Machine learning prediction of future land surface temperature from SAR optical fusion under urban expansion in Changsha, China.

In the context of growing urbanization and persistent cloud contamination in optical remote sensing, reliable large-scale land surface temperature (LST) monitoring in subtropical regions remains a significant challenge. To address this issue, this study develops an innovative SAR-optical collaborative framework that integrates Sentinel-1 dual-polarization features with Landsat-8 observations for improving land use and land cover (LULC) classification, restoring cloud-covered areas, and enabling high-quality reconstruction of cloud-free LST. Building on this foundation, future LULC dynamics were projected using the PLUS model, and LST variations were predicted with the XGBoost algorithm, which enabled quantification of LULC-specific contributions to urban thermal change. The prediction model achieved high accuracy (RMSE = 0.9940 °C, MAE = 0.4714 °C, R = 0.9819), underscoring the robustness of SAR-optical integration for LST reconstruction. The results further revealed a strong synchrony between built-up expansion and the increase in LST. Between 2024 and 2030, built-up land is projected to expand by 10.8%, accompanied by a 0.18% increase in extreme high-temperature areas. Overall, the proposed cloud-resilient LST retrieval and prediction framework offers practical value for urban climate adaptation, providing quantitative evidence to support heat mitigation planning, the optimization of green-blue infrastructure, and resilience-oriented spatial development.