Research On Convolutional Neural Network Based PolSAR Image Classification

With the special remote sensing principle,polarimetric synthetic aperture radar(PolSAR)can observe the earth all-day and all-weather,which is of great significance to military monitoring and situation awareness.As a classic and active research topic in the field of PolSAR image processing,classification is the most basic and important part for PolSAR image interpretation.Therefore,it has high research value.In recent years,convolutional neural network(CNN)has been widely applied to PolSAR image classification,and its research level has been developed in this process,but it also shows some shortcomings.After analyzing relevant researches,we find that for supervised image classification,traditional CNN-based classifiers are difficult to use the rich polarimetric information in PolSAR images,and there are many hand-crafted parts in the model.These factors limit the further improvement of classification accuracy.At the same time,although the classification performance of supervised CNN-based classifiers is satisfactory when the training samples are sufficient,it degrades rapidly in few-shot and cross-scene environments.Aiming at the above problems,in this thesis,we make targeted improvements to the network model and training algorithm of CNN from the perspective of supervised and weakly-supervised learning to improve the classification performance when training samples are sufficient and insufficient.The research content of this thesis can be divided into the following aspects:For the condition that the number of labeled training samples is sufficient,this thesis focuses on model design problems based on mature and effective training algorithms in supervised learning,and improves the supervised PolSAR image classification performance of CNN by optimizing the architecture and parameter of the model.For the model architecture design problem of supervised classifiers,a siamese network architecture is proposed,which learns deep representations from source data and polarimetric features of PolSAR images simultaneously,and makes comprehensive use and fusion of multi-stage information to improve the classification performance.For the model parameter design problem of supervised classifiers,an automatic parameter search method is proposed.By analyzing the key factors affecting the PolSAR image classification performance,combined with the two-stage search strategy developed in the continuous search space,the automatic parameter search method with low computational complexity is studied,so that the classification capacity of existing model architectures can be improved plug-and-play.Experimental results show that the two proposed methods can not only improve the classification accuracy,but also have good complementarity with each other.For the condition that the number of labeled training samples is insufficient,this thesis focuses on the training algorithms of CNN.Specifically,for two typical scenarios when lacking labeled data,i.e.,few-shot and zero-shot,we study the weakly-supervised algorithms which use other data than the data to be tested to aid the model training,so as to make CNN effective when the number of labeled training samples is insufficient.For the few-shot problem,this thesis proposes a pre-segmentation and contrastive learning based few-shot PolSAR image classification method.Considering the existing contrastive learning methods have some deficiencies when applied to PolSAR images,which leads to the performance bottleneck.We start with the particularity of PolSAR: On the one hand,a presegmentation based sample collection method is proposed to mitigate the sampling deviation of PolSAR images.On the other hand,we propose a transductive model pre-training method on basis of the polarimetric information based sample augment.Therefore,the proposed method can better mine the usable information in unlabeled data,and ensure the generalization performance in the few-shot environment.For the zero-shot problem,this thesis proposes a cross-scene zero-shot PolSAR image classification method based on polarimetric pseudo-label and adversarial doamin adaptation.The training is not only based on the domain discriminator to extract domain-invariant features with the help of adversarial learning,but also introduces the category semantics obtained based on image polarimetric information to extract the feature with discriminability.The two aspects jointly guide the model to extract shared features under the same category in existing and testing data,so as to achieve better zero-shot classification.The experimental results on multiple datasets show the effectiveness of proposed methods in the case of insufficient training samples.Through the above research,this thesis gives supervised and weakly-supervised solutions according to the requirements of PolSAR image classification in different environments.In the proposed methods,how to take advantage of the polarimetric information provided by PolSAR images,and how to fully integrate task characteristics with deep learning,are principles that we follow in our research.On this basis,the methods with better pertinence and performance are constructed to ensure that satisfactory deep learning based classification results can be obtained in different environments.