Semi-Supervised Hyperspectral Band Selection Based On Labeld Sample Expansion

The rich spectral information in hyperspectral images can bring great benefits to the application of various fields of social life,but the too high spectral dimension also makes data storage and subsequent data processing face certain challenges.Therefore,it is necessary to minimize the spectral dimension of hyperspectral images while preserving the original information of hyperspectral data.The band selection method is widely used in the dimensionality reduction technique of hyperspectral images because it retains the physical meaning of the original data.According to the criterion of whether the labeled samples are used in the band selection process,the band selection method is divided into three types:supervised,semi-supervised and unsupervised.In hyperspectral images,the acquisition of labeled samples requires a lot of manpower and resources.In order to overcome the problem of limited label samples in hyperspectral image preprocessing,this thesis makes full use of the concept of pseudo-label in the process of band selection,and studies the semi-supervised band selection method using labeled samples and unlabeled samples.The pseudo-tag is not the actual tag given in the data,but the tag predicted by the classifier trained by a part of the tag sample,which solves the problem that the tag sample is missing in the encapsulated band selection method to some extent.The paper mainly studies the following aspects.1.The encapsulated method is rarely used in hyperspectral image band selection methods because its accuracy is limited by the number of labeled training samples.According to the characteristics of local smoothness of hyperspectral image,this paper proposes a simple and effective packaged semi-supervised band selection method.Among them,the edgepreserving filter can be used to improve the performance of the pixel-level classification map,and the improved classification map can be used to evaluate the performance of the band subset.The method is characterized by the simultaneous use of rich unlabeled samples and valuable marker sample information.The experimental results in the actual hyperspectral dataset show the effectiveness of the proposed method.2.In the hyperspectral image classification task,dynamic classifier selection refers to selecting the best base classifier according to the characteristics of different test samples,wherein the base classifier refers to a group of classifiers trained using different criteria.In this paper,the concept of dynamic classifier selection is applied to the band selection method,and a packaged semi-supervised band selection method is proposed.In the proposed method,the band selection is converted to the choice of the base classifier.Specifically,the filtering process of the support vector machine classification map has provided high quality pseudo tags,and the K-nearest neighbor method is used to define local regions.Finally,the algorithm in the text serves as a selection criterion to select the band with the best performance.The experimental results of the three real hyperspectral data sets also show the practicability of the proposed method.3.High spectral band selection essential task is to choose the appropriate band subsets can retain the original data information,to a great extent because of hyperspectral data collection of high dimension,if calculated the permutation and combination on the original data set directly,then the amount of calculation is too big and cannot be used in practical application,this paper puts forward a kind of wavelength selection method based on the grouping and permutation and combination,with suitable clustering methods to group first,on the basis of the group to choose the appropriate band subsets,this method considers the mutual influence between band and at the same time the computational complexity of high dimension.