Research On Remote Sensing Image Feature Extraction And Anomaly Detection Method Based On Spatial Density

The hyperspectral images(HSIs)have high-dimensional characteristics,high correlation between bands,and spectral mixing,which makes traditional feature extraction poor in characterizing complex spatial spectrum structures.Therefore,how to accurately and quickly obtain the key and subtle characteristics of the ground objiects is a difficult problem to be solved at present.This paper deeply explores the spectral continuity and spatial correlation of HSIs and further fuses the local,statistical and spatial geometric structure information in the image,to propose the remote sensing image feature extraction and anomaly detection method based on spatial density,which can solve the problems of unsatisfactory classification and detection accuracy of HSIs.The main research content of this paper is divided into the following parts:(1)The traditional classification method based on the covariance matrix only uses the spectral information of the HSI.There is still a lot of noise in the classification map,and the classification accuracy needs to be further improved.In order to explore the rich spatial information and dig deeper into the spectral correlation in the image,a density peak covariance matrix(DPCM)method is proposed for the HSIs.First,the density peak clustering(DP)algorithm is employed to calculate the density value of each pixel,thereby effectively converting the spectral information in the image into density information.Then the Local Covariance Matrix Representation(LCMR)method is used to represent the center pixel with the covariance matrix.Each off-diagonal entry in the covariance matrix represents the correlation between two different features,which introduces a way to naturally merge multiple features that may be related.Finally,the covariance matrix based on the Riemannian manifold space is converted to the Euclidean space using the matrix logarithm function,and then fed back to the support vector machine(SVM)for classification.The experimental results show that this method makes full use of the spatial information between pixels,and effectively avoids the impact of the same spectrum and the same spectrum on the classification accuracy.Moreover,its classification performance is better than other spatial spectrum feature extraction algorithms.(2)Extracting knowledge from the background and comparing spectral differences to identify abnormal pixels is the main strategy nowadays.However,due to the similarity of the spectral features of adjacent ground elements,the precise extraction of features is affected.Thus,there are still potential abnormal pixels similar to the background pixels in the background.Directly introducing background information for calculation will reduce the detection accuracy.Therefore,a HSI anomaly detection algorithm with spatial density information background purification(SDBP)is proposed.Specifically,the local density information is calculated by the DP cluster algorithm to extract the pure background set in a single window,thereby effectively removing potential abnormal pixels in the background.Then,the collaborative representation detector(CRD)is used to obtain the detection result of the test pixel.Considering that the neighborhood of each pixel is a homogenous region,a dual-window strategy is adopted to improve the above algorithm.This makes the background information more refined,reduces the false alarm rate,and optimizes the detection performance.Experimental results indicate that the proposed SDBP method can remove noise and enhance the separability between the abnormal target and the normal background,to achieve the purpose of extracting the pure background set from the image and separating the abnormal target.The method can greatly improve the detection accuracy.(3)Window-based operation is a general technique for hyperspectral anomaly detection.However,the problem remains that background knowledge containing abnormal information often affects the attributes of test pixels.Based on this problem,a new framework based on a dual collaborative representation(DCR)algorithm for hyperspectral anomaly detection is proposed.First,the low-rank and sparse matrix decomposition(LRa SMD)algorithm is used to extract the low-rank background matrix to obtain the background pixel set with low-rank attributes.Then,the density peak clustering algorithm is used to extract the pure background set in the double window.Meanwhile,the single-window strategy is also used to extract the neighboring pixels of the test pixels,and collaboratively represent the test pixels.Finally,a decision function based on the residuals of this dual-stage collaborative representation is utilized to detect abnormal pixels.Experimental results demonstrate that the proposed DCR method can effectively extract background pixel sets with low-rank characteristics,achieve the purpose of prominently highlighting the target and suppressing background information,making its detection accuracy better than other commonly used anomaly detection methods.In summary,aiming at the problem of feature extraction in HSI,this paper studies the spatial and spectral dimension structural features of hyperspectral data,and deeply excavates the space spectral correlation of HSIs.Thus,the remote sensing image feature extraction and anomaly detection method based on spatial density are proposed.Experiments prove the superiority and robustness of the proposed algorithm.