Integrating multi-source remote sensing and ensemble machine learning to quantify the surface urban heat island in Oradea municipality, Romania

Source abstract

Land surface temperature (LST) is a key parameter for understanding thermal conditions in urban environments, where surface characteristics and anthropogenic activities strongly influence the intensity of the Urban Heat Island (UHI) phenomenon. Analysing the spatial and temporal variability of LST enables the assessment of how land-use and land-cover changes affect the local temperature distribution. This study proposes a machine-learning (ML) based framework for assessing the Surface Urban Heat Island (SUHI) effect in a medium-to-large city in Romania. Spectral indices derived from Landsat 8 imagery, including the Normalised Difference Vegetation Index (NDVI), Normalised Difference Built-up Index (NDBI), Normalised Difference Water Index (NDWI), Urban Index (UI), and Albedo, were integrated as explanatory variables. Non-parametric correlation analyses were applied to evaluate the relationships between these indices and LST. ML-derived LST estimates were validated against independent LST data from the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument on the Terra satellite to assess model accuracy and robustness. The results indicate a significant increase in LST between 2013 and 2024, with an average warming of approximately 1.6 °C across the study area, largely driven by urban expansion and structural land-use changes. Among the evaluated models, the Random Forest algorithm achieved the highest predictive accuracy (RMSE ≈ 2.70 °C, MAPE ≈ 6.47%, R 2 ≈ 0.49), followed by Extreme Gradient Boosting (XGBoost) (RMSE ≈ 2.73 °C, MAPE ≈ 6.55%, R 2 ≈ 0.48) and the Support Vector Machine (SVM) (RMSE ≈ 2.79 °C, MAPE ≈ 6.78%, R 2 ≈ 0.45). Although all models successfully delineated the spatial patterns of SUHI, their ability to reproduce absolute LST values was limited, and some biases were observed in the estimation of thermal extremes. Overall, the findings demonstrate that integrating multi-source remote sensing data with ML techniques provides a reliable and effective approach for monitoring SUHI dynamics.