Enhanced Urban Heat Island Modeling with Machine Learning and Regression Kriging in a Topographically Diverse Medium-Sized City

Urban heat islands (UHIs) represent a major environmental challenge, particularly in cities with complex topography and local air dynamics where spatial variability of the surface temperature is difficult to model. This study presents an enhanced UHI modeling framework that integrates Random Forest regression with Regression Kriging (RF–RK) in Cluj-Napoca, Romania. Using a network of 36 air temperature sensors and spatial predictors such as urban environment and location parameters, we evaluated the performance of RF–RK against traditional Multiple Linear Regression (MLR). Results show that RF–RK achieved substantially higher accuracy (R2 = 0.844, RMSE = 0.241 °C) compared to MLR (R2 = 0.653, RMSE = 0.359 °C). Spatial patterns revealed that a notable thermal gradient can be seen between the western and eastern parts of the city and the extension of the heat island core eastward. The combined approach effectively captured both the non-linear relationships of predictors and the spatial autocorrelation of residuals, outperforming single-method models. These findings highlight the potential of hybrid machine learning–geostatistical frameworks for urban climate research and provide insights for urban planning and heat mitigation strategies in topographically diverse cities.