Research On Anomaly Detection Methods For Hyperspectral Remote Sensing Imagery

Hyperspectral images(HSIs)contain rich spatial-spectral information.Compared with other remote sensing techniques,HSIs have widely applied in precision agriculture,military reconnaissance,marine monitoring and so on.However,HSIs also produce some challenges in image acquisition and analysis.With the improvement of the spectral resolution,the data dimension has largely increased.A large amount of redundant information limits the interpretation performance of traditional methods.Different objects have the same spectrum and the same objects have the different spectrum.With the existence of this phenomenon,it is difficult to use only the spectral difference to distinguish different objects.With the existence of the mixed pixel,the pixel-level feature fails to utilize spectral-spatial information effectively.Due to the existence of flares and noises,traditional anomaly detectors fails to identify the oil spill area.To solve these problems,this paper summarizes and analyzes the related research works.Then,we introduce image denoising and anomaly detection methods.Image denoising is a technique,which can reduce noises in images.Anomaly detection is an unsupervised technique for automatic target detection,which aims to discriminate anomaly targets from their surrounding background.The research contents are as follows:(1)Aiming at solving the mixed noise contamination problem in degraded HSIs,this paper proposes a group low rank tensor recovery method.This method first searches the non-local similarity regions in HSIs,and combines these regions to generate group tensors.Then,lowrank tensor recovery is employed to remove the mixed noise in group tensors.The advantages of the method are as follows.First,the spectral-spatial information in HSIs can be exploited effectively by adopting the nonlocal similarity strategy.Then,the low rank tensor recovery technique can effectively utilize the 3D spectral-spatial property in HSIs.As a result,the mixed noise will be effectively removed from observed HSI while the important spectral-spatial information will be preserved well.Compared with both the low-rank matrix recovery and other classical spectral-spatial restoration methods,the proposed method can always achieve an improvement in terms of both visual impression and quantitative measurements.(2)Due to the redundant information in HSIs,detecting anomalies effectively is a challenging task.This paper introduces a game theory-based anomaly detection method.First,by capturing the spectral-spatial property in anomalies,we define the anomaly detection problem as an anomaly game.When the anomaly game reaches the Nash equilibrium state,based on each player strategy,an optimal anomaly detection results can be generated.Then,based on the saliency in spectral-spatial features,the decision fusion technology is employed to combine three features.With the complementary information of spectral-spatial features,the fused detection image can be obtained.Experimental results validate that our approach can outperform some state-of-the-art anomaly detection methods.(3)Different objects have the same spectrum and the same objects have the different spectrum.To solve this problem,we introduce an anomaly detection technique based on the manifold ranking.This method not only uses the difference between anomaly pixels and background pixels,but also uses the similarity between anomaly pixels to detect.Firstly,a set of query points are generated automatically by the difference estimation and thresholding.Then,we design a closed-loop graph to describe the similarity between adjoining nodes,in which each node is a superpixel region.Finally,according to the similarity between nodes and exception queries,the ranking value of each node is calculated by the manifold ranking algorithm to estimate the final detection result.The experimental results demonstrate the proposed method can achieve good detection performance in the complex environment.(4)The single feature fails to exploit the spectral-spatial information in HSIs.To solve this problem,this paper introduces a subpixel-pixel-superpixel guided fusion method for hyperspectral anomaly detection.First,subpixel-,pixel-,and superpixel-level features are extracted from an HSI by employing the spectral unmixing,morphological operation,and superpixel segmentation techniques,respectively.Then,based on the spatial consistency of three features,a guided filtering based weight optimization technique is developed to construct weight maps for fusion.Last,an effective decision fusion method is adopted to utilize the complemental information of three features,and then produces a fused detection result.The proposed technique is tested on real scene HSIs and one synthetic HSI.Experimental results validate the advantages of the SPSGF method.(5)The oil spill and sea are strong absorbers of the natural light.Their spectrums are difficult to distinguish.Besides,due to the existence of the sunglint and noise,traditional anomaly detectors fail to detect oil spill effectively.Therefore,this paper proposes a novel spectralspatial hyperspectral oil spill detection method.First,the subpixel spectral feature in HSIs is extracted with spectral unmixing.Then,the low-rank matrix decomposition and domain transform recursive filtering technologies are employed to remove the environmental interferences,such as sunglint and noise.Finally,according to the anomalous characteristics of the oil spill area,the Mahalanobis distance between each pixel and the background is used to estimate the oil spill area.Experimental results show the advantages of the proposed technique even in complex marine scenes with the interference of the sunglint and noise.