Cluster Algorithm Based On Density Peaks For Hyperspectral Image Classification

In recent years,hyperspectral remote sensing technology is developing rapidly,which has attracted the attention of many researchers.Compared to the traditional multispectral remote sensing,Hyperspectral remote sensing demonstrates some obvious advantages.Its spectral resolution has been greatly enhanced while retains a higher spatial resolution which provides the possibility for fine classification and object recognition.However,it also brings a higher processing complexity.Moreover,it also creates challenges to the later image processing and classification with a series of problems,such as insufficient training samples,strong correlation between bands,and the variance within different classes problem.Therefore,it is an urgent issue that how to overcome the interference of the above problems so as to dig out the information quickly and accurately and achieve high-precision classification.On the basis of analyzing the current research situation,two kinds of problems in hyperspectral classification are studied and validated on three different hyperspectral data.The specific work of this paper is as follows:(1)For the spectral-metric-based methods,in view of the fact that the traditional spectral matching method and extension algorithm can not adapt to the intra class variation,a method based on density peak clustering and spectral matching fusion for hyperspectral image classification was proposed in this paper.Firstly,the hyperspectral image was pre-process.And then the cluster center was quickly found by the density clustering method,and the remaining pixels were divided into blocks,After that,the objects were identified by matching the spectra of clustering center and the spectra library.(2)In view of the fact that many classification methods require sufficient training samples,a method based on density peak clustering was proposed to extract samples and classify hyperspectral images.First of all,the two typical methods based on spectral features and statistical features,including spectral matching method,neural network method and support vector machine method,were analyzed.Experiments show that spectral matching method is not sensitive to sample size,while the classification accuracy of neural network and support vector machine depend on sufficient sample to some extent.Therefore,the clustering method based on density peak was adopted to classify the sample points around the cluster centers into training samples.The experimental results show that the accuracy of the neural network classification and support vector machine was greatly improved with the new added samples.It can be concluded that the sample points selected by the method based on density peak clustering can be used to increase training samples,thus effectively solving the problem of small samples or g hardly obtaining samples.