Research On Probability And Generative Model Based Hyperspectral Anomaly Detection

The development of an integrated monitoring and forecasting system for Earth observation is crucial for current advancements in Earth science.One of the core scientific challenges in establishing such a system is detecting anomalies and their background quickly and ac-curately under Earth observation.Hyperspectral remote sensing images,which are highly sensitive to spectral information of objects on the ground,have facilitated the application of hyperspectral anomaly detection for obtaining information about anomalies and their environment.The anomalies here mainly refer to small artificial targets containing metal materials in Earth observation scenes,including aircraft and buildings.In practical appli-cations,hyperspectral remote sensing images often suffer from the problem of anomalies being”lost”in complex and diverse backgrounds due to factors such as spectral confusion and distortion,and unbalanced samples.Moreover,the degradation mechanisms associated with the measurement process and atmospheric effects inject striping noise into the images,making anomalies susceptible to noise contamination.The introduction of deep generative models can enhance the detector’s ability to understand and model anomalies but also pos-es technical challenges,such as limited samples and lack of prior knowledge.This paper presents a comprehensive analysis of the various categories of data in hyperspectral images,including anomalies,background,and noise,as well as their relative relationships.Rely-ing on the spatial probability theory of the number of anomalies being much smaller than the number of background samples,the physical properties theory of spectra of anomalies being distinct from that of background samples,and the mathematical statistics assumption that the background obeys a multivariate Gaussian distribution,generative network models are systematically constructed to create anomaly-background reconstruction representation models based on unsupervised learning and weakly supervised learning.Firstly,a background discriminated reconstruction model under unsupervised learning is proposed,which aims to indirectly reconstructing the background to detect anomalies.The model utilizes latent adversative learning,consistency constraint,and shrink constraint to ensure that the network focuses on reconstructing background samples.The spectral dis-criminator prevents the generation network from generating variants outside the original image category.They collectively guarantee that the model can suppress anomalies while effectively reconstructing the background spectral vector,so that the residual image between the original image containing both background and anomalies and the reconstructed back-ground image can effectively highlight the anomaly target information.The AUC score of((P_d,P_f))of this model is 0.9849,which is 0.99%higher than the second best algorithm,especially 1.18%higher on the Los Angeles dataset.However,it still faces limitations due to the lack of prior information resulting in limited detection performance.The model has explored the potential of using generative adversarial networks for hyperspectral anomaly detection,and addressed the issue of anomalies getting”lost”in the background by leverag-ing background reconstruction.The weakly supervised anomaly-background spectral constraint separation model introduces the concept of weak supervision,relying on the fact that the initially given labels are not always true.The model developed two novel coarse label generation models,namely a category probability-based coarse label generation model and a spectral distance metric-based coarse label generation model.The subsequent network learning under weak super-vision incorporates a spectral distance constraint in the generative network to construct an anomaly-background spectral constraint separation model with background homogeniza-tion and anomaly salience.Compared to other comparison methods,this approach not only achieves the optimal detection performance(with the average((P_d,P_f))AUC score of 0.9860)but also demonstrates superior performance on different datasets tested using the same trained network.The key focus of this work is to utilize spectral constraints to enhance the differentiation between anomalies and background,addressing the issue of anomalies being easily”lost”in the background.While the previous two models have fully considered the relationship between anomalies and backgrounds,they have not addressed the interference of striping noise on these factors.Therefore,this paper proposes a background reconstruction model inspired by sparse coding under weak supervision.This method is committed to restoring the quality of hyperspectral anomaly detection while minimizing the influence of striping noise in the spectral bands.It introduces weakly supervised learning to enhance detection performance and employs s-parse coding in representation theory to increase the robustness of the network to noise and prevent overfitting.Additionally,two different reconstruction error calculation strategies are explored:reconstruction in the original space and reconstruction in the latent space.The ex-periments show that the latent space enables the network to focus more on the differences in intrinsic spectral properties,resulting in reconstruction errors with higher noise robustness.The average AUC score of((P_d,P_f))by the proposed method is up to 0.9864,which exceeds the maximum value of 0.9765 in the comparison algorithm.This method mitigates the issue of anomalous susceptibility to noise pollution by reconstructing sparse codes and computing reconstruction errors in the latent space.After constructing the above network models,a joint spatial-spectral anomaly detector with a structure tensor and guided filtering has been proposed.This detector fully incorporates spatial information,taking into the context of each pixel and its surrounding pixels.This approach avoids spectral misguidance caused by single spectral dimension information.As a result,anomalies can be more accurately identified within complex backgrounds,thus improving the detection performance.The detector achieves both fast and accurate detec-tion and location of anomalies.The image decomposition and background removal steps based on attribute filtering can effectively separate the bands with anomaly prominence and background suppression;the adaptive weighted fusion step based on the structure tensor adaptively fuses the bands by measuring the contribution of different bands;the final filter is guided to the final detection result map by examining the correlation between adjacent pixels.The detector achieves AUC scores of((P_d,P_f))of 1.19%and 0.69%higher than the second-best method on the Los Angeles and San Diego datasets,respectively.This work addresses the issue of anomalies easily getting”lost”in the background by incorporating spatial information.Finally,the proposed joint spatial-spectral anomaly detector is applied to the reconstructed or residual images obtained from all the network models,resulting in seven detection results.It is found that 1)the detection performance of the three network models with the spatial-spectral detector is higher than that of the original spectral RX detector,indicating that the proposed spatial-spectral anomaly detector performs better on network output images than a single-dimensional spectral detector.2)The ability of the weakly supervised anomaly-background spectral constraint separation model with the cascaded spatial-spectral detector to detect abnormal targets is the strongest,with((P_d,P_f))AUC scores of 0.9966 and 0.9937on the Gulfport and San Diego datasets,respectively,approaching the optimal value of 1.3)The abnormal-background spectral constraint separation method could most effectively suppress the background information of the Gulfport dataset,achieving an((P_d,τ))AUC score of 0.0154,while the background suppression ability of the background reconstruction method inspired by sparse coding was the strongest for the San Diego dataset,with an((P_d,τ))AUC score of 0.0052.