Machine learning-based fusion and downscaling of multi-sensor satellite data for urban heat island analysis in Damascus, Syria

Abstract

Effective climate and environment monitoring requires high-resolution land surface temperature (LST) data, while satellite-derived LST data have coarse spatial and temporal resolutions that pose major challenges to detailed analysis. This study attempts to overcome this gap by integrating remote sensing and Geographic Information Systems (GIS) with machine learning algorithms to enhance LST accuracy, achieving a spatial resolution of 10 m. Data from MODIS, Landsat-9 and Sentinel-2 were processed using the Google Earth Engine (GEE) platform. Machine learning models, including Random Forest (RF), Support Vector Machines (SVM), Gradient Tree Boosting (GTB), K-Nearest Neighbours (KNN) and Naïve Bayes (NB), were applied to downscale LST data. RF model showed the highest accuracy, achieving an R2 of 0.91 and an RMSE of 1.003°C when validated against Landsat-9 data. It effectively preserved spatial details, capturing fine thermal variations in urban and industrial zones. Moreover, fusion of multi-sensor data within RF framework improved LST spatial resolution and enhanced UHI detection. Results also revealed elevated temperatures in densely populated and industrial areas, providing actionable insights for urban planners to mitigate UHI through vegetation, reflective materials and sustainable design.