Research On Semi-supervised Classification Algorithm Of Hyperspectral Remote Sensing Image

With the rapid development of imaging spectrometers,hyperspectral remote sensing has become an important method for ground observation.Compared with multispectral remote sensing images,hyperspectral remote sensing images are characterized by a large amount of data,high correlation between bands,and difficulty in obtaining label samples.Using traditional remote sensing image processing techniques to analyze hyperspectral images faces many difficulties and challenges.Hyperspectral image classification technology is the main content of hyperspectral remote sensing data analysis.Traditional supervised classification requires enough labeled sample data as prior information.However,obtaining labeled samples requires a lot of manpower and material resources,which makes supervised classification methods are limited.Unsupervised classification is based on the premise that there is no prior information,and the corresponding classifier is constructed based on the hidden structure in the unlabeled data.As a result,he classification accuracy cannot be guaranteed.Semi-classification combines supervised and unsupervise classification,which makes up for the shortcomings of traditional methods and makes full use of a large amount of unlabeled data and a small amount of labeled data.Collaborative training and self-training are two commonly used algorithms in semi-supervised classification.Self-training is prone to mislabeling in the process of labeling samples.Collaborative training labels multiple unlabeled samples by training multiple classifiers.The small number of label samples leads to insufficient differences between the classifiers,which limits the classification accuracy of the collaborative training algorithm.In view of the above problems,this paper improves the collaborative training and self-training classification algorithms.The specific work is as follows:1.Aiming at the problems of insufficient difference between multiple classifiers and small initial label samples in collaborative training,a collaborative training classification algorithm combining active learning and differential evolution algorithm is proposed.Improve the original active learning algorithm to obtain a multi-criteria multi-edge sampling algorithm with support vector machine as the technical support,and introduce an adaptive operator in the search process of the differential evolution algorithm to select rich information from the set of unlabeled samples and different samples.The advantage of the aboveclassification algorithm is that through active learning of the sample selection strategy,better use of unlabeled samples and effective expansion of the training sample set can largely solve the problem of less prior information in traditional semi-supervised classification methods.The problem is to introduce strong differences for the three base classifiers in collaborative training through the differential evolution algorithm,which makes the decision result of the sample more accurate.The experimental results show that this algorithm effectively improves the classification accuracy.2.Aiming at the problem of mislabeling in self-training and insufficient mining of spatial information in traditional spatial spectrum classification,a self-training classification algorithm based on spatial-spectral clustering is proposed.The spatial texture information is extracted through bilateral filters,and the spatial boundary texture information is retained,and it is cascaded with the spectral information.In the spectral clustering algorithm,the spatial correlation information is fully introduced,a new method for constructing a spatial-spectral similarity matrix is proposed.The advantages of the above classification algorithm are that it breaks the limitation of classification based on spectral information only,makes full use of spatial-spectral information for classification,and uses a mechanism that combination probability-based support vector machines and spectral clustering algorithms during self-training.The mislabeled samples are effectively removed,and the experimental results show that this algorithm achieves higher classification accuracy.