Researches On Hyperspectral Image Classification By Semi-supervised Learning And Deep Learning

With the rapid development of satellite imaging technology,hyperspectral images(HSIs)are widely obtained and applied.HSIs are usually composed of hundreds of spectral bands,with spectral resolutions up to nanometers,and are extremely rich in spectral information and spatial structural features.Therefore,HSIs are widely used in many fields such as precision agriculture,ocean monitoring,military investigation and many other fields.Among them,HSI classification technology has always been a research hotspot in the field of HSI processing because it can achieve accurate recognition of ground objects.Recently,with the deep learning technology gradually applied in the field of image processing and achieved remarkable results,a large number of deep networks are used to solve the related problems of HSI classification.However,the traditional classification network for HSIs is usually a single path structure.With a limited number of network layers,it is difficult for single path classification network to extract sufficient deep semantic features.In addition,the classification method based on deep learning requires a large number of labeled samples to train the network model.For HSIs with small samples,the deep network cannot obtain the optimal parameter model,and the network is prone to overfitting,which leads to the result of misclassification.In view of the above difficulties,this paper proposes two HSI classification algorithms based on deep learning.The main research contents of this article are as follows:1)Aiming at the problem that the traditional single path network is difficult to extract sufficient deep features of ground objects under the limited layer network,this paper proposes a HSI classification method based on the deep dual-path network.In this method,a dual path connection block is constructed by sharing features,and effectively combines the residual block and the dense connection block to form a dualpath network structure,so that the residual network and the dense connection network are combined to learn features.It can realize the reuse of features and constantly exploring discriminative HSI features,which effectively improves the performance of network classification.2)Aiming at the problem that the deep network is prone to overfitting with a small number of training samples,this paper proposes a HSI classification method based on semi-supervised deep learning.Firstly,the multi-scale edge preserving feature and multi-attribute profile feature of HSI are extracted by multi feature learning,and the corresponding classification probability results are obtained by support vector machine(SVM)classification based on different features,so as to prediction for a large number of unlabeled samples.Then,the classification probabilities are sorted and filtered by joint probability decision,and a high confidence sample set is selected from the predicted unlabeled sample set to expand the training set.Finally,a deep feature fusion network is trained based on the acquired extended training set.Experimental results show that the proposed semi-supervised learning strategy can better alleviate the overfitting problem of deep networks with small samples and effectively enhances classification performance.3)According to the practical application requirements,this paper applies the above two classification algorithms to the real hyperspectral data set of Changsha City,Hunan Province,which is obtained by Gaofen-5,and makes a detailed interpretation analysis.