Research On Adaptive-region Anomaly Detection Algorithm For Hyperspectral Remote Sensing Images

In recent decades,space optical remote sensing technology has developed rapidly.The optical remote sensing images develops from multispectral images to a new stage which is named as hyperspectral images.Hyperspectral images have the advantages of high spectral resolution and space-spectrum combination,which have been widely used in civil and military fields and have promoted rapid progress in geological prospecting,agricultural and forestry monitoring,military reconnaissance and other fields.Anomaly detection is one of the important research topics in hyperspectral image processing,which is characterized by the separation of outliers without prior information and can quickly locate the area of interest in the image.Lately,many scholars at home and abroad have made numerous achievements in the field of hyperspectral anomaly detection.However,due to the complicated real ground scenes,it is inapposite to simply construct statistical model for the background of hyperspectral images.And some texture information in hyperspectral images can cause variability in the spectra of pixels belonging to the same type of ground objects,which will make some existing anomaly detection algorithms insufficient.In this paper,starting from the inherent characteristics of hyperspectral images,we will carry out research on the anomaly detection methods of hyperspectral remote sensing images.The main research content and innovative results are described as follows:(1)At present,most hyperspectral anomaly detection algorithms need to describe the background of hyperspectral images.However,it is often impossible to accurately model the background since real ground scenes are usually complex,which leads to the unsatisfactory detection performance of anomaly detection algorithms.In addition,only considering the spectral difference between the abnormal pixels and the background pixels as well as ignoring the spatial information in the image usually cannot obtain the best anomaly detection effect.In response to these two issues,chapter 3 of this thesis proposed a novel spectral-based selective searching hyperspectral anomaly detection method(Triple-S).The graph-based image segmentation method is adapted to hyperspectral images,and a dynamic spectral similarity operator is proposed to realize the initial segmentation of hyperspectral images with several regions having adaptive shape and size.The regions after the initial segmentation are used as the abnormal detection processing units.The adjacent regions with similar spectral characteristics will be fused through the greedy algorithm and the calculation of the spectral similarity of adjacent regions.And the regions that cannot be merged with the adjacent regions are the abnormal targets to be detected.The accurate detection of abnormal targets in the hyperspectral image is realized.Experimental results show that the spectral-based selective searching anomaly detection method not only achieves superior anomaly detection performance,but also obtains an ideal separation degree between background and anomalies.The area between the lower right of the receiver characteristic operating curve and the coordinate axis(AUC)of the test images AVIRIS-I and URBAN are respectively 0.9997 and 0.9998,which quantitatively characterizes the superior anomaly detection performance of the Triple-S algorithm.(2)Aiming at the problem that the local background is contaminated by anomalies in the existing local hyperspectral anomaly detection algorithm based on the dualwindow structure,the fourth chapter of this thesis proposed a background screening process of dual-window anomaly detection method to improve local anomaly detection algorithms.The spectral-based selective searching algorithm was used to make a preliminary judgment on the background or anomaly of the pixels in the hyperspectral image.Then the abnormal pixels in the local background framed by the double window were filtered out,thereby improving the detection performance of the local anomaly detection algorithm based on the dual-window.Taking the local RX algorithm and the collaborative representation algorithm as an example,the application of background screening shows that it can effectively alleviate the situation that the local background in the dual window is contaminated by anomalies,and then improve the existing dual window anomaly detection method for large anomaly targets The anomaly detection ability of local hyperspectral anomaly detectors can be greatly improved,while the dependence of anomaly detection effect on the size of the dual window is significantly reduced.The AUC value of the anomaly detection for the test image AVIRIS-I under different double window sizes increased by an average of 24.46%,and the AUC value of the anomaly detection results of the test image Airport under different double window sizes increased by an average of 12.11%,which indicates the great reduction of the dependence of the anomaly detection effect on the size of the dual-window.The standard deviation of the AUC value of the test image AVIRIS-I under different dualwindow sizes is 0.15,and the improved standard deviation is 0.04.The standard deviation of the AUC value of the test image Airport under different dual window sizes is 0.082,and the improved standard deviation is 0.009.A decent anomaly detection effect can be achieved under a variety of double window sizes.(3)In view of the problem that the complex texture information in the hyperspectral image may increase the spectral variability of pixels belonging to the same type of feature,and the construction of the k-means-based background dictionary will cause the instability of the sparse representation background dictionary,this thesis proposes a hyperspectral anomaly detection algorithm based on two-scale relative total variation combined with low-rank sparse representation in the fifth chapter.This algorithm applies the relative total variation model to the field of hyperspectral anomaly detection for the first time,and uses the dual-scale relative total variation model to extract the structural information in the hyperspectral image.At the same time,the graph-based image segmentation method was used to replace the k-based image segmentation method.The segmentation method generates a series of regions with similar spectral characteristics and combines with structural information to build a more accurate and stable background dictionary.Finally,the alternate iterative direction method is used to solve the background matrix and background sparse representation coefficients to obtain the residual matrix.The judgement of whether the pixel is an abnormal pixel was based on the response of each pixel.Experimental verification is performed on four typical hyperspectral images.Compared with the average AUC value(50 times running)of the low-rank sparse representation anomaly detection algorithm using k-means to build a background dictionary,the AUC value obtained by the proposed algorithm is increased by 4.63%.The experimental results show that the hyperspectral anomaly detection algorithm of dual-scale relative total variation combined with low-rank sparse representation can not only effectively detect the anomalies in the image,but also have a good suppression effect on the background in the image.The problem of unstable detection performance caused by the construction of a background dictionary based on the k-means algorithm was also solved.