Understanding Land Surface Temperature Patterns in Istanbul: The Role of Blue–Green and Built-Environment Predictors Using Explainable Machine Learning

Rapid urbanization and the replacement of vegetated surfaces by impervious materials can substantially alter urban thermal environments, while water bodies, topography, and coastal setting may further modify these patterns. Understanding how these factors contribute to land surface temperature (LST), particularly across large and geographically heterogeneous cities, is important for spatially targeted heat mitigation. This study investigates the spatial variation in summer LST across Istanbul, Türkiye, using an explainable geospatial machine learning framework that integrates satellite observations with blue–green, built-environment, topographic, and coastal variables. Landsat 8/9 observations from June to August 2020–2025 were integrated with multi-source land cover, imperviousness, topographic, and coastline data within a common 500 m grid. Following predictor screening and comparison of five tree-based algorithms, XGBoost was applied to the full study area and evaluated using nested spatially blocked cross-validation. The model achieved a mean R2 of 0.9531 ± 0.0083, RMSE of 1.0616 ± 0.0293 °C, and MAE of 0.7734 ± 0.0201 °C across the five outer folds of the nested spatial cross-validation. Predictive performance decreased under the larger 20 km block configuration, indicating sensitivity to the spatial scale of validation. SHapley Additive exPlanations (SHAP) identified the Normalized Difference Vegetation Index (NDVI), used here as a spectral proxy for vegetation greenness rather than as a direct measure of vegetated area, as the largest contributor to the fitted model, followed by imperviousness, water body cover, and grassland. The importance of water body cover was sensitive to coastal grid geometry and decreased substantially when low-land-fraction coastal cells were excluded. Because vegetation-related information is also used in the emissivity adjustment of the Landsat Level-2 surface temperature product, the magnitude of the NDVI attribution is interpreted as model-based rather than as an independent estimate of vegetation cooling. A sensitivity model excluding the NDVI retained substantial predictive performance (R2 = 0.8063), with imperviousness becoming the leading predictor. A sensitivity analysis replacing the NDVI with independently derived tree cover percentage yielded lower model performance (R2 = 0.9160). Dependence analyses revealed nonlinear relationships between the main predictors and LST. Spatial SHAP mapping further showed that although the NDVI had the largest local SHAP attribution across most of Istanbul, water-related, coastal, and topographic factors became locally important under specific geographic conditions. These findings show the value of spatially explicit explainable machine learning for identifying both city-wide and local factors associated with urban surface temperature.