Soil Moisture Retrieval From Remotely Sensed Data: Neural Network Approach Versus Bayesian Method
Neural network (NN) approaches and statistical methods, based on a Bayesian procedure, are applied and compared in soil moisture (SM) retrieval from remotely sensed data. The principles and the practical implementations of Bayesian procedures and NNs are briefly discussed in terms of the advantages and disadvantages of each. Experimental tests are carried out by using the same set of training and test data for each method. The methodologies have been applied to two sets of data to retrieve SM from bare soils and to verify their accuracy. One data set contains scatterometer and radiometer data acquired on a variety of agricultural fields in different polarizations, frequencies, and incidence angles. The other is made up of five experiments carried out with a C-band scatterometer on rough and smooth soils at different polarizations and incidence angles. There are significant similarities in the performance of each method; they both retrieve the same features and trends in the analyzed data sets. Algorithm performances change according to SM level and data configuration. The main difficulties are found in retrieving low SM values, and in this case, the error on estimates is reduced when the data with two polarizations or two incidence angles are inserted in the inversion procedure. One major difference between the methodologies is that the NN performance improves, with respect to the Bayesian method, when more inputs are presented as two polarizations or two incidence angles in the training phase.