Hyperspectral Image Classification Based On Cascade Markov Random Field And Self-supervised Learning

In the field of remote sensing,hyperspectral image processing has occupied a vital position for many years,and in the hyperspectral field,hyperspectral image classification is the most popular topic.There are many spectral bands in hyperspectral images,which contain abundant ground object information.It is widely used in many fields,including agriculture,military and environmental monitoring,which is of great practical significance to the development of human society.Since the birth of remote sensing technology,higher spectral resolution has been the goal of researchers in various countries,and spectral images have undergone the evolution from multispectral to hyperspectral.From the beginning of a dozen bands to the present hundreds of bands,the information contained in hyperspectral images has also increased significantly,which provides more basis and opportunities for the analysis of ground object targets based on spectral data.However,at the same time,the classification of hyperspectral images is faced with difficulties,with the increase of the number of bands,not only increases the cost of classifier training,dimension disaster can not be ignored,which puts forward new requirements for the development of hyperspectral image classification.In recent years,with the rapid development of machine learning,the field of hyperspectral images has also ushered in new opportunities.Many methods based on traditional algorithms and deep networks have been proposed and applied in the field of hyperspectral images,which have achieved certain results,but also left some problems.Based on the current status of hyperspectral image classification,this paper proposes an algorithm that is more suitable for hyperspectral image classification,aiming at the problems of less hyperspectral data samples and unbalanced data.The main work of this paper is as follows:First,inspired by the cascade forest,this paper fuses the excellent image segmentation model algorithm of markov random field(MRF)into the cascade structure,and proposes a cascade markov random field applied to the classification of hyperspectral images.In this algorithm,the prediction vector generated by support vector machine(SVM)is combined with the MRF model,and then the fused feature vector is splited with the original spectral feature to obtain the enhanced feature vector with stronger recognition ability.In subsequent phases,the enhanced eigenvectors are used as the next level input to the cascading MRF model.By fusing multiple MRFS into a cascade framework,the utilization efficiency of MRF on hyperspectral data information is improved,and the classification accuracy of samples is effectively improved.Then,SFT is applied to hyperspectral image processing.Band-by-band filtering of hyperspectral data can introduce spatial features into the sample features and effectively remove noise information,which is a common method in hyperspectral image processing.However,most filtering methods will put the sample point in the center of the filter window during operation,which will cause a part of the edge information to be lost.As the number of filtering increases,the image edge will become blurred.By placing the sample point on the edge of the filter window,SFT can better solve this problem.This chapter will explore the application effect of SFT on hyperspectral images,and provide more basis for the classifier training through the multi-layer filtering of data,which effectively reduces the difficulty of training the classifier and improves the classification effect.Finally,a hyperspectral image classification method based on self-supervised learning is proposed.Self-supervised learning is one of the most popular methods in deep learning.In this paper,self-supervised learning is applied to hyperspectral image field to solve the problem of insufficient hyperspectral data samples.Through the construction of the selfsupervised auxiliary task,the middle layer feature vector is extracted as the enhancement feature of the classification task while the auxiliary task is completed.By changing the loss function of the deep network in the auxiliary task,several different enhancement features can be obtained.Finally,these enhanced features were fused and classified using SVM and MRF.