Soil parameters retrieval from remotely sensed data: efficiency of neural network and Bayesian approaches

Abstract

Six remote sensing experiments are analyzed in order to study the feasibility of soil parameters extraction from active and passive microwave data. The inversion process has been carried out through two methodologies: a Bayesian and a neural network approach. Two different sets of data have been analyzed: one experiment with active and passive data on a smooth soil and five experiments carried out with a C-band scatterometer on rough and smooth soils at different polarizations and incidence angles. In the case of active and passive data, using a Bayesian algorithm, the correlation coefficients between the extracted and the measured values of soil moisture are R=0.83, R=0.84 and 0.72 for the three analyzed data configurations. In the neural network approach, the correlation coefficients are R=0.72, R=0.83 and 0.79. The best performance is achieved when two different frequencies, 4.6 GHz for active data and 2.5 GHz for passive data are employed where the neural networks produce the lowest errors in the estimates. For the second group of data, the neural network makes fewer mistakes and overestimates only the values of epsiv that originated from backscattering coefficients acquired on the rougher field. The Bayesian approach tends to overestimate the values of epsiv with an average bias of 5%