Soil roughness retrieval from TerraSar-X data using neural network and fractal method

The purpose of this study is to estimate the surface roughness (rms) using TerraSar-X data in HH polarization. Simulation of data is carried out at a wide range of moisture and roughness using the Integral Equation Model (IEM). The inversion method is based on Multi-Layer Perceptron neural network. Inversion technique is performed in two steps. In the first step, the neural network is trained using synthetic data. The inputs of the first neural network are the backscattering coefficient and incidence angle, and the moisture is the output. In the next step, three neural networks are built based on a prior and without prior information on roughness. The inputs of three neural network are backscattering coefficient, estimated moisture in the first step and incidence angle and the roughness is output. The validation of the proposed methods is carried out based on synthetic and real data. Ground roughness measurements are extracted from Digital Terrain Model (DTM) using the fractal method. The accuracy of moisture from synthetic data is 6.1 vol% without prior information on moisture and roughness. The roughness (rms) accuracy of synthetic datasets is 0.61 cm without prior information and is 0.31 cm and 0.38 cm for rms lower than 2 cm and rms between 2 and 4 cm, with prior information on roughness. The result's analysis of the simulated data showed that the prior information on roughness strongly improves the accuracy of roughness and moisture estimates. The accuracy of rms estimates for the TerraSar-X image in the HH polarization is about 0.9 cm in the case of no prior information on roughness. The accuracy improves to 0.57 cm for rms lower than 2 cm and 0.54 cm for rms between 2 and 4 cm with prior information on roughness. An overestimation of rms for rms lower than 2 cm and an underestimation of rms for rms higher than 2 cm are observed. The results of the accuracy of the synthetic and real data showed that the X band in HH polarization has a very good potential to estimate the soil roughness. (C) 2019 Published by Elsevier Ltd on behalf of COSPAR.