An Open-Source Deep Learning Framework for Scalable Urban Heat Island Detection Using Geospatial Data

Abstract

Urban Heat Islands (UHIs), where urban areas exhibit elevated temperatures relative to their rural surroundings, pose growing challenges in the context of climate change, particularly for densely built, vegetation-scarce cities. Traditional methods for UHI detection, often based on empirical indices or statistical regressions, lack spatial resolution, scalability, and adaptability across diverse urban environments. This study introduces an open-source deep learning framework that integrates multi-source satellite imagery and urban geospatial data to detect, map, and analyse UHIs with high spatial fidelity. The framework leverages a U-Net convolutional architecture with attention mechanisms to predict land surface temperature (LST) and delineate UHI hotspots. Input features include NDVI, impervious surface area, building density, and land use classifications, processed through a reproducible pipeline built with open-source tools such as QGIS, TensorFlow, and GDAL. Applied to Lagos, Nigeria, a rapidly urbanizing tropical megacity, the model achieved high predictive performance, successfully identifying critical hot zones and spatial correlations with urban morphology. The results reveal strong associations between UHI intensity and impervious surfaces and inverse correlation with vegetation. The framework’s open architecture, combined with publicly released datasets and modular code, ensures adaptability for use in both data-rich and resource-limited settings. This research contributes a transparent, scalable, and participatory approach to UHI detection, offering actionable insights for climate adaptation, heat risk mitigation, and sustainable urban planning. It underscores the importance of open geospatial AI tools in promoting equitable and data-driven environmental governance. Copyright: