A Multi-Channel Convolutional Neural Network for Soil Moisture Estimation Using Combination of Polarimetric Channel in PolSAR Data

Soil moisture estimation using polarimetric synthetic aperture radar (PolSAR) data is of great interest and difficult issue in SAR quantitative research. One of the key issues to separate the effects of vegetation and soil roughness. In this paper, the polarization parameters used to describe the roughness of vegetation and soil are obtained by different combinations of polarization channels. A multi-channel convolutional neural network (MCCNN) is proposed to extract different polarization parameters for soil moisture classification and regression. In the network, we subtract the features extracted from different polarization parameters so as to effectively weaken the influence of vegetation and soil roughness on soil moisture estimation. The experimental results show that the classification network can achieve up to 94 % classification accuracy. Meanwhile, the root mean square error and the determination coefficient of the regression network are 2.46% and 0.92 respectively.