Research On Hyperspectral Remote Sensing Anomaly Detection Based On Background Refinement

Hyperspectral imagery is a new type of remote sensing image.It contains both the spatial information and spectral information of the objects.In general,its spectral resolution ranges from10 nm to 20 nm.Compared with multispectral imagery,hyperspectral imagery has more spectral bands.The hundreds or even thousands of spectral bands can provide much more subtle differences for different objects,which benefits the development of target recognition and object classification.Based on the availability of priori target information,the target detection can be roughly divided into two categories: supervised target detection and unsupervised target detection.We usually call the unsupervised target detection as anomaly detection.Anomaly detection analyzes the spectral difference between the pixel and its near neighborhood,and then determine directly that the pixel is an anomaly or not.In general,if there are few anomalies in background set,it would weaken the spectral difference between the target and the background pixels.The probability of detection for anomaly detection algorithms can be improved effectively by obtaining a purer background.This paper focuses on the background refinement method,which means that the potential anomalies are removed from the initial background.And we combine the background refinement method with the current anomaly detection algorithms to improve their probability of detection.The main research contents in this paper are as follows:First,the classical RXD(Reed-Xiaoli detector)has the problem that the estimation of background statistics will distort when the initial background contains some anomalies.It causes that the detection result of RXD performs with a high false alarm rate.To solve the problem,a new background refinement method based on local density is proposed.The local density of the pixel is considered as its anomaly degree label.Through the Otsu's method,we remove the potential anomalies in the initial background.Finally,we can get a purer background.Using the refined background to get more accurate background information,can effectively reduce the false alarm rate of the RX algorithm.Secondly,aforementioned background refinement method based on local density needs to select hypersphere radius through experiments to achieve the best detection performance.For the problem,we add a radius adaptive procedure to the background refinement.The improved background refinement method is then combined with RXD and collaborative-representation-based detector to detect the anomalies in hyperspectral image.The experimental results show that the improved method can effectively improve the probability of detection.Finally,combined density estimation and collaborative representation anomaly detector is proposed.Compared with the local density model,each background pixel has a contribution to the estimation value.A smaller pixel probability density represents that the pixel is more likely to be an anomaly.Firstly,the probability density of each pixel in the initial background is calculated.Then,the pixels with larger anomaly degree are removed by Otsu's method.Finally,the refined background is combined with CRD to detect the anomalies in hyperspectral imagery.