Semi-supervised Classification Of Hyperspectral Images Combined With Convolutional Neural Network And Conditional Random Fields

Hyperspectral images contain rich spectral information,which can describe subtle differences of surface materials,and thus it has a high application value for accurate classification of ground objects.However,the high dimensionality of hyperspectral images,insufficient training samples,and information redundancy have brought challenges to traditional hyperspectral image classification methods.Researchers generally believe that the extraction of spectral-spatial(SS)features is essential to improve the accuracy of hyperspectral classification.Deep learning classification methods based on convolutional neural networks(CNN)has been proved to be able to better extract SS features,but traditional deep learning methods still need to balance the limitation of samples and network depth.To effectively exploit SS features of hyperspectral images to achieve accurate classification and overcome the problem of limited samples,this paper has done the following work:(1)A classification model combining depthwise separable convolution and residual network(Res Net)is proposed.The number of parameters in this model remains in a low level as the network gets deeper,and the deeper network is useful for SS features extraction.Another Res Netbased classification model with full convolutional structure(DFRes)also be proposed,which takes the above network as the baseline.DFRes takes the full images as input and output the predicted maps in real time,and overcomes the problem of small receptive field and data redundancy of general CNN-based methods.(2)To solve the problem that general CNN-based methods are less prone to preserve details and samples are limited,a semi-supervised classification model is designed based on the proposed DFRes,which introduced the convolutional conditional random field(Conv CRF)and region grow(RGW)strategy and is called DFRes-CR.More accurate soft labels are obtained from training DFRes and Conv CRF,which are input into RGW to implement the expansion of pseudo samples so as to finish the semi-supervised training.(3)Instead of traditional 4 or 8 neighborhood,the adaptive neighborhood is used to build RGW to overcome the problem that invalid pseudo samples increase with iterations,and improve the proposed semi-supervised method DFRes-CR.In this paper,we design a network to measure spectral similarity,which can provide a basis for the determination of adaptive neighborhood by learning the similarity between selected neighborhood pixels and the center pixel.In this paper,three hyperspectral datasets are selected to experiment on the proposed two supervised classification methods and one semi-supervised classification method and its improved version.It is verified that proposed methods have better performance than other classification methods.