Utilizing Self-Regularized Regressive Models to Downscale Microwave Brightness Temperatures for Agricultural Land Covers in the SMAPVEX-12 Region

A novel algorithm is developed to downscale microwave brightness temperatures (TB), obtained at satellite scales of 10-40 to ≤1km, meaningful for agricultural applications. Downscaling TB directly bypasses the errors induced by inverse modeling encountered while downscaling satellite-based soil moisture products. This algorithm is based upon self-regularized regressive models (SRRM) and uses higher order correlations between auxiliary variables, such as precipitation (PPT), land cover, leaf area index, and land surface temperature, and horizontally polarized TB observations. It includes information-theoretic clustering based on auxiliary variables to identify areas of similarity, followed by kernel regression that produces downscaled TB. The algorithm was evaluated using a multiscale synthetic dataset over North Central Florida for one year, including two growing seasons of corn and one growing season of cotton. Compared to the true TB, the downscaled TB had a root-mean-square error (RMSE) of 5.76 K with standard deviation (SD) of 2.8 K during the growing seasons and an RMSE of 1.2 K with an SD of 0.9 K during nonvegetated. The SRRM algorithm effectively captured the variability in TB at 1 km through the auxiliary variables. This algorithm was implemented to downscale SMOS observations available for five days during the SMAPVEX-12 experiment. Spatially averaged rootmean-square difference (RMSD) between the downscaled TB and the airborne TB observations from the airborne passive-active Lband sensor was 6.2 K, with Kullback-Leibler divergences of up to 0.91. For the SMAPVEX-12 dataset, better downscaling results are obtained for days when there was no PPT due to regional biases in the remotely sensed PPT from the NASA Tropical Measurement Mission. The RMSDs were lower when in-situ PPT data were used.