Hyperspectral Imagery For Anomaly Detection Research Based On Analysis Of Regression

Hyperspectral imagery is a three-dimensional data cube whose length and width correspond to the spatial dimension of the image,while the height corresponds to the spectral dimen-sion and is a remote sensing image with a good combination of map and spectrum proper-ties.Hyperspectral imagery can provide tens or even hundreds of continuous spectral that are narrow.Compared to multispectral images,hyperspectral imagery have higher spectral resolution and more abundant image and spatial information.Anomaly target detection with spectral differences from the surrounding environment without prior information becomes a wide research hotspot in the field of hyperspectral imagery for target detection,and has important applications in military,agriculture,and forest fire monitoring.Hyperspectral imagery anomaly detection method based on sparse representation and col-laborative representation are a popular anomaly detection methods in recent years.The main idea of these methods is that to linearly represent the spectral of the pixel to be detected with the spectral of the surrounding pixels to obtain the pixel to be detected.Spectral estimation,and then determine whether or not the pixel to be detected as an anomaly pixel based on the residual error and a preset threshold,and achieved good results.The linear representation method ignores the non-linear information implied in hundreds of bands of hyperspectral imagery.In order to fully utilize the rich nonlinear information provided by hyperspectral imagery,this paper mainly studied the hyperspectral imagery's kernel model for representa-tion dictionary design and coefficient solving.Collaborative representation model.In view of the manual selection of nuclear parameters in many methods in the past.Firstly,this pa-per studied the nuclear collaborative representation model that adaptively selects nuclear parameters as the background changes.Secondly,we studied the effect of nuclear cooper-ative coefficient on detection performance and finally consider the nuclear parameters and the coefficient.Based on this,a coefficient-normalized and adaptive kernel collaborative representation of hyperspectral imagery for anomaly detection model is proposed.The local background kernel parameters are determined according to the standard covariance of all the bands,indicating that the normalization of the coefficients can substantially enhance the separation ability of the target and the background,that is,whether or not the coefficients are normalized for the background pixels and its estimated residuals.The range of variation is basically the same;but for the anomalous pixel and its estimated residual,the coefficient is normalized and the range of residual variation is larger,which indicates that when the anomaly pixel is determined by the preset threshold,the coefficient is normalized and the residue is As the difference increases,abnormal pixels are easily detected.In addition,based on the hyperspectral spectrum and image space,the joint collaborative representation model for abnormal detection of hyperspectral images was studied,ie,anoma-ly detection was performed in the spectral and spatial domains,respectively.To synthesize spectral domain and spatial domain detection,this paper proposes an anomaly detection method for collaborative representation of air-spectrum combination based on band cluster-ing.To provide different intensity of abnormal target information for different band sets,this paper uses subspace clustering method to cluster all the bands of hyperspectral images to form several band sets.By calculating the standard deviation of each band set,the most important anomaly targets are extracted.The set of information bands,and then the de-tection of the spectral domain on the band set;in the spatial domain,this article uses the orthogonal subspace projection transform to perform background noise reduction process-ing,and then uses the denoised hyperspectral data to perform spatial domain detection;The spectral domain and spatial domain detection results were fused to obtain the final detection results.Numerical experimental results show that this paper has achieved better detection performance and effect than previous methods.