Machine Learning-Based Prediction of Land Surface Temperature in a Mining-Intensive Region Using Google Earth Engine

Abstract

This study examines the spatiotemporal dynamics of land surface temperature (LST) across Chhattisgarh, India, one of Asia’s largest coal-mining regions, from 1 January 2023 to 1 January 2024, using Google Earth Engine and a multi-variable dataset comprising NDVI, Albedo, Elevation, rainfall, nighttime lights, population density, and atmospheric constituents (CH4, AOD, NO2, SO2, O3). Results reveal pronounced thermal anomalies in active mining clusters driven by elevated CH4 (1899–1928 ppb) and AOD (0.73–0.84), while higher vegetation cover (NDVI 0.12–0.71) exerts a strong cooling effect. Pearson correlation analysis identified NDVI (r = −0.84), population density (r = 0.70), and elevation (r = 0.55) as dominant controls. Four tree-based machine learning models were developed; Random Forest and Extra Trees Regressor achieved the highest predictive accuracy (R2 = 0.96 and 0.94, RMSE = 0.17°C and 0.15°C, respectively), successfully capturing non-linear interactions among drivers. Ground validation at a central mining-zone station yielded moderate agreement (R2 = 0.53), highlighting residual surface-heterogeneity effects. These findings, along with the openly reproducible modeling framework, provide actionable, high-resolution insights for heat-risk zoning, prioritized reforestation, and emission-controlled mining practices, directly supporting Sustainable Development Goals 11 (Sustainable Cities) and 13 (Climate Action) in rapidly industrializing resource regions.