Machine Learning–Driven Assessment of Urban Heat Island Intensity in Athens: Integrating Remote Sensing, Landscape Metrics, and Green Infrastructure for Climate-Resilient Urban Planning

Sammanfattning

Environmental and health risks are increasing in large Mediterranean cities like Athens, where high urban density and limited vegetation intensify the Urban Heat Island (UHI) effect. This study analyzes the spatial patterns of surface urban heat in Athens, Greece, using satellite-derived Land Surface Temperature (LST) as a proxy for UHI conditions. LST datasets were derived from Landsat 8 thermal imagery and Sentinel-2 forecasts within Google Earth Engine (GEE) to achieve a high-resolution 10 m grid. Environmental variables, including vegetation (NDVI), building intensity (NDBI), and topography (DEM and slope), were analyzed alongside landscape metrics (PLAND, LPI, FRAC, ED, and LSI) to quantify the influence of both urban composition and spatial configuration. The research implemented a dual-track predictive framework: neighborhood-level machine learning models—Random Forest (RF), Support Vector Regression (SVR), Decision Tree (DT), and XGBoost—and a pixel-level Convolutional Neural Network (CNN) for spatial prediction. Results demonstrate that building intensity (NDBI) is the dominant driver of higher surface temperatures, while vegetation (NDVI) and topography contribute to localized cooling. Among the neighborhood-level models, Random Forest demonstrated the most consistent performance, achieving a maximum coefficient of determination (R2) of 0.5153 in August and a Root Mean Squared Error (RMSE) of approximately 1.75°C. In contrast, although the CNN model had low statistical accuracy (R2 < 0.20), due to data configurations and sensor noise, it managed to capture thermal points that were invisible in tables. Finally, the study highlights the great importance of integrating remote sensing, GIS-based landscape measurements, and machine learning for the changes that the climate crisis will bring. It thus appears that urban planning, at least in Mediterranean cities, should go a step beyond simple green space and, to more effectively address heat, should take into account spatial geometry and topographic factors. However, the study was limited by the small sample size of 53 administrative neighborhoods, thus limiting the statistical power of high-variance models such as XGBoost and CNN.