Research On Hyperspectral Image Classification Based On Data Gravitation

Hyperspectral image(HSI)usually has dozens or even hundreds of bands,and rich spectral information can provide the basis for accurate recognition of ground targets.Therefore,hyperspectral images are widely used in agricultural monitoring,environmental management and other fields.As the basic work of remote sensing analysis and application,hyperspectral image classification has become an important research hotspot.However,traditional image classification methods face many problems and challenges,such as heterogeneous pixels and outliers in the image itself,and limited labeled samples,etc.,which will restrict the classification performance of hyperspectral images.In recent years,inspired by Newton’s universal gravitation,Data gravitation based classification(DGC)has been proposed to deal with the binary classification problem of UCI data sets.This method simulates the universal gravitation between data points,which is called data gravitation,and uses it as a similarity measure to guide the classification process.Data gravitation is directly proportional to the mass of data points and inversely proportional to the square of distance between data points.Therefore,there is often a larger data gravitation between similar data points.DGC overcomes the defect of treating data with the viewpoint of local relation in traditional data classification,and shows the advantage of data gravitation aggregation,which makes it possible to solve the classification problem of hyperspectral images.However,the advantages of DGC in binary classification is not directly suitable for multi-label hyperspectral data.In addition,it is difficult to combine data gravitation with the spatial information of hyperspectral images.Therefore,how to effectively use the characteristics of data gravitation for hyperspectral classification is a challenge.To solve the problems of limited labeled samples,heterogeneous pixels and outliers in hyperspectral images,a hyperspectral image classification method based on data gravitation is proposed based on the characteristics of " image-spectrum merging " of hyperspectral images,the effectiveness of the proposed method is verified on several hyperspectral data sets.It mainly uses different data mass definition ways to comprehensively utilize the spatial-spectral information of hyperspectral images,and then constructs joint data gravitation for classification.The main research work of this paper is described as follows:(1)In the classification of hyperspectral images,it is difficult to make full use of the spatial information of the labeled training samples.Therefore,how to make effectively use of the spatial-spectral information of the hyperspectral images to improve the classification accuracy is particularly important.In response to this problem,this paper proposes a bayesian gravitation-based classification(BGC)method for hyperspectral image classification,which effectively utilizes spatial context information and prior information about the spatial distribution of training samples.Specifically,the test sample is assumed to be a mass object with a unit volume and a specific density,and based on bayesian theorem,the data quality is expressed as a combined function of the spectral distribution of its neighborhood and the spatial distribution of the surrounding training samples.On this basis,a joint data gravitation model is established for classification.This method can effectively mine the prior information of training samples and successfully integrate the spatial-spectral information of hyperspectral images.The proposed algorithm was validated on four benchmark datasets,Indian Pines,Salinas,Pavia University and Grss＿dfc＿2014.Comparative experiment results show that BGC has obvious advantages in high-resolution hyperspectral image classification,and shows flexibility in the classification with limited samples.(2)Traditional joint learning models,such as joint sparse representation(JSRC)and joint nearest neighbor(JNN),are based on the assumption that neighboring pixels are completely homogeneous,and that all spatial neighbors have the same weight.This assumption is unreasonable,because when there are heterogeneous pixels or outliers in the neighborhood,the same weight will restrict the classification accuracy.In response to this problem,a local similarity-based data gravitation classification method(LSDGC)is proposed.LSDGC assigns a local data mass to each pixel and constructs joint data gravitation to suppress the interference of noise and outliers.Specifically,the local data mass of pixels in the neighborhood is defined by the cosine similarity with the center pixel,and pixels with lower similarity contribute less to the local quality.The definition of local quality can effectively alleviate the interference of heterogeneous pixels and outliers on classification.Experiments on three benchmark data sets,Indian Pines,Pavia University,and Grss＿dfc＿2014 verified the effectiveness of LSDGC antiinterference。(3)To the problem of the limited number of training samples,a training set construction method based on the connection center evolution(CCE)is proposed.The connectivity in CCE reflects the local similarity of neighboring pixels,and the representative pixel are selected in the highly homogenous neighborhood and added to the training set,and then build an narrowing the size of neighborhood iteratively construction way to fully mine more potential training samples.The experimental results show that the training set based on CCE has high quality and can be significantly improve the classification performance of the classifier.Combining the constructed training set with LSDGC,the experimental results on the three data sets of Indian Pines,Pavia University and Griss＿dfc＿2014 show the robustness of CCE-LSDGC for classification under limited samples.