Research On Key Problems Of Anomaly Detection In Hyperspectral Remote Sensing Images

The space-air-ground remote sensing platforms will produce a large amount of remote sensing data.Remote sensing technology has become an important means of monitoring geographical conditions,which can provide targeted data support and scientific basis for future informational society.Anomaly detection for hyperspectral remote sensing image is an important research content in the field of intelligent processing and analysis of remote sensing data.It does not need any prior information about the object and background in the observation scene,which can directly locate the potentially suspicious object in the image scene by analyzing the difference of spectral characteristics.It can provide the region of interest for the follow-up accurate target recognition and analysis.Therefore,it has important application value in precision agriculture,national defense security,geological survey and so on.However,with the rapid development of hyperspectral remote sensing technology,the high-dimensional characteristic of hyperspectral images is more significant,and the correlation between spectral bands is more complex.At the same time,the geographical space covered by the images is vaster,the material contained in the scene is more diverse,and the interference characteristic between different substances is more obvious.These characteristics bring new challenges to the accurate detection of hyperspectral abnormal targets.The key problems include: 1)effective feature selection;2)interference of abnormal targets;3)accurate modeling of complex scenes.Therefore,a series of researches on hyperspectral anomaly detection are carried out focusing on the three key issues mentioned above.The main research contents and contributions are summarized as follows:(1)To address the problem of effective feature selection,a hyperspectral anomaly detection method via discriminative feature learning with multiple-dictionary sparse representation is proposed.In this dissertation,a feature selection framework driven by anomaly detection task is constructed,and a reciprocal relationship between anomaly detection and feature selection is established to ensure that the selected representative features can satisfy the requirements of anomaly detection.In order to select the representative spectra that can enlarge the difference between background and target,this dissertation uses background information and abnormal information to jointly restrict the feature selection process.Considering the unique spectral attributes of different material types,this dissertation further constructs a multiple-dictionary feature learning model to fully discover the differences of material attributes and improve the discrimination of representative features.(2)To address the problem of abnormal target interference,a hyperspectral anomaly detection method via sparse dictionary learning with capped norm is proposed.In the traditional sparse representation based anomaly detection model,the learning of background dictionary is affected by the abnormal target.In this dissertation,the capped norm is used to restrict the learning process of background dictionary,which can cap potential anomalies with large reconstruction error.Only specific background pixels are adaptively selected for effective coding,and abnormal targets are avoided participating in the dictionary coding process.Therefore,the dissertation can effectively suppress the interference characteristics of abnormal targets,and learn a more accurate background dictionary to improve the detection performance.(3)To address the problem of accurate modeling of complex scenes,a hyperspectral anomaly detection method based on separability-aware sample cascade is proposed.According to the separability of different material samples in complex background data,this dissertation constructs a sample cascade selection framework.By perceiving the separable degree of different samples,the various materials in complex scenes are gradually sifted out layer by layer,which are divided into sample sets with different background attributes according to the difficulty of detection.This dissertation can effectively balance the traditional model's preference for easy-to-divide samples and improve the representation ability of hard-to-divide samples.It can effectively perceive the difference between complex background and abnormal target,and realize the accurate modeling of complex background data.Compared with the classical method,the average detection accuracy on the public available data sets is improved by 11.11%.