Hybrid Bayesian deep learning model for predicting urban heat island intensity in African cities

Abstract Urban Heat Islands (UHIs), in which urban areas have higher temperatures than adjacent rural areas, lead to increased energy demand, poor air quality, and greater public health risks. These impacts are particularly significant in rapidly urbanizing regions of Africa, where climate-adaptive infrastructure has yet to be fully established. Here, we propose a hybrid predictive model derived from Bayesian Neural Networks (BNNs) for deep learning-based forecasting of Surface Urban Heat Island (SUHI) intensities across the African continent. The model employs attention mechanisms and Gradient Boosting Regressors (GBRs), using global climate and urban datasets from over 10,000 cities, incorporating variables such as land surface temperature, land use, and population density. This hybrid ensemble approach outperforms conventional deep learning methods, such as Convolutional Neural Networks (CNN) and Long Short-Term Memory networks (LSTM), when applied to structured climate data. The model was validated using ten-fold cross-validation, yielding a Mean Absolute Error (MAE) of 1.47 °C, Mean Squared Error (MSE) of 3.56 °C 2 , and an R 2 score of 0.84. Furthermore, it improved the overall accuracy of trained models by 12.3% in suburban areas and achieved an 8.6% reduction in MAE compared to baseline neural architectures. These findings highlight the value of uncertainty-aware and interpretable machine learning tools for climate adaptation. The proposed model can support urban planners in mitigating the thermal burden of rapidly growing urban populations in Africa, ultimately contributing to more livable, climate-resilient cities in the future.