Research On Anomaly Detection Algorithm For Hyperspectral Image Based On Superpixel Segmentation

Hyperspectral anomaly detection is one of the research hotspots in the hyperspectral image processing field and has practical application value in military reconnaissance,pollution monitoring,mineral exploration,and many others.Although many achievements have been obtained in hyperspectral anomaly detection field,there is still room for improvement in its detection performance.RX is the most classic hyperspectral anomaly detection technology.This paper mainly studies the superpixel segmentation methods applicable to hyperspectral images and the superpixel segmentation technologies to improve the RX detection performance.The major works of the thesis are as follows:(1)The application of the entropy rate superpixel(ERS)segmentation method in hyperspectral images is studied.On the one hand,using euclidean distance(ED),spectral angular distance(SAD),and spectral correlation coefficient(COR)as similarity indicators,the edge weights in ERS are calculated by full-band data.Three extended ERS methods are constructed: ERS-ED,ERS-SAD and ERS-COR.Experiments and analysis are conducted to compare them with the traditional method PCA-ERS,which adopts principal component analysis(PCA)as the pre-processing step of hyperspectral images and ERS as the superpixel segmentation method.On the other hand,in order to make the superpixels relatively regular and compact,a new method called PCA-ERS-FA is proposed to finely adjust the boundary pixels of each superpixel obtained by PCA-ERS.(2)Aiming at effectively determining the local background area in the RX algorithm,a superpixel-based double-window(SPDWRX)algorithm is proposed.A double-window construction method based on superpixels is designed to enable the background area not to contain the target pixel and be significantly different from the target pixel.The corresponding calculation method of the double-window RX detection value is given.Finally,two comparative experiments are designed: the experiment using different superpixel segmentation methods in SPDWRX and the experiment comparing SPDWRX with other local RX anomaly detection algorithms.The experimental results show that the proposed anomaly detection algorithm based on superpixel segmentation can improve the detection efficiency of the algorithm and maintain the accuracy of the algorithm.(3)The method of band selection and its effect on the performance of SPDWRX were further studied.A band selection algorithm based on fisher description analysis is proposed to use the RX anomaly detection algorithm to select background pixels,and then use the Gaussian differential function to calculate the differential background subset,the background scattering matrix and the total scattering matrix respectively.The scattering matrix ratio of the background scattering matrix to the total scattering matrix are consequently obtained.Then the target bands are obtained by forward selection and filtered by particle swarm optimization algorithm.Experiments results verified that the proposed band selection method can greatly improve the detection performance of RX based on dual windows and SPDWRX.