Machine Learning-Based Modeling of Land Surface Temperature in Lagos, Nigeria: Integrating Canopy Structure, Built Environment, and Surface Reflectance Variables

Abstract Urban heat is an escalating environmental challenge in tropical megacities, where rapid urbanization and declining vegetation cover intensify surface warming. This study applies a machine learning approach to predict and map Land Surface Temperature (LST) across Lagos, Nigeria, by integrating multisource remote sensing variables within a Random Forest (RF) framework. Three models of increasing complexity were developed using combinations of vegetation, structural, and spectral predictors derived from Landsat 8, GlobeFCH canopy height, ESA WorldCover, and SRTM data. Model 1, using vegetation variables (NDVI and canopy height), achieved an R² of 0.51, while Model 2, incorporating built-up and elevation variables, improved performance to R² = 0.66. The final model (Model 3), combining NDVI, canopy height, built-up percentage, elevation, NDBI, LULC, and albedo, achieved the best accuracy (R² = 0.74; RMSE = 1.77°C; MAE = 1.21°C). Partial dependence analysis revealed that NDVI and canopy height exert strong cooling effects, whereas NDBI and albedo were positively associated with surface warming. Spatial predictions highlighted pronounced thermal gradients, with high LST values concentrated in industrial and densely built-up areas like Ikeja, Apapa, Lagos Island; and cooler conditions in vegetated and coastal zones. These findings underscore the role of vegetation structure in mitigating urban heat and provide actionable spatial insights for urban greening and climate adaptation planning. The reproducible workflow demonstrates the potential of machine learning and Earth observation data for urban climate monitoring in data-limited tropical regions.