Dual-Channel Convolutional Neural Network for Bare Surface Soil Moisture Inversion Based on Polarimetric Scattering Models

Abstract

The polarimetric synthetic aperture radar (PolSAR) can be used to obtain soil moisture by inverting scattering models at high resolution. The convolutional neural network (CNN) has been recently introduced to estimate soil roughness for PolSAR data, which need to be driven by a large amount of data. In this paper, a dual-channel CNN based on polarimetric models is proposed for soil moisture inversion, and it aims to further expand the applicable range of roughness in the X-Bragg model by integration with the integral equation model (IEM). Meanwhile, it fully utilizes the spatial information of PolSAR images to relax the number of required training samples when real data on the surface are difficult to obtain. Besides, we designed a framework based on this network. Coarse-grained inversion and fine-grained inversion of soil moisture are carried out through the qualitative classification network and the quantitative regression network, respectively. Experiments on simulated and airborne E-SAR data show that the proposed network can accurately fit the nonlinear relationship between polarization parameters and soil moisture, so as to improve the inversion accuracy with a small number of samples. In our experiments, the average inversion accuracy reached 95.39%, and the root mean square error (RMSE) of the regression network was 0.98%. This method can be applied to a wide range of soil moisture monitoring applications.

Keywords

• soil moisture inversion
• polarimetric synthetic aperture radar
• deep learning
• X-Bragg model
• integral equation model