Performance evaluation of a neural network model and two empirical models for estimating soil moisture based on sentinel-1 sar data

Abstract

The objective of this paper is to propose an inversion model of soil moisture using a neural network and compare the performance of this method with two empirical models in soil moisture inversion. A wide dataset of backscattering coefficients extracted from Sentinel-1 images and in-situ soil surface parameter measurements (moisture content and roughness) are used. Since the available backscattering models have limited performances of describing the nonlinear relationship between soil parameters and backscatter coefficient, the retrieval of soil parameters from radar backscattering coefficient remains challenging. The proposed inversion method of a neural network is used for establishing this relationship. At the same time, two empirical models are employed to estimate the soil moisture for comparison. The results show that for most of the six measuring stations the inverted soil moisture with the neural network model has higher correlation coefficient with the in-situs oil moisture than those by the empirical models. Moreover, the neural network model inversion results under multipolarization input conditions are discussed in this paper. The results of stations 2, 4, and 5 show that R2 of multi-polarization inputs are increased by 0.1928, 0.4821, and 0.2758, respectively, compared with those of single-polarization inputs.