Enhancing Land Surface Temperature Downscaling with Machine Learning: A Case Study in Bangkok’s Heterogeneous Urban Environment

Precisely estimating high-resolution land surface temperature (LST) is crucial for understanding urban heat dynamics and developing climate adaptation strategies. This study applies machine learning to enhance LST resolution from 30 m (Landsat 8/9) to 3 m by integrating PlanetScope multispectral imagery and elevation data. To ensure model robustness and interpretability, we assess the relative importance of predictors before performing Regression. We utilize Random Forest Regression to combine high-resolution predictors and refine temperature estimations, capturing complex urban heat patterns more accurately. LST is derived from Landsat thermal data, while high-resolution multispectral and topographic features serve as independent variables. After applying a stratified random sampling method, we train and evaluate the model using independent test data. The model achieves R² = 0.603 and MAE = 1.03°C, demonstrating its effectiveness in improving spatial LST resolution. Our findings highlight the potential of machine learning-based downscaling for urban climate research. By generating a 3m-resolution LST map, we provide key insights into urban heat islands, temperature variations, and land-atmosphere interactions. This approach empowers policymakers and researchers to develop data-driven heat mitigation strategies and climate-resilient urban designs. Ultimately, our research shows that combining high-resolution optical and thermal remote sensing with machine learning significantly enhances urban climate analysis, particularly in data-scarce environments.