Research On Hyperspectral Anomaly Detection Algorithm Based On Joint Spatial-Spectral Information

Hyperspcctral images are characterised by their unity of image and high spectral resolution,and they contain rich spectral and feature information.The abundant information in hyperspectral image gives important practical significance to target detection.Target detection of hyperspectral image is a hot research topic in the field of hyperspectral image processing at this stage,widely used in fields such as camouflage identification,resource exploration and agricultural monitoring.In hyperspectral target detection tasks,hyperspectral anomaly detection is an important branch,it does not require a priori information to be discriminated by the difference between the anomalies target from background pixels.For the deficiencies of hyperspectral data features and existing anomaly detection algorithms,this paper proposes the study of hyperspectral anomaly detection based on joint spatial-spectral information.The details of the study are as follows:(1)For the problem of low spatial resolution and blurred detai ls of hyperspectral images,a hyperspectral anomaly detection method based on local gradient profile transformation is proposed.This algorithm uses the gradient profile transform for the first time to guide the spatial enhancement of spatial information in the image,while applying the gradient profile transform preprocessing only on the locally potential anomaly points to reduce computational complexity.Firstly,some ppssible anomaly pixels are coarsely selected in the input hyperspectral image,and the the local anomaly pixels are processed using the gradient profile transformation algorithm.Secondly,the gradient profile transformation is performed on the located anomaly pixels,and the transformed gradient profile is used to guide the spatial information enhancement of the original hyperspectral image.Finally,the enhanced hyperspectral image is detected.By introducing local gradient profile transform in the process of enhancing hyperspectral image,the reduction in detection accuracy caused by loss of spatial structure details is avoided.Comparative experiments demonstrate the effectiveness of the algorithm in improving detection accuracy.(2)In order to address the problem of redundant and noisy bands in hyperspectral images,a band selection algorithm is introduced to increase spectral discrimination.Meanwhile,considering the characteristic that the proportion of anomaly pixels is small in hyperspectral images,a small number of anomaly pixels can be ignored in the spatial clustering process.By removing clustering results,the effect of highlighting anomaly target pixels can be achieved.This paper proposes a hyperspectral anomaly detection based on band selection and spatial clustering.Firstly,in order to avoid that the selected bands have similar information and unable to represent the entire spectral range,an iterative band selection strategy is applied in this method to divide the bands into subsets with a certain step size,so that the selected bands have high resolution and a certain degree of diversity.Secondly,a clean background image is constructed by ignoring a small number of anomalous pixels through spatial clustering operations,and the background clustering results are removed from the band selected hyperspectral images,thus highlighting the anomalies.Finally,the differenced image is detected.The detection accuracy of the algorithm is improved by spectral and spatial features respectively,which improves the detection performance of the algorithm.(3)For the problems of anomalies sample contamination and redundancy of spectral information in hyperspectral images,a hyperspectral anomaly detection based on background cleaning and feature extraction is proposed.The algorithm introduces a domain transformation strategy to enhance the discrimination between anomalies and background pixels,while dividing the subspace by a sliding window strategy to reduce spatial correlation and select highrepresentative and low-redundancy band sets.Firstly,the sample features with more differentiation are extracted by combining with domain transformation processing,and the lowrank sparse decomposition using "row constraint" to obtain a low-rank background matrix which constitutes relatively clean background information,and the anomalies are highlighted by eliminating the background information.Secondly,a sliding window strategy is adopted to select the most representative bands by linear dimensionality reduction in the local area.Finally,the final anomaly detection result is obtained for local area anomaly detection.The algorithm can well suppress the interference of background information and can use of the spectral information,thus improving the detection accuracy of the anomaly detection algorithm.