Polarmetric SAR Image Terrain Classification Based On Metric Learning

Polarmetric Synthetic Aperture Radar(SAR)imaging systems can detect different scattering data of target based on different manners of receiving and transmitting electromagnetic wave.Therefore,polarimetric SAR images contain abundant target polarization information,which makes it possible to classify images accurately.However,there are still problems in the polarimetric SAR image classification.On the one hand,With the increase of data scale brought by the development of imaging technology,obtaining a large number of labeled samples requires labor-intensive and time-consuming works;On the other hand,the radar imaging mechanism causes inherent speckle noise,which has an impact on image classification.It's a difficult problem in current polarimetric SAR image terrain classification that how to achieve accurate image terrain classification under a small number of labeled samples.The measurement of the distance between pixels of an image is used to implement the image terrain classification in most existing unsupervised/supervised method.Thus,this study is motivated by the aim for finding good metrics and achieving accurate and robust polarmetric SAR image terrain classification under a small number of labeled samples.Focusing on this issue,the main contents are as follows:1)We design a polarimetric SAR image classification method via Divergence-Chebyshev neighbors pursuit.The influence of noise can be reduced by designing the Divergence distance between polarimetric SAR pixels in feature space and the Chebyshev distance between pixels in space location and combining information of feature and pixel spaces.By the two-stage neighbor pursuit,the most similar feature neighbors are found in the spatial neighbors,which has a low-complexity character.Experiment results on six polarimetric SAR datasets show that the proposed metric is superior to the existing statistical distribution distances,and can maintain spatial consistency.Under a small number of labeled samples,the method can achieve fast and accurate polarimetric SAR imagery classification.2)We design a polarmetric SAR image classification method based on Triplet deep metric network.By constructing a Triplet dataset,the training sample set is expanded;Three concentric neural networks with the same weight sharing and the same structure are used to construct a Triplet deep metric network;we design corresponding Triplet loss functions of the same class and different class to learn strongly discriminative nonlinear metric subspace.Experiments on six polarimetric SAR data show that the method can learn strongly discriminative measures and achieve accurate polarmetric SAR image classification under a small number of labeled samples.3)We design a polarmetric image classification method based on Triplet deep adversarial metric network.Firstly,Triplet dataset are constructed to realize the training sample set expansion.We design the generator loss function and the Triplet discriminator loss function and construct adversarial metric network.The discriminative nonlinear metric subspace is dynamically learned in the adversarial training.Experimental results on six polarimetric SAR datasets show that the auxiliary training of the generator can enhance the discriminability of the extracted features of the Triplet discriminant network.The classification results are improved compared to the Triplet deep metric network under a few labeled samples.