Polarimetric SAR Image Terrain Classification Based On Multi-layer Neighborhood Preserving

Polarimetric synthetic aperture radar(Pol SAR)has unique imaging feature of all-weather,all-day,multi-band,multi-polarization,and can provide large area high resolution images.Further comprehensive study on Pol SAR image post-processing and interpretation can extract more valuable information for environmental monitoring,earth resource survey and military systems,etc.Deep learning is a new field in machine learning research.Its motivation lies in the establishment and simulation of the human brain to analyze and study the neural network.The deep learning method can extract the more abstract and intrinsic features of the target data automatically.The multi-layer deep network has been applied to Pol SAR processing.However,the traditional deep network model is complex,the parameter adjustment is difficult,and is not necessarily fit for Pol SAR data.Moreover,for most of the supervised network model,the training of the network model is insufficient when there is very few labeled samples.Based on the conventional models of deep learning method,this thesis proposes three novel deep networks using neighborhood preserving for Pol SAR terrain classification.A deep learning method based on sparse filtering and neighborhood preserving is proposed.This method combines sparse filtering with neighborhood preserving,and extends the idea of deep learning to construct a novel deep sparse filter network.Compared with the traditional depth learning network,the network has less adjustment parameters,and the neighbor relationship of the data is preserved.The neighborhood preserving regularization is used to optimize the network parameters with labeled samples and partially unlabeled samples.The experimental results show that the feature learned by the depth network is more favorable for the classification of Pol SAR images.The accuracy is greatly improved.A deep learning method based on distance metric learning is proposed.This method first adds the popular learning regularization term on the basis of the large margin nearest neighbor algorithm(LMNN),which overcomes the flaw of the supervised distance measurement method when the labeled samples are insufficient.And then through thegreedy layer-wise pre-training and the fine-tuning of the back-propagation algorithm.It creates a semi-supervised depth distance measurement network.Demonstrated by a large number of experiments,the deep network can effectively improve the terrain classificationaccuracy of Pol SAR data.A new classification method based on convolution neural network for Pol SAR data is proposed.As the demand of labeled samples for the traditional convolution neural network is large,a semi-supervised convolution neural network is designed,and the convolution layer is carried out by sparse filtering and neighborhood pre-training,and then a small number of label samples are utilized to fine-tune the network at the end of the network,which aims at enhancing performance of convolution neural network.Moreover,unlike the conventional fixed window based division method,a new scheme which uses the padded superpixel is developed.The experimental results show that the semi-supervised convolution neural network can obtain superior Pol SAR terrain classification accuracy in the case of fewer labeled samples.