Spatially Explicit Model for Statistical Downscaling of Satellite Passive Microwave Soil Moisture

Abstract

We introduce a spatially explicit statistical downscaling (SESD) method that fuses multiscale geospatial data with the soil moisture (SM) product from NASA's SM Active and Passive (SMAP) satellite. The multiscale data included the 9-km resolution SMAP SM image, 1-km resolution normalized difference vegetation index (NDVI), 1-km digital elevation model (DEM), 1-km resolution MODIS land surface temperature (LST), 500-m resolution gross primary productivity (GPP), 30-m resolution topographical water index (TWI), and West Texas Mesonet (WTM) station data. We used the random forest (RF) machine learning method to make a downscaled SM prediction at the 1-km resolution. Then, a regression kriging was applied to model the unpredicted variability at local scales to produce downscaled SM using the WTM station data. Due to the low number of ground truth samples, the validation was based on Monte -Carlo cross validation (CV) to calculate the unbiased root-mean-square deviation (ubRMSD), root-mean-square deviation (RMSD), and bias of the test set randomly separated from the training set from the WTM station data. Model validation showed that the downscaled SM data at the 1-km resolution can significantly improve the accuracy of the SM product as well as enhancing its spatial resolution. This article has its novelty in using the spatially explicit model to reconcile the scale difference from satellite data and ground observations.