Retrieval of bare soil surface parameters from simulated data using neural networks combined with IEM

Many attempts have been made to retrieve soil surface parameters, such as soil moisture(SM), surface roughness parameters, by regressions or other statistical methods and some other techniques like Neural Networks (NN) and genetic algorithm. And NN is proved to be an effective method for retrieval problem; much effort has been devoted to it. In this study, our goal is to estimate the bare surface soil moisture and surface roughness at Advanced Microwave Scanning Radiometer (AMSR/E) frequencies. First, a preliminary analysis was conducted based on a sensitivity analysis of surface parameters to surface parameters by simulating AMSR/E emissivity data of V, H polarizations, which were generated by the Integral Equation Model (IEM) for AMSR/E viewing angle of 55 degree. We employed NNs to be first trained with part of the sensitive data determined by the above sensitivity analysis, then the trained NNs were used to retrieve the parameters that we need, especially soil moisture, from the simulated data. Analysis of the difference between the retrieved parameters and the simulated ones is presented here. In addition, because the retrieval accuracy of NN is supposed to be extremely sensitive to "noise"- the difference between the model and measurements, we introduced random noise to the simulated data. At the same time, we carried out a sensitive analysis about the input noise. We also selected the most sensitive frequencies 6.9, 10.7 to soil moisture in our retrieval scheme. This study demonstrates the great potential of NN in estimating soil surface parameters from passive microwave remotely sensed data again.