Using ensemble learning to take advantage of high-resolution radar backscatter in conjunction with surface features to disaggregate SMAP soil moisture product

Abstract

ABSTRACT Remote sensing based retrieved of soil moisture from low frequency passive microwave observations is preferred in different aspects such as better spatial coverage and more measurement compared to traditional ground-based measurements. However, due to coarse spatial resolution of the observations, their applications are limited in local to regional studies. This paper provides a framework using random forest regression to disaggregate the daily SMAP enhanced soil moisture (SPL3SMP_E) utilizing several ancillary data to overcome the spatial resolution limits and cloudiness effects. Ancillaries were acquired from sentinel-1 radar, MODIS monthly NDVI, land cover, topography, and surface soil properties. To validate the downscaled results with 1-km spatial resolution, the OZNET soil moisture measurements and sparse TDR ground soil moisture measurements were collected from Murrumbidgee catchment (Australia) and Firozabad catchment (Iran), respectively. Downscaled soil moisture product unbiased root-mean-square error (UnbRMSE) of ensemble learning demonstrated a range of 0.023 and –0.07 cm3/cm3. The produced downscaled soil moisture exhibited better local heterogeneity when compared to the coarse data and tracked the dynamics of temporal changes in soil moisture. Furthermore, cumulative distribution function (CDF) analysis showed good accuracy of downscaled soil moisture in grassland and cropland. Taken together, the findings supported usefulness of the suggested methodology in downscaling the medium- resolution SMAP soil moisture product.