Disentangling Vegetation Water Content and Surface Soil Moisture from Sentinel-1 Sar Time Series Through Asynchronous Convolutional Neural Network

This paper introduces and assesses an algorithm tailored for the discrimination of Vegetation Water Content (VWC) and Surface Soil Moisture (SM) using time series data acquired from the Sentinel-1 Synthetic Aperture Radar (SAR). The backscatter signals captured from the Earth's surface encompass contributions from both soil surface scattering and multiple interactions involving both the soil surface and canopy constituents. The conventional water cloud model encounters challenges in soil moisture retrieval due to the intricate interplay of moisture in both vegetation and soil, rendering direct retrieval from the backscatter coefficient a complex task. Employing the principle of blind signal separation, we deploy an asynchronous convolutional neural network to distinctively retrieve vegetation water and soil moisture from the Sentinel-1 SAR VV polarization backscatter time series. Our findings validated at OzNet network reveal an unbiased Root Mean Square Error (RMSE) of 0.054 m
³
/m
³
for soil moisture retrieval, with the accuracy of vegetation water content closely aligning with SMAP auxiliary VWC data.