The Mlp Radar Parameter Retrieval And Classification Of Sar Images

This paper describes the application of neural networks to surface parameters retrieval and targets classification from multi-polarization ENVISAT-ASAR datas and AirSAR datas. It is an important advancement to use neural networks to perform inversion and classification in Remote Sensing recently. The combination of a scattering model (SM) and neural networks make it possible to perform inversion and classification accurately and in real time. The used neural network is multilayer perceptron (MLP) with fast learning (FL), which is fully interconnected network. Simulated data sets based on the Integration Equation Model (IEM) are used to train the neural network. Accordingly, the training data sets may be viewed as taken from a completely known randomly rough surface. The input to the neural network is the set of values (σ0(θ)) with angles and polarizations, the output of the neural network is the set of surface scattering parameters. The layer permittivity (ε), surface correlation length (kl) and surface roughness (kσ) are retrieved from σ0(θ) using the trained MLP.As above, this method can be used into the classifier. For this aim, we suggest the proposed fully interconnected MLP with FL, in which the training data sets are values (σ0(θ)) with polarizations from some identified targets. The trained neural networks is used in target classification in the ENVISAT-ASAR datas and AirSAR datas. And finally, the results of proposed method are compared with that of the unsupervised classification one, the in situ test data are from Zhaoqing in Guangdong Province and Taichung in Taiwan Province in China.