Evaluating Multivariable Statistical Methods for Downscaling Nighttime Land Surface Temperature in Urban Areas

For research and practice in fields such as the environment and meteorology, nighttime land surface temperature (LST) images at fine resolution provide important basic data. Since only a few satellite sensors can take fine-resolution thermal infrared images at night, such images are rather scarce. Downscaling of coarse-resolution LST images is a potential method for obtaining high-resolution LST images. However, downscaling methods have been mostly proposed for daytime LST images. This study aimed to evaluate the performance of methods that combine multiple explanatory variables and machine learning algorithms for downscaling coarse-resolution nighttime LST images in urban areas. Verification showed that the errors in the downscaling results were acceptable (mean absolute errors within 2 K). The resulting images could depict the spatial pattern of night LST in the study area in great detail. It was demonstrated that visible-near infrared images taken in the daytime could be used for downscaling of nighttime LST images, the rationality of which was deduced. It was also demonstrated that the performance of the spectral explanatory variables in nighttime LST downscaling was not lower than that of the mechanism explanatory variables. This method has high application value in many academic and practical efforts, such as land-atmosphere interface radiation budget studies, suitability assessments of human settlement, and urban environmental planning.