Research On Hyperspectral Remote Sensing Image Classification Algorithm Based On Joint Spatial-aware Collaborative Representation

Hyperspectral image(HSI)classification is one of the main methods for distinguishing between ground features and reading ground objects in remote sensing images and is also one of the hot topics in the field of remote sensing.HSI that usually contains abundant continuous narrow band and spectral information can be widely used in fields like mineral exploration,agriculture and environmental monitoring and so on.These applications promote the optimization and improvement of HSI classification algorithms.The classification algorithm based on collaborative representation(CR)uses labeled training samples to linearly represent the test pixels,and determines the class labels of the test pixels according to the representation residuals.There is no need to assume density distribution of any data nor need to train the classifier model with training samples.Therefore,CR algorithms have attracted much attention in the field of remote sensing image classification.However,traditional CR algorithms only input spectral information and spatial information as classification features into the model for classification,and do not optimize the classification model.Aimed at the problems that need to be optimized about the CR algorithm classification model,this paper proposes two HSI classification methods based on joint spatial-aware CR(JSaCR).To address the insufficienty of texture information-based classification features to classify samples.This paper proposes a texture regularized-based joint spatial-aware collaborative representation(TRJSaCR).First,the algorithm obtains denoised hyperspectral dataset,HSI coordinate values and texture prior data.Second,the texture feature is added as a regular term to the constraint expression coefficient in the objective function of JSaCR,and the test sample is reconstructed by finding the analytical solution of the objective function.Finally,the class labels of the test pixels are determined according to the representation residuals to achieve classification.Experimental results show that TRJSaCR has better classification effect than JSaCR algorithm,and is more universal.For classifiers,the problem of misclassified pixels still exists after classifying images.In this paper,spatial information-assisted discrimination rules(SIDR)method is added to the TRJSaCR algorithm.This algorithm combines SIDR in the post-processing of TRJSaCR,and further proposes a SIDR coupled with TRJSaCR(TRJSaCR-SIDR)for classification method.First,the algorithm encapsulates its neighborhood spectral information according to the test sample after splitting the dataset.Second,more precisely,the label information of the test samples and their corresponding neighborhoods are specified by TRJSaCR-SIDR and the final labels are determined by considering their neighborhood label distribution.Experimental results show that compared with SVM,CR,SR,NRS,SaCR,JSaCR,and TRJSaCR algorithms,the TRJSaCR-SIDR algorithm has the best classification effect.The innovation of this article is to improve the JSaCR algorithm from the texture feature as a regular item,and for the problem of misclassified pixels after classification,the JSaCR algorithm is combined with the texture regularization to improve and optimize the algorithm during and after the classification.Significance of research work,using spatial information to expand the idea of JSaCR classification method in HSI classification.Even when the training sample is a small one,the accuracy of the classification results can still be guaranteed.Experimental results based on two benchmark hyperspectral datasets,Indian Pines and Pavia University,indicate that the proposed algorithms are superior to other state-of-the-art classifiers.