He et al. (2025) Machine learning prediction of future land surface temperature from SAR optical fusion under urban expansion in Changsha, China

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-12-02
• Authors: Peng He, Zhihui Chen, Lin Zhang, Chengjun Ma, Chen Luo
• DOI: 10.1038/s41598-025-30976-5

Research Groups

• School of Architecture, Changsha University of Science and Technology, Changsha, Hunan, China
• Hunan Planning Institute of Land & Resources, Hunan Key Laboratory of Land Resources Evaluation and Utilization, Changsha, China
• Malanshan New Media College, Changsha University, Changsha, Hunan, China
• Central South University, Changsha, China

Short Summary

This study developed an innovative SAR–optical collaborative framework to reconstruct cloud-free land surface temperature (LST) and predict future LST under urban expansion in Changsha, China, demonstrating high accuracy and revealing a strong synchrony between built-up expansion and LST increase.

Objective

• To develop an integrated, model-based SAR–optical collaborative framework for reconstructing and forecasting urban LST in cloud-prone subtropical cities.
• Specifically, to leverage SAR’s all-weather scattering signals to generate continuous, cloud-free historical LST and couple this dataset with machine-learning-based forecasting to anticipate future thermal evolution under ongoing urban expansion.

Study Configuration

• Spatial Scale: Urbanized core of Changsha City, China, encompassing seven administrative divisions, with a total land area of 3,906.39 km². LST prediction and LULC classification were performed on 500 m × 500 m spatial grids, with underlying data at 20 m (Sentinel-1) and 30 m (Landsat-8/9, DEM) resolutions.
• Temporal Scale: Historical data from 2018, 2021, and 2024 (August). Future LULC and LST predictions for 2027 and 2030.

Methodology and Data

• Models used:
  • Land Use and Land Cover (LULC) classification: Random Forest algorithm
  • LULC simulation: Patch-Generating Land Use Simulation (PLUS) model, incorporating a Markov chain model and Random Forest for driving factor analysis.
  • Land Surface Temperature (LST) interpolation (cloud removal): Random Forest regression (using SAR features).
  • LST prediction: XGBoost algorithm.
• Data sources:
  • Satellite: Sentinel-1 dual-polarization data (VV, VH backscatter, texture features, polarimetric decomposition parameters), Landsat-8 and Landsat-9 thermal infrared (B10) and optical data.
  • Ancillary: Digital Elevation Model (DEM) data, slope, aspect (from Google Earth Engine/SRTM), OpenStreetMap (OSM) data for roads, railways, water bodies, and Points of Interest (POIs) such as hospitals, factories, and schools.

Main Results

• The SAR–optical integrated LST prediction model achieved high accuracy with an RMSE of 0.9940 °C, an MAE of 0.4714 °C, and an R of 0.9819 for 2024 validation.
• LULC classification for 2018, 2021, and 2024 showed high accuracy, with overall accuracy, user’s accuracy, producer’s accuracy, and kappa coefficients all exceeding 0.87.
• Between 2018 and 2024, grassland decreased by 14.30% (164.52 km²) and forest by 2.34% (28.36 km²). Residential building areas increased by 10.92% (95.08 km²), and industrial building areas by 18.07% (88.09 km²).
• From 2018 to 2024, areas with LST below 40 °C decreased by 4.32% (169.48 km²), while moderate- and high-temperature zones (40–60 °C) expanded by 4.12% (1,177.56 km²). Extreme high-temperature areas (>60 °C) expanded from 0.48% to 0.69% of the total area.
• Predictions for 2024–2030 indicate continued urban expansion: residential built-up areas are projected to increase by 4.70% (40.9 km²), and industrial built-up areas by 6.06% (29.57 km²). Grasslands are projected to decline by 4.80% (55.22 km²), and forests by 1.42% (17.25 km²).
• Predicted LST for 2024–2030 shows a clear trend of intensifying surface temperatures: areas exceeding 60 °C are projected to expand from 0.6845% (26.68 km²) to 0.8622% (33.61 km²), representing a 25.9% increase. The 50–60 °C range is projected to expand by 58.07 km².
• SHAP analysis revealed that surface texture (ASM, Anisotropy) and vegetation structure (VH) are associated with lower LST, while increased structural complexity, heterogeneity, and radar backscattering intensity (LIA, GLCM entropy, VV, Alpha, Contrast) contribute to higher LST.

Contributions

• Developed an innovative SAR–optical collaborative framework for LST retrieval and prediction, specifically addressing cloud contamination challenges in subtropical regions.
• Systematically leveraged Sentinel-1 dual-polarization SAR features (backscatter, texture, polarimetric decomposition) to generate continuous, cloud-free historical LST, enhancing data availability and accuracy beyond traditional optical-only methods.
• Integrated this SAR-optical fused dataset with machine learning (XGBoost) for robust forecasting of future LST dynamics under urban expansion, providing a novel methodological pathway for urban thermal studies.
• Quantified the LULC-specific contributions to urban thermal change and identified industrial building areas as principal hotspots of extreme thermal emissions, offering actionable insights for urban climate adaptation and heat mitigation planning.
• Provided a generalizable and transferable methodology for advancing LST reconstruction and prediction in cloud-prone megacities, supporting resilience-oriented planning and sustainable spatial development.

Funding

• Natural Science Foundation of Hunan Province of China (2024JJ8332, 2025JJ80036)
• Open Topic of Hunan Key Laboratory of Land Resource Evaluation and Utilization (2025)

Citation

@article{He2025Machine,
  author = {He, Peng and Chen, Zhihui and Zhang, Lin and Ma, Chengjun and Luo, Chen},
  title = {Machine learning prediction of future land surface temperature from SAR optical fusion under urban expansion in Changsha, China},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-30976-5},
  url = {https://doi.org/10.1038/s41598-025-30976-5}
}