A High-Resolution Land Surface Temperature Downscaling Method Based on Geographically Weighted Neural Network Regression

Abstract

Spatial downscaling is an important approach to obtain high-resolution land surface temperature (LST) for thermal environment research. However, existing downscaling methods are unable to sufficiently address both spatial heterogeneity and complex nonlinearity, especially in high-resolution scenes (<120 m), and accordingly limit the representation of regional details and accuracy of temperature inversion. In this study, by integrating normalized difference vegetation index (NDVI), normalized difference building index (NDBI), digital elevation model (DEM), and slope data, a high-resolution surface temperature downscaling method based on geographically neural network weighted regression (GNNWR) was developed to effectively handle the problem of surface temperature downscaling. The results show that the proposed GNNWR model achieved superior downscaling accuracy (maximum R2 of 0.974 and minimum RMSE of 0.896 °C) compared to widely used methods in four test areas with large differences in topography, landforms, and seasons. We also achieved the best extracted and most detailed spatial textures. Our findings suggest that GNNWR is a practical method for surface temperature downscaling considering its high accuracy and model performance.