Estimation of Surface Soil Moisture from Multi-Source Remote Sensing Data Based on Deep Learning Model

Soil moisture content (SMC) plays an important role in the study of global and regional climate change, energy balance and water cycle, and is an important parameter affecting hydrology, ecology and climate. In this study, it is utilized the advantages of deep learning algorithm in solving nonlinear problems and the characteristics of strong penetration of synthetic aperture radar(SAR) data are utilized improve the operational efficiency and maximize the inversion accuracy of soil moisture content. The inversion method is established and verified using datasets of In-situ measurements. The results show that Sentinel-1 data was more suitable and obtained a higher accuracy in the inversion of SMC. The SMC inversion method using the Back Propagation Neural Network (BPNN) model after feature selection is feasible and the operation efficiency is higher. The technique was tested at cropland with indicates the correlation coefficient (R) between the estimated and measured soil moisture was 0.71 and that the root mean square error (RMSE) was 0.083 cm3/cm3;The SMC of the study areas showed a strong spatial differentiation and had a strong consistency with the actual situation, which demonstrated the effectiveness of feature selection in the optimal collection selection of soil water inversion, and provided a new method for collaborative SMC inversion research of multi-source remote sensing.