A systematic literature review of machine learning and deep learning for urban heat island modelling

Abstract The urban heat island (UHI) effect is a growing environmental and public-health challenge intensified by climate change and rapid urbanisation. While physical models have traditionally been used to study UHI processes, they often struggle to represent the complex spatial–temporal structure of urban environments. Recent advances in machine learning (ML) and deep learning (DL) offer new opportunities to address these limitations; however, their application in UHI research remains limited. This systematic literature review analyses 85 peer-reviewed studies published between June 2019 and December 2024, structured around four research questions addressing model typology, spatial–temporal scale, evaluation and explainability practices, and remaining research gaps. Adopting a model-centric perspective, this review analyses how different ML and DL model classes are designed, evaluated, and applied to represent multi-scale UHI processes, rather than treating UHI modelling as a purely descriptive comparison of algorithms. The results show that static spatial snapshot models remain dominant, particularly for surface-level classification and mapping tasks, while spatio-temporal and hybrid DL architectures are increasingly applied for short-term forecasting and dynamic UHI analysis. Most studies focus on single-city case studies, with limited evidence of cross-city or cross-climate generalisation. Regarding evaluation practices, only nine of the 85 studies incorporated physics-informed or hybrid approaches, and the adoption of XAI techniques remained limited. This review identifies critical methodological gaps and highlights the need for transferable and generalisable models, robust spatio-temporal evaluation strategies, and integrated explainability frameworks to move ML- and DL-based UHI modelling beyond experimental case studies toward reliable, interpretable, and decision-ready tools for urban planning and heat mitigation.