Hyperspectral Image Classification Based On Self-supervised Learning

In recent years,hyperspectral image processing plays an increasingly important role in the field of remote sensing,and hyperspectral image classification,as an important part of it,plays a decisive role.Hyperspectral images,which contain rich information of the ground object,have a broad practical application value in the fields of military and civilian,so it has an extremely important research value.The goal of hyperspectral image classification is to accurately classify each pixel in the hyperspectral image according to the spatial and spectral features presented by the ground objects.Hyperspectral image is a comprehensive carrier containing rich radiation,spatial and spectral information.Abundant information makes ground object analysis obtain more opportunities,but at the same time makes classification face many difficulties such as information redundancy and dimensional disaster.In recent years,the application of machine learning and deep learning in the field of hyperspectral image has been fruitful.However,hyperspectral image has few prior samples,which poses great difficulties in supervised classification which relies heavily on the quantity and quality of samples.To solve this problem,this paper proposes a series of algorithms for hyperspectral image classification from the perspective of self-supervised learning,using the information from a large number of unlabeled samples.The main work of this paper is described as follows:Firstly,a method of hyperspectral image classification is proposed by using contrastive learning.The method trains the model in the way of self-supervised learning.In the pre-training stage,combined with data enhancement methods,a large number of unlabeled samples are used to construct positive and negative sample pairs for contrastive learning.The purpose is to enable the model to identify positive and negative samples.On the basis of the pretraining model,the features of hyperspectral image are extracted for classification,and a small number of prior samples are used to fine-tune the classifier.The experiment shows that the features extracted by the self-supervised learning approach achieve good expected results in the downstream classification task.Secondly,using the idea of clustering,combining deep learning with clustering,a hyperspectral image classification method based on deep clustering is proposed.Following the paradigm of self-supervised learning,the method takes clustering as an proxy task of selfsupervised learning and classification as a downstream task.The features extracted by the encoder are clustering in the deep neural network,and then the category labels obtained by clustering are used as ’pseudo labels’ to update the network parameters,and then the network is made to predict these ’pseudo labels’.Updating network parameters and updating cluster parameters are carried out in turn.The method focuses on the correlation between samples,extracts valuable feature by virtue of the powerful representation ability of deep neural network.At last,a small number of labeled samples are used for supervised classification.Finally,a hyperspectral image classification method based on multi-level feature fusion is proposed.The idea of feature fusion considers fusing features of different levels into more discriminative new features.This method continues the idea of self-supervised learning.On the basis of contrastive learning network,the features obtained from different convolution layers are combined together through a feature fusion module.The extracted new features have the multi-level feature of complementary deep and shallow layers,and perform well on the classification task.