An intelligent data-processing and machine learning workflow for urban heat assessment in smart cities

Abstract

Smart city management relies on practical and repeatable ways to turn spatial data into information that planners can use. Sofia, Bulgaria, was selected as a test area for the proposed satellite-based urban heat assessment. The work uses 14,400 grid observations prepared in Google Earth Engine for June 1–September 15, 2024. The dataset combines Landsat 8/9 land surface temperature, NDVI, NDBI, MODIS albedo and geographic coordinates. These variables were examined through correlation analysis and used for regression model comparison with train-test validation, five-fold cross-validation and spatial block cross-validation. Gradient Boosting achieved the best result with R² = 0.886, mean cross-validation R² = 0.890 ± 0.004 and spatial block validation R² = 0.830 ± 0.035. An additional 2023 check confirmed the same best-performing model but with lower validation scores, showing that model performance is year-dependent. The workflow supports early-stage identification of priority areas for green infrastructure.