Physics-Informed Explainable Machine Learning for Surface Urban Heat Island Driver Analysis and Cooling Strategy Optimization in Ahmedabad using remote sensing

In rapidly urbanizing semi-arid cities, Surface Urban Heat Islands (SUHIs) are increasing thermal stress and public health hazards, yet there are limited physically interpretable frameworks for identifying heat causes and assessing focused mitigation solutions. This study creates a physics-informed explainable machine learning framework for surface heat hotspot detection, cause attribution, and cooling strategy optimization in Ahmedabad, India.

Multi-source geospatial datasets like Sentinel-2 spectral indices (NDVI, NDBI, NDWI, MNDWI, NDMI), ERA5-Land atmospheric variables, ESA WorldCover land-use/land-cover data, OpenStreetMap-derived urban morphology metrics, and surface energy balance parameters estimated using the SEBAL approach were integrated with Land Surface Temperature (LST) derived from Landsat 8 OLI/TIRS Collection 2 imagery (April–June 2024). A total of 69,312 pixels with LST values ranging from 31.8°C to 55.2°C (mean 46.8°C) were included in the analysis. Strong sensible heat dominance was found in the SEBAL analysis, as evidenced by a mean Bowen ratio of 4.6 and a net radiation of 421 W m⁻².

An XGBoost regression model trained on 20 predictor variables demonstrated excellent predictive accuracy (R² = 0.906; RMSE = 0.89°C). Along with the established energy balance relationships, model interpretation using SHAP revealed surface albedo, net radiation, built-up intensity (NDBI), and surface moisture deficit (NDMI) as the main influences on urban thermal variability. A total of 13,863 hotspot pixels (the top 20%) were found using a composite Heat Stress Index that integrated LST and the Bowen ratio. These pixels formed an approximately 83 km² contiguous thermal hotspot, located within Ahmedabad's dense historic core.

Urban greening, water-body expansion, and cool roofing lowered temperatures by 1.5°C, 1.3°C, and 0.6°C, respectively, according to scenario-based models; a combined intervention produced a mean cooling of 4.7°C. A total of 9,192 priority pixels with cooling potentials higher than 3°C were found through spatial optimization. Streamlit was used to create an interactive dashboard that shows hotspot locations, driver relevance determined by SHAP, and intervention results for evidence-based ward-level heat mitigation planning and climate-resilient urban policy actions.