Retrieval of soil moisture content from radar backscatter coefficients

Presents a neural network for inverting radar backscatter coefficients into soil volumetric moisture content. The soil surface is characterized by a Kirchhoff rough surface. The permittivity of the soil is estimated from the moisture content using a semi-empirical model from Dobson et al. (1985). Radar backscatter coefficients are simulated from a forward model by invoking the scalar approximation to the Kirchhoff rough surface formulation. The limits of the applicability of such model are studied. Using a wide range of soil parameters, about 500 backscatter patterns are generated. The simulated data are used in a back-error propagation neural network which constitutes an inverse scattering model, for soil moisture content. Random noise is introduced to the test sets of backscatter coefficients to simulate measured responses. The results of the inversion are presented. Various issues such as the optimum network configuration, the optimum learning rate and momentum rate, the optimum training iterations and the problem of overtraining are also investigated