Detection of Surface Urban Heat Islands in Warsaw Using Satellite Remote Sensing and Machine Learning

Urban heat islands (UHI) intensify as cities expand, exposing residents to elevated thermal stress and complicating urban climate adaptation planning. Existing satellite-based approaches to detecting surface urban heat islands (SUHI) typically rely on a single class of data and narrow temporal windows, limiting their ability to capture the full range of processes driving surface overheating. This study develops and evaluates a random forest model for SUHI detection in Warsaw, Poland, integrating two classical spectral indices (NDVI, NDBI) derived from Landsat 8/9 Collection 2 imagery with three land-cover probability layers (built-up, tree, water) from the Dynamic World deep-learning product, processed in Google Earth Engine. Both a multi-year summer median composite (2020–2025) and individual annual summer composites were used, the latter enabling a leave-one-year-out temporal validation. Heat island pixels were defined as those whose land surface temperature anomaly exceeded +3 °C relative to the study area mean, a local criterion rather than a city-versus-rural contrast. The model achieved high and stable performance (accuracy = 0.831, AUC = 0.910 on the test set; AUC = 0.907 ± 0.004 in five-fold cross-validation and 0.905 ± 0.018 in leave-one-year-out validation). An ablation analysis showed that combining the probability layers with the spectral indices clearly outperformed the indices alone (AUC = 0.852 vs. 0.905), whereas the additional gain over the Dynamic World layers alone remained within uncertainty. Vegetation-related predictors (NDVI and tree probability) contributed more to classification than built-up indicators. These results indicate that vegetation deficit, rather than built-up presence alone, is the primary driver of surface overheating in Warsaw and that the proposed open-data workflow offers municipalities a low-cost screening tool for identifying priority areas for climate adaptation and, thanks to its reliance solely on open data, can be adapted to other cities, subject to further validation.