Research On Methods For Anomaly Detection In Hyperspectral Remote Sensing Imagery

Over the past three decades,with the rapid development of aerospace and imaging spectrometer technology,the method of Earth observation based on hyperspectral remote sensing has attracted more and more attention.Compared with conventional remote sensing images,the most significant feature of hyperspectral imagery(HSI)the ability to acquire the spatial and spectral information of objects simultaneously,which provides support for analyzing the characteristics of objects accurately and quantitatively.Anomaly detection is one of the most important research topics in the hyperspectral remote sensing field.It can find regions of interest with spectral differences from their surrounding background without any prior knowledge.However,the complex background components,the existence of subpixel anomalies and the background regions with small areas may have negative influences on the detection performances of many anomaly detection methods.In order to overcome or alleviate these issues aforementioned,we consider the spectral difference between anomalies and the background as well as the spatial distribution characteristics of anomalies simultaneously,and propose two new anomaly detection algorithms and a new performance boost strategy in the thesis,which promotes the development of the hyperspectral anomaly detection filed to some extent.The main contributions of the thesis include:1)A hyperspectral anomaly detection method based on background estimation and adaptive weighted sparse representation is proposed.The basic motivation of the method is that the spectra of background pixels can be well reconstructed based on a background over-complete dictionary,while the spectra of anomalous pixels cannot.Then the background and the anomalies can be distinguished based on the reconstruction errors.In order to construct the background dictionary accurately,a new strategy based on endmember extraction and structure element sliding process is utilized in the method.And for further increasing the discrimination between the background and anomalies in the detection result,it is weighted adaptively from the global and local domain based on the anomaly endmembers' abundance images set.Finally,a more accurate and reasonable detection result can be obtained.Experiments on three HSI datasets validate that the proposed method achieves an outstanding detection performance compared with the other anomaly detection methods,especially in the suppression of the global background response energy.Moreover,the method is also quite robust to the transformations of the parameters.2)A hyperspectral anomaly detection method based on low-rank and sparse matrix decomposition(LRaSMD)and cluster weighting is proposed.Considering the low rank characteristic of background components and the sparse characteristic of anomalous targets,the method uses the LRaSMD technique to decompose the HSI data set.Since the obtained sparse part contains the major anomaly information,the method generates an initial detection result bases on this part.In addition,since some background regions with small areas may also have large response energies on the sparse part,which can affect the detection performance of the method to some extent,thus a cluster weighting process is added to locate the background areas,and then pixels in these areas are assigned corresponding weights.Finally the weighted detection result is obtained.Experiments on three hyperspectral datasets demonstrate the proposal can alleviate the influence of small area background regions on the detection results.Compared with a series of state-of-the-art detectors,the proposed method possesses better detection performance.In addition,it also performs well on the computational efficiency as well as the robustness to the transformations of the parameters.3)A simple weighting strategy for improving hyperspectral anomaly detection is proposed.It is a modular approach that can be equipped on the results of generic anomaly detection methods as a postprocess,which is used to improve the detection performances.The tensor decomposition is introduced in the method,which can remove the main background information while preserving the intrinsic spatial structure information in the data set.Therefore,the anomalies and the background have more distinct spectral differences on the remaining part after the decomposition.Then the strategy applies a parameter adaptive k-means clustering method on the remaining part,and segment the cluster image.Afterwards,each pixel of the segmented image is assigned a corresponding weight according to an improved Gaussian weight function.Finally a weight matrix is obtained.The matrix can be combined with the results of the conventional anomaly detectors to improve their detection performances.Experiments on two hyperspectral datasets validate that the proposed strategy can locate background positions accurately and then suppress them for improving the performances of detectors.Moreover,the strategy also shows the strong robustness and adaptability in terms of the transformations of the parameter and different noisy conditions,respectively.