Machine learning-based modelling of land surface temperature dynamics in response to landscape fragmentation in industrial areas

Abstract

Urbanisation and industrialisation often replace vegetation with impervious surfaces, affecting thermal capacity and evaporation rates, and leading to increased land surface temperature (LST). This study develops predictive models for LST using biophysical parameters such as the Normalised Difference Vegetation Index (NDVI), Normalised Difference Built-up Index (NDBI), Enhanced Built-up Index (EBBI), and albedo, along with landscape fragmentation metrics from land use and land cover (LULC) maps for the years 2006, 2011, 2016, 2021, and 2024. We developed a Composite Fragmentation Index (CFI) using Principal Component Analysis (PCA) based on metrics such as the number of patches and cohesion. Random Forest (RF) and XGBoost (XGB) models are trained on this data for a peri-urban region, with XGB showing slight improvement over RF. The models achieved a strong performance, with an R2 of approximately 0.78 and an RMSE of around 1.7. Feature importance analysis indicates that higher NDVI significantly reduces LST, while increased NDBI and EBBI raise it. Uncertainty, quantified through conformal prediction, showed 90% prediction intervals of approximately ±2.8 °C. Additionally, Piecewise regression identified an NDVI threshold of 0.32, beyond which LST reductions become more pronounced. Overall, our findings highlight the need to consider both land cover composition and configuration in formulating effective strategies to mitigate heat stress in industrialising areas. © 2025 The Author(s). Published by IOP Publishing Ltd.