Reconstruction Of Time-Series Soil Moisture From Amsr2 And Smos Data By Using Recurrent Nonlinear Autoregressive Neural Networks

Soil moisture (SM) is a key variable in describing land surface characteristics. However, most passive microwave sensed soil moisture products are spatially and temporally discontinuous. In this study, a recurrent autoregressive neural network was investigated for its capability to reconstruct time-series soil moisture. The train dataset was collected from the observations of AMSR2 and SMOS, along with the daily NDVI, land surface temperature (LST), precipitation (PRC) and DEM information. Then, the trained neural network was used to predict time-series soil moisture at a spatial resolution of 0.25 degrees. Result shows that this approach is promising in providing time-series soil moisture. Moreover, compared to ground soil moisture measurements, the predicted dataset tends to have lower root-mean-square error (rmse) and higher correlation coefficient (R) than the original soil moisture product of AMSR2 and SMOS.