An improved surface soil moisture downscaling approach over cloudy areas based on geographically weighted regression

This study proposed a methodological framework for downscaling AMSR-2 surface soil moisture (SSM) products over cloudy areas using MODIS LST/NDVI datasets. The experiment was conducted in a relatively large area of 430,000 km2 in the middle and lower reaches of the Yangtze and Huaihe rivers in China, which is characterized by humid climate and frequent cloudy weather conditions. As MODIS LSTs suffer from serious pixel loss due to cloud interference in this area, an effective LST interpolation method was preliminarily applied to achieve daily LST datasets with quasi-full covers. And rather small RMSEs in the range 1.5 K–3.5 K were obtained when the interpolated LST datasets were validated against a reference LST dataset built from observed relationships between LST and ground-based near-surface air temperatures on clear sky days. A regression equation was then established between AMSR-2 SSM and spatially resampled MODIS datasets using "Geographically Weighted Regression (GWR)" to implement the SSM downscaling process. SSM estimates downscaled by the GWR-based method showed a better performance over those downscaled by the traditional "universal triangle feature (UTF)" based method in view of their "non-biased RMSEs (ubRMSEs)", correlation coefficients, and mean biases with respect to ground-based soil moisture validation data. Comparisons between SSM estimates from MODIS LST inputs and those from interpolated LST inputs were conducted, and they showed that the SSM estimates downscaled by interpolated LST inputs performed only slightly poorer (with an ubRMSE difference no larger than 0.02 cm3/cm3) than those by MODIS data. Time series analysis further showed that the GWR-based downscaled SSM estimates with reconstructed LST data inputs are in phase with the variation in ground-based soil moisture with the exception of areas of extremely high vegetation cover or low temperatures. The framework proposed in this study thus proved feasible for the derivation of reliable downscaled high spatial resolution SSM estimates, an essential application in mitigating pixel loss under cloudy weather conditions.