Deep Clustering Methods For Hyperspectral Remote Sensing Imagery

Hyperspectral remote sensing,which collects continuous spectral bands of objects with nanometer resolution,is one of the most advanced Earth observation technologies.Hyperspectral imagery(HSI)often consists of rich spatial,radiation,and spectral information,allowing users to analyze objects of interest precisely.HSI has been widely applied in various fields including geological exploration,fine agriculture,and military.HSI classification is a fundamental task for the application and analysis of HSIs.Nowadays with the flourishing development of artificial intelligence technologies,particularly deep learning,as well as hyperspectral imaging technologies,HSI classification is faced with many challenging issues.On the one hand,advances in imaging sensors have greatly facilitated the collection of remote sensing data.The availability of the massive data enables us to design more complicated algorithms but also issues new challenges to the efficiency and effectiveness of learning methods.On the other hand,the training of deep learning models relies on high-quality labeled data.However,it is often expensive to manually label samples in practice,which has become one of the most important factors restricting the development of HSI classification.Based on the above analysis,HSI clustering is expected to be a promising solution to overcome the problems caused by the scarcity of label information.Traditional HSI clustering approaches usually belong to shallow methods,and thus would suffer from the following problems in handling complex HSI scenes.(1)Hyperspectral imaging mechanism often leads to high-dimensional,nonlinear,and noisy HSI data.Traditional clustering methods use either handcrafted features or shallow features.They often not only ignore the high-level and nonlinear semantics but also fail to consider local and global structure relationships,thus resulting in poor robustness and low clustering accuracy.(2)Many existing HSI clustering methods,such as subspace clustering,generally adopt pixel-level or off-line strategies,which require quadratic or higher computational and storage complexity,resulting in poor scalability in large-scale hyperspectral image scenes.To address the above-mentioned problems,this thesis proposes a series of effective theories and methods for HSI analysis based on deep clustering models.More specifically,the goals of this thesis are to improve the robustness and scalability of clustering algorithms.The main contributions of this thesis are summarized as follows:(1)The traditional subspace space clustering method only considers the linear subspace of HSI,making it fail to exploit high-level nonlinear subspace and structural information.To circumvent these drawbacks,this thesis introduces a graph regularized residual subspace clustering network.The proposed approach models the nonlinear selfexpressiveness based on deep embedding obtained by a residual convolutional autoencoder.The non-local structural information is incorporated into the subspace representation by leveraging a graph regularization technology,resulting in improved robustness of the resulting model.Extensive experimental results show that the proposed method can accurately learn the subspace structure and achieve state-of-the-art results.(2)To generalize existing subspace clustering approaches into the nonEuclidean space,this thesis proposes a graph convolution subspace clustering framework based on spectral graph convolution.The resulting model framework can efficiently aggregate and propagate structural information,thereby reducing the intraclass variance and increasing the discriminative ability.Furthermore,an efficient optimization strategy is introduced by calculating closed-form solutions,based on which a linear and a nonlinear version of the proposed method are implemented.Experiments on several HSI datasets demonstrate that the proposed methods are significantly better than previous works in terms of robustness and accuracy.(3)To scale up deep subspace clustering to large HSI datasets,this thesis presents a superpixel contracted neighborhood contrastive subspace clustering network.First,a novel superpixel pooling auto-encoder is devised to reduce the computational complexity of deep self-expressiveness by leveraging the homogeneousness of HSI.Second,inspired by contrastive learning,a novel neighborhood contrastive regularization is introduced based on the context relationship of HSI,which can enhance the subspace structure to be learned.The experiments indicate the proposed model not only has good scalability but also can achieve superior results.(4)A one-stage online clustering framework called spectral-spatial contrastive clustering network is introduced for large-scale HSI clustering problems.The network follows a self-supervised contrastive learning architecture,which consists of sematicpreserving augmentation strategies and intra-and inter-class contrastive losses.The resulting model can be efficiently trained with mini-batch in an end-to-end fashion,resulting in good scalability,flexibility,and generalization ability.The state-of-the-art experimental results of the proposed model signify that it holds immense potential for large-scale HSI processing.Through the above four aspects,a series of new approaches for hyperspectral image clustering are proposed,which effectively improves the robustness and extensibility of deep clustering approaches and provides effective ways for the processing and analysis of hyperspectral images.