Retrieving surface roughness and soil moisture from synthetic aperture radar (SAR) data using neural networks

Abstract

An inversion technique based on neural networks has been implemented to estimate surface roughness and soil moisture over bare fields using European Remote Sensing (ERS) and RADARSAT data. The neural networks were trained with a simulated data set generated from the integral equation model. Later the networks were applied to a field data set spanning a wide range of surface roughness and soil moisture, with backscattering coefficients for three radar configurations (VV 23°, HH 39°, and HH 47°). Approaches based on two and three radar image configurations were examined and tested. Although the three-image configuration produces slightly more accurate results, a two-image configuration gives results of comparable accuracy when a favourable combination of incidence angles is adopted. The introduction of a priori information on the range of soil moisture (mv) improves mv estimation. Soil moisture and surface roughness errors were estimated at about 7.6% and 0.47 cm, respectively, using the root mean square error (RMSE).