Downscaling Land Surface Temperature With A Moving Window Approach

The relationships between land surface temperature (LST) and the other surfaceparameters, e.g. the Normalized Difference Vegetation Index (NDVI), make it to bepossible to downscale the thermal images and the estimated LSTs. Many approacheshave been proposed by literatures based on the negative relationship between LST andNDVI. In this dissertation, the Disaggregation procedure for radiometric surfacetemperature (DisTrad) modelis improved through modeling the regression relationshipbetween LST and NDVI with a moving window method.Findings are:(1)The DisTrad model is proven to be applicable in the two study areas. The rootmean square errors (RMSEs) of the downscaled LST are4.09K in the middle reaches ofHeihe Reiver and5.67K in Sichuan, respectively.(2) By considering the spatial variability of LST caused by vegetation at the localand global scales, a moving window method is developed. DisTrad model for each pixelbased on NDVI and LST values in the pixel block is constructed. RMSE for downscaledLST is4.18K in Heihe area and5.29K in Sichuan area, respectively. Hence,the movingwindow method is has better performance than the traditional DisTrad model, especiallyin Sichuan area with high vegetation abundance.(3) Regarding the influences of topography on the spatial variability of LST, thedigital elevation model (DEM) and other land surface parameters are incorporated tobuild multiple regression model in the moving window method. The application inSichuan area shows that the RMSE of the sharpened LST is4.93K, demonstrating thatthe scheme is capable to estimate the sub-pixel of original LST using both NDVI andDEM.(4) Significant positive correlations may exist among the surface parameters.Therefore, the stepwise regression is utilized to select the candidate parameters in eachmoving window. Validation of this method in Sichuan area indicates that thedownscaled LST yield RMSEs of5.54-5.58K. The minimum RMSE value is4.93Kwhen both the NDVI and DEM are selected as the downscaling kernel. Therefore, the moving window method with the stepwise regression model has a good ability toimprove the spatial resolution of LST.