Hybrid modelling of random forests and kriging with sentinel-2A multispectral imagery to determine urban brightness temperatures with high resolution

Abstract

ABSTRACT Herein, a random-forest-based Kriging downscaling method (RFRKriging) was tested to increase the spatial resolution of Landsat 8 thermal-infrared remote sensing (TIRS) brightness temperature (T b) over the central areas of Guangzhou, China. The 240-m Landsat 8 T b was upscaled from the resampled 30-m Landsat 8 T b by using the spatial averaging method. T b was downscaled based on its relationship to spectral reflectance in visible/near-infrared and shortwave infrared bands, spectral indices, including normalized difference vegetation index (NDVI), normalized difference building index (NDBI), and modified normalized difference water index (MNDWI), digital elevation model (DEM), longitude, latitude, impervious surface, vegetation, and soil fractions. The linear regression Kriging (LRKriging) and NDVI-based thermal sharpening algorithm (TsHARP) regression Kriging (TsHARPKriging) were also implemented. Relative to the reference data from the resampled Landsat 8 T b, the RFRKriging produced satisfactory downscaling results with respect to the maximum coefficient of determination (R 2), the minimum root-mean-square error (RMSE), and the minimum ratio between RMSE and the correspondent Landsat 8 T b standard deviation (RSD). RFRKriging captured T b differences over roads, forests, water bodies, and buildings and enhanced the spatial details of downscaled T b in all urban areas. Furthermore, integrating the multi-type predictor variables from 10-m Sentinel-2A and 30-m Landsat 8 image sets, RFRKriging was further applied to downscale 30-m Landsat 8 T b to 10 m. Then, it was upscaled to 30 m and assessed with the original 30-m Landsat 8 T b. RFRKriging was an effective and practical method for increasing the spatial resolution, with a low RSD of 0.254, a low RMSE of 0.525 K and a high R 2 of 0.935. The RSD of less than 0.5 indicates that RFRKriging could achieve an acceptable downscaling accuracy. Furthermore, RFRKriging enhances the spatial details of a 10-m T b in complex urban environments while maintaining the intrinsic information and spatial patterns of a downscaled T b. Due to the unavailability of the 30-m and 10-m T b truth, this study may only provide a quantitative assessment from some theoretical experiments. Also, a comparison of the RFRKriging downscaled T b with the resampled Landsat 8 T b was performed. The results suggest that the proposed RFRKriging downscaling method shows better downscaling performance with more spatial details than the cubic convolution method applied by the National Aeronautics and Space Administration (NASA) Centre.