Development of downscaling technology for land surface temperature: A case study of Shanghai, China

Abstract

Rapidly urbanizing megacities face multiple challenges such as heat island effect and ecological degradation. High-precision land surface temperature (LST) data is critical for optimizing urban planning and environmental management. However, the spatial resolution of LST data obtained by satellite alone is low, which has certain limitations in urban-scale analysis. Based on ECMWF ERA5-Land reanalysis data, Landsat, Sentinel and other remote sensing data, as well as ground station observation data, this paper takes Shanghai, China as a case study, uses two machine learning algorithms, Random Forest (RF) and eXtreme Gradient Boosting (XGBoost), and Multiple Linear Regression (MLR) method, to downscale and monitor LST with fine resolution. Results show that the three downscaling methods all have good fitting effects, with XGBoost emerging as a standout performer, with an impressive coefficient of determination (R²) of 0.97, a minimal root mean square error (RMSE) of 1.14 °C and a mean absolute error (MAE) of 1.85 °C. MODIS data is further upgraded from low resolution to higher resolution, and finally realizes multi-level downscaling from 1000 m to 30 m and 10 m, which greatly improves the monitoring accuracy of LST in urban areas, and supports the identification and evaluation of subtle spatial differences in heat island effect and microclimate characteristics. In addition, the results of this study have been successfully transferred to the Google Earth Engine (GEE) platform to achieve rapid update and analysis. This innovative application provides technical support for real-time and dynamic urban thermal environment monitoring, helping to optimize the management and decision-making of environmental resources.