Land surface temperature (LST) is a key parameter in numerous thermal environmental studies. Due to technical constraints, satellite thermal sensors are unable to supply thermal infrared images with simultaneous high spatial and temporal resolution. LST downscaling algorithms can alleviate this problem and improve the spatiotemporal resolution of LST data. Spatial nonstationary and spatial autocorrelation coexist in most spatial variables. The spatial characteristics of the LST should be fully considered as a spatial variable in the downscaling process. However, previous studies on LST downscaling considered only spatial nonstationary, and spatial autocorrelation was neglected. In this article, we propose a new algorithm based on the geographically weighted autoregressive (GWAR) model for LST spatial downscaling. The digital elevation model and normalized difference build-up index were chosen as explanatory variables to downscale the spatial resolution of the moderate resolution imaging spectroradiometer LST data from 1000 to 100 m, and Lanzhou and Beijing were taken as the study areas. The performance of the GWAR model was compared with that of the thermal data sharpening (TsHARP) model and the geographically weighted regression (GWR) model. The Landsat 8 LST was used to verify the downscaled LST. The results indicate that the GWAR-based algorithm outperforms the TsHARP-and GWR-based algorithms with lower root mean square error (Beijing: 1.37°C, Lanzhou: 1.76°C) and mean absolute error (Beijing: 0.86°C, Lanzhou: 1.33°C). Index Terms-Geographically weighted autoregressive (GWAR), land surface temperature (LST), Landsat 8, moderate resolution imaging spectroradiometer (MODIS), spatial downscaling. I. INTRODUCTIONL AND surface temperature (LST) is one of the key parameters that influences the environment and ecological systems Manuscript

Land surface temperature (LST) is a vital physical parameter in geoscience research and plays a prominent role in surface and atmosphere interaction. Due to technical restrictions, the spatiotemporal resolution of satellite remote sensing LST data is relatively low, which limits the potential applications of these data. An LST downscaling algorithm can effectively alleviate this problem and endow the LST data with more spatial details. Considering the spatial nonstationarity, downscaling algorithms have been gradually developed from least square models to geographical models. The current geographical LST downscaling models only consider the linear relationship between LST and auxiliary parameters, whereas non-linear relationships are neglected. Our study addressed this issue by proposing an LST downscaling algorithm based on a non-linear geographically weighted regressive (NL-GWR) model and selected the optimal combination of parameters to downscale the spatial resolution of a moderate resolution imaging spectroradiometer (MODIS) LST from 1000 m to 100 m. We selected Jinan city in north China and Wuhan city in south China from different seasons as study areas and used Landsat 8 images as reference data to verify the downscaling LST. The results indicated that the NL-GWR model performed well in all the study areas with lower root mean square error (RMSE) and mean absolute error (MAE), rather than the linear model.

No abstract