Research On Intelligent Detection Methods For Spectral Imagery

Intelligent detection of spectral images is an active research area in the field of spectral image processing and analysis.How to use spectral images to more effectively serve the major national needs of geographic environment monitoring,food safety,medicine and health,and precision agriculture has become a core problem in the field of intelligent detection of spectral images.Spectral anomaly detection and change detection are hot research topics of intelligence detection of spectral images.Therefore,this dissertation will focus on the anomaly detection and change detection of spectral images.Spectral anomaly detection and change detection are widely used in many fields,such as precision agriculture,food and drug testing in public safety,urban expansion monitoring,land cover change,resource management and damage assessment.Since spectral anomaly detection and change detection have important applied values,spectral anomaly and change detection are the important research directions in the field of intelligence detection of spectral images,and have achieved great progress.However,some problems still exist: 1)inaccurate complex background model;2)insufficient exploration of the intrinsic structure of spectral images;3)insufficient use of spatial information and spectral information.In view of the above problems,this dissertation studies intelligent detection of spectral images from four perspectives.The main research contents and contributions are summarized as follows:(1)Similarity constrained convex nonnegative matrix factorization for hyperspectral anomaly detection.The existence of mixed pixels cannot be ignored due to the low spatial resolution of hyperspectral data,which makes the difference between background and anomalies not significant enough.Therefore,from the perspective of endmembers,this dissertation builds a basis matrix by endmembers obtained by the similarity constrained convex nonnegative matrix factorization algorithm,which can improve the detection accuracy of samples hard to classify.(2)Manifold constrained auto-encoder network based hyperspectral anomaly detection.The latent representations learned from the auto-encoder or deep belief network are not always able to reflect the intrinsic structure of hyperspectral data because the locality property is not considered during the learning process.Manifold learning method is employed to constrain the learning of latent representations in this dissertation,which aims to make the latent representations with locality characteristic.Then,the global and local reconstruction errors are combined to improve the anomaly detection accuracy.The detection accuracy on the Kunshan dataset has increased by 4% compared to other algorithms.(3)Coarse-to-fine semi-supervised change detection for multispectral images.In general,the multispectral image is only composed of a few bands,therefore,spectral information of multispectral images is relatively inadequate and cannot represent objects very well.This dissertation first builds the multiscale feature that both considers the spectral and spatial information;then acquires more discriminant metric by strengthening the contribution of training samples easy to be classified and weakening the contribution of training samples hard to be classified to the trained model;finally,employs a coarse-to-fine strategy to detect testing samples from the viewpoint of distance metric and label information of neighborhoods in spatial space to improve the change detection accuracy.(4)Spectral-spatial joint learning for change detection in multispectral imagery.Spectral-spatial joint learning network proposed by this dissertation contains three parts: spectral-spatial joint representation,feature fusion,and discrimination learning.Spectral-spatial joint learning network explores the underlying information and highlevel semantic information by spectral-spatial joint representation and fusion.Moreover,A new loss function that considers both the losses of the spectral-spatial joint representation procedure and the discrimination learning procedure is proposed to make the learned representations more discriminative in order to improve the change detection accuracy.