Downscaling Land Surface Temperature Using Multiscale Geographically Weighted Regression Over Heterogeneous Landscapes in Wuhan, China

Abstract

The deficiency of fine-resolution satellite-derived land surface temperature (LST) has impeded the meticulous exploration of the urban thermal environment at micro spatial scales. Although the LST downscaling methods are well-documented, the scale dependence between LST and environmental factors has been rarely considered in the regression establishment. This article proposes a new method (MGWRK) coupling multiscale geographically weighted regression (MGWR) and area-to-point kriging (ATPK) to generate fine-resolution LST maps over heterogeneous landscapes. Technically, normalized difference vegetation index (NDVI), normalized difference built-up index (NDBI), and albedo are selected using random forests (RF) as scale factors. First, the spatial and temporal robustness of MGWRK is evaluated by downscaling aggregated Landsat-8 Thermal Infrared Sensor images (1000-m) into finer resolutions of 500, 400, 300, 200, and 100-m in different seasons (spring, summer, and winter). Results reveal that MGWRK is superior to GWR and DisTrad methods for spatial details improvement (SSIM > 0.93) and LST fidelity ( ${\boldsymbol{R}^2}$ > 0.91) under all scales and in the seasons. Second, MGWRK is utilized to downscale the synchronous 1000-m Moderate Resolution Imaging Spectroradiometer (MODIS) LST map into 100-m. The quantitative and qualitative result substantiates that MGWRK obtains ideal performance when applied on MODIS data with few LST distortion ( ${\boldsymbol{R}^2}\; = \;0.915$ , RMSE = 1.607 K) and obvious spatial information enrichment (SSIM = 0.903).