Research On Hyperspectral Anomaly Detection Via Robust Background Modeling

Compared with the traditional remote sensing methods like infrared and multispectral images,hyperspectral imagery（HSI）can provide more discriminative spectral information.This property encourages HSI to be applied in many fileds like target detection,mineral exploration and fine agriculture.As an essential part in HSI processing,anomaly detection can automatically detect potential targets with no a prior spectral information,which is advantageous in remote sensing applications.One of the most important steps in anomaly detection is background modeling,which has a great impact on the detection result.In real scenes,however,HSI data often suffers from some problems like spectral variation,indiscriminative features and insufficiency in exploitation of relationship among samples.These bring about severe challenges to the robustness of the model construction and thus degrade the detection performance.Aiming at the robustness of background modeling,this thesis focus on the problems encountered in hyperspectral anomaly detection,and analyzes the characteristics and drawbacks of state-of-the-art methods.Then,several hyperspectral anomaly detector based on robust background modeling are proposed in this thesis.Specifically,the main contributions of this paper are listed as follows:（1）Starting from the hyperspectral anomaly detection and the basic idea behind it,this thesis analyzed the major problems commonly existed in the procedure of hyperspectral imaging and the influences on background modeling,before summarizing the characteristics of anomalous targets in hyperspectral remote sensing data;Several typical anomaly detection methods based on probability theory,optimization theory and deep learning theory are deduced in detail,and the characteristics and drawbacks of each algorithm are introduced;Finally,the hyperspectral data sets and detection evaluation metrics utilized in this thesis are briefly introduced.（2）Addressing the failure of robust modeling by spectral variation and pullution of anoamlies,a hyperspectral anomaly detection method based on feature extraction and background purification（FEBPAD）is proposed.To extract more distinctive features,FEBPAD introduces fractional Fourier transform（Fr FT）to fuse the information from both the original space and the corresponding Fourier domain.Moreover,the low rank and sparse matrix decomposition algorithm is modified to preserve the physical meaning of sparse matrix in HSI applications,taking advantage of the row-sparsity constraint of targets in HSI.Last but not least,a background purification strategy is proposed to obtain background data with high confidence in a coarse-grained level,thus avoiding the pollution of abnormal samples to the statistical information of background and ensure the robustness of model.The experimental results validate the fact that the method performces better than original methods,and the computational cost is acceptable in real scenes.（3）Addressing the failure of robust modeling by the weakness of representation ability and the insufficiency in exploitation of relationship among samples,a hyperspectral anomaly detection method based on spatial-spectral collaborative autoencoder（SSc AE）is proposed.Inspired by the infrared patch image（IPI）model,SSc AE obtains anomaly responses via hyperspectral patch image（HPI）model,a modified version of IPI,in spatial domain.The spatial response is also used for guidance to select background pixels with high confidence,taking advantage of the excellent ability of HPI in background suppression.This strategy can help autoencoder to be robust to anomalous samples during training.Further,a weighting operator based on spatial response is designed to fuse the spatial and spectral information on decision level.Experiments on five data sets demonstrate the effectiveness and superiority of the proposed method,and the result can effectively improve the contrast between the detected anomalous target and the background.（4）Addressing the failure of robust modeling by sub-optimal solution and the weakness of representation ability,a hyperspectral anomaly detection method based on robust graph autoencoder（RGAE）is proposed.RGAE proposes an end-to-end robust autoencoder anomaly detection framework,which automatically pays more attention to the background samples with high confidence during training,thus alleviates the influence of anomalies and noise on the training process,and eventually ensures the stability of the whole training process.At the same time,aiming at the problem that autoencoder often ignores the relationship between samples,RGAE tend to utilize the graph regularization constraint to preserve the geometric structure information among pixels,anda fast graph construction strategy,Super Graph,is designed.Compared with the traditional graph construction method,it not only greatly reduces the amount of calculation,but also introduces space information,which further improves the representation ability of the network.Experiments show that this method is more robust to anomalies and noise than the traditional autoencoder framework,and can extract more discriminative features,so as to better detect potential abnormal targets in the background.