Research On Anomaly Detection In Hyperspectral Remote Sensing Images By Isolation Forest

With the development of remote sensing technology,hyperspectral remote sensing images(HSIs)have been widely used in civil and military fields because of its advantages such as high spectral resolution and spectrum-space combination.The detection of the target of interest is one of the important research topics in the field of HSIs processing,and anomaly detection is the main way to achieve target detection.Most of the existing algorithms involve higher-order operations such as matrix inversion,calculating covariance and solving higher-order derivatives in the solution process,which makes algorithms unable to achieve fast anomaly detection when processing large fields-of-view HSIs.There are also some algorithms that statistically model the HSIs,and then detect the pixels with large differences from the background as anomalies.However,real-world HSIs are usually complex and it is difficult to model accurately.In summary,most methods are more or less deficient in terms of the separation of anomaly pixels from background,the robustness and the computational efficiency of the algorithm.Therefore,the existing anomaly detection techniques still face challenges in the following two aspects: on the one hand,how to analyse and obtain initial anomaly detection results for large fields-of-view HSIs rapidly;on the other hand,how to make full use of spatial and spectral information for fine anomaly detection in certain areas.Based on the above two challenges,this dissertation,taking into account both spatial relations and spectral information,carries out the research on the anomaly detection in HSIs based on the Isolation Forest(i Forest).The main innovative work are introduced as follows:(1)A fast anomaly detection algorithm for large field-of-view HSIs is proposed:Considering the problem of false detection and false alarm of local anomaly pixels,the RMi Forest algorithm is proposed.Compared with existing algorithms,RMi Forest improves the limitations of high complexity and insensitivity to local anomaly pixels.Compared with the existing algorithms,the i Forest algorithm does not involve higherorder operations such as solving higher-order derivatives and calculating covariance matrices,and has low complexity.Also,the i Forest has linear complexity.Second,compared with the i Forest algorithm,the RMi Forest algorithm no longer defines the anomaly measure based on the path length from the leaf node to the heel node where the pixels is located.It is improved to a local anomaly measure based on the Relative Mass.RMi Forest partly solves the limitations of missed detection and false alarm for certain anomaly pixels that are normal points in the global image but outliers in the local area.In addition,the training and generation of each Isolation Tree(i Tree)in this algorithm are independent of each other.Hence,the RMi Forest algorithm has the application prospect of being deployed in large-scale distributed systems to accelerate the operation,which is convenient for the engineering implementation with high hardware portability and real-time requirements.(2)An anomaly detection algorithm in HSIs based on High Dimension Isolation Forest(HDi Forest)is proposed: Considering the higher accuracy requirements of anomaly detection in specified areas of HSIs,the improved strategy for anomaly target detection of high-dimensional data is given;the concepts of Spectral Useful Information Rate(SUIR)and Target-Background Separation(TBS)are proposed,and the band selection based on SUIR and TBS is operated.Compared with the existing i Forest-based anomaly detector,HDi Forest can select the spectral bands of HSIs for analysis optionally.Hence,HDi Forest detects the anomaly pixels more accurately while achieving effective suppression of the background,and improves the differentiation between the anomaly pixels and the background quantitatively and qualitatively.Also,HDi Forest has excellent robustness.Moreover,HDi Forest achieves good anomaly detection performance in both full-pixel anomaly detection and subpixel/small target anomaly detection.(3)A novel spectral-spatial anomaly detection framework based on multi-scale spatial constraint and i Forest(SSIIFD)is proposed.The multi-scale spatial constraint method is proposed based on Gabor filter and ERS algorithm.Specifically,Gabor features and segmented HSIs by ERS algorithm are employed to construct RMi Forest and HDi Forest,respectively.The advantages of the proposed SSIIFD method are threefold: first,the method fully utilizes spectral and spatial information in HSIs;second,this method fully employs global and local information in HSIs;third,this method detects anomaly pixels more clearly and accurately at a lower false alarm rate(FAR).The experiments on four real hyperspectral data sets reveal that SSIIFD is stable and superior to other state-of-the-art methods in terms of both objective and subjective evaluations.