Research On Hyperspectral Anomaly Detection Based On Ensemble Learning And Fuzzy Clustering

Hyperspectral image contains rich spectral information and spatial information,which brings good opportunities for the fine identification of ground objects.It has high application value and broad development prospect in civil and military fields such as disaster prevention,mineral exploration,national defense and security.However,the characteristics of hyperspectral image,such as large amount of data,strong background uncertainty,high characteristic dimension and strong spectral correlation,bring a series of challenges to its processing and analysis.Anomaly detection is an active research topic in the field of hyperspectral image processing.Compared with the target detection task,anomaly detection does not need any prior information and has higher practical application value.Therefore,from the perspective of hyperspectral anomaly detection,this thesis studies its development status and related theoretical basis,and proposes two new hyperspectral anomaly detection algorithms.The main research contents of this thesis are as follows:（1）In order to make full use of spatial and spectral information of hyperspectral images and reduce the pollution of anomaly pixels on background modeling,an ensemble and random RX with multiple features（ERRX MF）anomaly detector is proposed in this thesis.Firstly,the spectral feature,Gabor feature,extended morphological profile（EMP）feature,extended multiattribute profile（EMP）feature are extracted and apply to hyperspectral anomaly detection.Multi-feature fusion can make full use of the complementarity of features and provide more complete information for anomaly detection.Then,considering the pollution of anomaly pixels to the background modeling,the algorithm uses random sub-sampling method to model the background for each feature multiple times.Finally,the final anomaly score of each pixel in hyperspectral image is obtained by ensemble learning.Experiments on four real hyperspectral data sets show that the proposed ERRX＿MF can get better anomaly detection performance and more suitable background suppression ability.In addition,the running time is quite fast,which is very competitive compared with other algorithms and suitable for practical scenarios.（2）Due to the large imaging range and complex background of hyperspectral images,we hope to divide the background into multiple clusters,and then detect anomalies in the cluster with relatively single background.However,due to the high dimension and large amount of redundant information,the traditional clustering method can not deal with the hyperspectral data well.Therefore,this thesis proposes a hyperspectral anomaly detection algorithm based on projection fuzzy clustering（PFCAD）.Firstly,a new projection fuzzy clustering method is used to analyze hyperspectral images.The projection fuzzy clustering algorithm can learn projection matrix and cluster indicator matrix at the same time to find the optimal clustering subspace adaptively.Then,RX anomaly detection is performed on each cluster.Finally,the anomaly detection results of all clusters are integrated to obtain the anomaly detection results of the whole image.The anomaly detection experiments on four real hyperspectral data sets show that the proposed PFCAD has better anomaly detection performance than other comparison algorithms.In addition,we also prove the clustering performance and convergence of the projection fuzzy clustering algorithm through experiments.