabwe

Abstract

Zimbabwe ranks among the most flood-prone countries in Southern Africa; however, its existing flood warning systems lack adequate coverage and depend on insufficiently advanced technologies. In this study, we propose an applicable deep learning-based framework combining several sources of satellite, topographic, and in-situ data that would facilitate flood predictions and early warning systems implementation in three Zimbabwean catchments, the Save, Manyame, and Mazowe. Seven models were created and evaluated in this study; one of them, hybrid CNN-LSTM model, demonstrated better performance results with 95.9% accuracy, F1-score of 95.0%, and AUC-ROC of 0.981 for the independent test set. In addition, spatial cross-validation was applied to prove the generalization capacity of the proposed model. According to SHAP analysis, the following predictors were determined as the most influential: antecedent rainfall within 72 hours, distance to the closest river channel, and Terrain Wetness Index, all of which coincide with real-life features of Zimbabwe's hydrology. As for the end user, the suggested model could be incorporated into a four-level flood early warning system (advisory level, watch level, warning level, and emergency level).