Spatiotemporal Neural Networks for Forecasting Climate-Related Disasters in Sub-Saharan Africa

Abstract Climate-related disasters such as floods, droughts, and heatwaves are significant threats to livelihoods and development in Sub-Saharan Africa. Accurate prediction of these hazards is difficult due to inadequate meteorological data and spatial inconsistencies, as well as the intricate interaction of various climatic parameters. In this paper, spatiotemporal neural networks (STNNs) are explored as a tool in disaster prediction, and their various architectures, namely Convolutional LSTMs (ConvLSTMs), Graph Neural Networks (GNNs), and Transformers, are compared, and a novel hybrid fusion framework is proposed. Datasets from satellite imagery (MODIS and Sentinel), reanalysis (ERA5 and CHIRPS), and ground-based weather stations are preprocessed to create multivariate input datasets consisting of rainfall, temperature, soil moisture, and vegetation indices. Composite loss functions are utilized in training all models, and a series of metrics are utilized to evaluate performance, namely RMSE, precision, recall, F1-score, and lead-time accuracy. k-fold cross-validation is utilized in validating results in regional case studies in Nigeria, Ethiopia, and South Africa. It is established that ConvLSTMs are optimal in flood prediction, GNNs in drought monitoring, and Transformers in heatwave prediction. Nevertheless, it is also established that a hybrid fusion model is superior to all individual models in all metrics, achieving a minimum RMSE of 0.31, a maximum F1-score of 0.89, and a maximum lead-time accuracy of 8.1 days. Regional evaluations also demonstrate that hybrid STNNs are adaptable to various regional hazards, and variable-based evaluations indicate that rainfall and soil moisture are key predictors of disaster occurrences.