Estimating Surface Soil Moisture from AMSR2 Tb with Artificial Neural Network Method and SMAP Products

In this study, we present a research to transfer the merits of SMAP (Soil Moisture Active Passive) to AMSR2 (Advanced Microwave Scanning Radiometer 2) with using machine learning method-artificial neural network. The surface soil moisture (SSM) products of SMAP were set as the reference data, while brightness temperature (TB) of various channels and the microwave vegetation index (MVI) obtained or derived from AMSR2 were input into an Artificial Neural Network (ANN). During training period (2015–2017), the ANN product (NNsm) can reproduce the SMAP SSM accurately, with a correlation coefficient (CC) of 0.74, Root Mean Square Error (RMSE) of 0.033 m3/m3, and Bias of −0.00008 m3/m3. It was found that machine learning method failed to provide reliable SSM over moderate vegetated areas where SMAP works well. With these trained networks, we developed a global soil moisture data set (named as NNsm) using AMSR2 TB from 2012 to 2018. Comparing to the in situ SM observations from all SCAN (Soil Climate Analysis Network) sites (named as SCANsm), NNsm has a good agreement with CC = 0.44, RMSE = 0.113 m3/m3 and Bias = 0.030 m3/m3, which is much better than those of the AMSR2 SSM products from JAXA and LPRM.