Work

A Scalable Remote Sensing and Machine Learning Framework for High-Resolution Land Surface Temperature Modeling

Year: 2025

Type: article

Abstract: Understanding urban heat trends, biological changes, and effects of climate change depends on knowing Land Surface Temperature (LST) in dynamic coastal settings. This work integrates multi-sensor satellite images with socio-environmental characteristics using machine learning to provide a scalable framework for high-resolution LST estimation in the coastal area. We combined thermal data from the Moderate Resolution Imaging Spectroradiometer (MODIS), optical imaging from Sentinel-2, and auxiliary information like precipitation, elevation, and population density using the Google Earth Engine (GEE) platform into a single geographic model. A Random Forest (RF) regression approach is used to forecast LST, designed on 4,305 samples and tested on 1,076 randomly dispersed sites at 100-meter resolution. The model shows good predictive ability with a Root Mean Squared Error (RMSE) of 0.624°C and a Mean Absolute Error (MAE) of 0.484°C. Significant correlations are found between surface temperature, topography, rainfall, and human activity through spatial analysis. The generated fine-scale LST maps provide practical information for reducing urban heat, adapting to climate change, and planning in highly populated coastal areas. This study highlights the potential of cloud-based geospatial platforms and machine learning for precise, effective, and scalable environmental monitoring in complex coastal systems.

Source: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

DOI: 10.1109/jstars.2025.3597395