Neural network adaptive algorithm applied to high resolution C-band SAR images for soil moisture retrieval in bare and vegetated areas

In general algorithms for soil moisture retrieval from high resolution satellite data cannot be easily extended to areas where they have not been calibrated and validated. This paper presents the application of an innovative approach for the detection of soil moisture from high resolution SAR images in order to overcome this main limitation by introducing a priori information. During the training phase, extensive data sets of SAR images and related ground truth on four areas characterized by very different surface features have been analyzed in order to understand the ENVISAT/ASAR responses to different soil, environmental and seasonal conditions. From preliminary analyses, the comparison of the backscattering coefficients in dependence of soil moisture values for all the analyzed datasets indicates the same sensitivity to soil moisture variations but with different biases, which may depend on soil characteristics, vegetation presence and roughness effect. These bias values have been used to introduce an adaptive term in the electromagnetic formulation of the backscattering responses from natural bare surfaces. The simulated data from this new model have been then used to train a neural network to be used then as an inversion algorithm. Preliminary results indicate an improvement in the accuracy of soil moisture retrieval with respect to the use of a traditional neural network approach. The results have been also compared with the estimates derived from the application of a Bayesian approach. © 2010 Copyright SPIE - The International Society for Optical Engineering.