Research On Hyperspectral Anomaly Detection Methods Based On Intrinsic Component Representation

Hyperspectral image anomaly detection technology can complete the positioning of potentially interesting targets in an unsupervised manner without any prior knowledge.It has been a research hotspot in the field of hyperspectral image processing,which has extensive applications in military reconnaissance,precision agriculture,geological exploration,and other fields.Due to the susceptibility of anomaly targets to background and noise,and the difficulty in predicting and modeling anomalies,accurate characterization of the background becomes a key technology for hyperspectral anomaly detection.Although existing hyperspectral anomaly detection methods mostly focus on background estimation,there are still phenomena such as inaccurate background representation,difficulty in separating noise and anomaly components,and insignificant background suppression and target enhancement capabilities,which affect the application significance of hyperspectral anomaly detection.To address these issues,this article conducts research on hyperspectral anomaly detection methods from the perspective of intrinsic component representation of data.The main research work is summarized as follows:(1)Aiming at the problem of anomaly signal leakage and poor robustness when constructing the dictionary of low-rank and sparse model,a hyperspectral anomaly detection method based on sparse component representation is proposed.By calculating the local outlier factor,the anomaly prior probability of the sample is excavated.Incorporating with the model,the anomaly components of the data are obtained after an iterative update process under the condition of multi-knowledge constraints,and an adaptive matched filter is designed to achieve the effects of background suppression and target enhancement.The relevant experimental results demonstrate that the algorithm owns powerful detection capability and good robustness.(2)Aiming at the problem of difficulty in modeling both the background and anomaly targets uniformly,a hyperspectral anomaly detection method based on the probability distribution representation with the framework of variational autoencoder is proposed.Due to the characteristics of network,the samples are characterized as multivariate Gaussian distributions in the joint space of background and anomalies.To obtain accurate representations of the background and anomalies,the knowledge constraint of subspace orthogonality is introduced into the loss function to attempt to recover relatively orthogonal background and anomaly subspaces from the original sample space.The statistical characteristics of the local background are estimated by designing Chebyshev neighborhood and a anomaly detector based on the Wasserstein distance metric is constructed to compute the discrepancy between the evaluated samples and their expected distribution.The relevant experimental results demonstrate that the proposed algorithm possesses strong detection capabilities and good computational efficiency.(3)Aiming at the problem of mutual influence between different signals caused by the fact that background and anomaly subspaces are not strictly orthogonal,a hyperspectral anomaly detection method based on the probability modeling of anomaly components is proposed,which combines the low-rank sparse representation model with probability distribution representation.The low-rank sparse representation model is exploited to acquire the intrinsic anomaly components of the data,which are the training data for the network and reflect the anomalous probability of the samples to some extent.After optimizing the weights of the variational autoencoder network,the intrinsic probability representation of the samples in the anomaly subspace is constructed.The final detection map is obtained through decision-level fusion,incorporating prior knowledge contained in the anomaly components.The relevant experimental results demonstrate that the proposed algorithm owns powerful detection capability and good background suppression ability.(4)Aiming at the problem of limited representation ability for background and anomaly in joint space,a hyperspectral anomaly detection method based on normalizing flow and Gaussian mixture estimation is proposed.Considering the diversity of sample space in practical scenarios,we introduce normalizing flow and transform the Gaussian distribution of the latent variables into distributions with stronger fitting ability through probability distribution transformation,significantly improving the generative and expressive power of the model.From the statistical and local spatial characteristics of the data,we design a anomaly detector based on Gaussian mixture estimation theory.The relevant experimental results demonstrate that the proposed algorithm possesses strong detection capability and good background suppression ability.