Hyperspectral image (HSI) subspace clustering re- mains a challenging task due to the poor spatial and rich spectral resolutions of HSIs. Most of the existing HSI subspace clustering approaches just extract the spatial and spectral features, ignoring the intrinsic distribution information of data and leading to low accuracy of clustering generally. To solve this problem, this paper presents a diffusion subspace clustering model (DiffSC) which learns distribution information of HSI data simultane- ously through a diffusion module (DM). Specifically, due to the diffusion probabilistic model (DPM) learning raw object data distribution to generate data of the same distribution, which has received wide attention in generation tasks and outperforms other generative models significantly, we attempt to apply the DPM in the field of feature extraction. DiffSC performs distribution information extraction of hyperspectral images by DM and fuses them with spatial-spectral features extracted by deep subspace clustering for training jointly. Experiment outcomes demonstrate that intermediate activation of specific timestep in the inverse diffusion process captures latent distribution information of images effectively and improves the HSI clustering accuracy significantly. Since the DPM is simplified, it can be easily trained from scratch. We evaluate the presented DiffSC on five real HSI datasets, and the experiments indicate that DiffSC can obtain the most advanced clustering outcomes that notably outperform most existing HSI subspace clustering approaches.

Suffering from rich spectral and spatial information, the hyperspectral images (HSIs) that embed low-dimensional nonlinear manifolds lead to a challenging clustering task. In this article, we propose a promising clustering method for HSIs, termed as graph convolutional optimal transport (GCOT). For capturing the intrinsic geometric structure of data, we develop optimal transport (OT) on their graph embeddings. The OT problem learns discrete transport coupling to form a natural affinity matrix used for spectral clustering. Different from most other methods of building a graph with the Euclidean distance among neighborhoods, we leverage the OT probability to seek the edges of the graph that characterizes the local nonlinear structure of the original feature. For profound understanding of the proposed model, we prove the coupling holds sparsity and low-rank property that are beneficial to spectral clustering and provide its relationship with subspace clustering. To the best of our knowledge, this work could be one of the first spectral clustering methods using the OT theory. Several experiments on three widely used datasets with subjective and objective evaluation show that the proposed method dramatically outperforms the common existing methods and yields a comparable result with the state-of-the-art methods. Especially, it is significantly superior to the latter ones on computational efficiency.

Acknowledging different wavelengths by imaging mechanisms, optical images usually embed higher low-dimensional manifolds into ambient spaces than SAR images do. How to utilize their complementarity remains challenging for multimodal clustering. In this study, we devise a conditional dual diffusion (CDD) model for multimodal clustering of optical and SAR images, and theoretically prove that it is equivalent to a probability flow ordinary differential equation (ODE) having a unique solution. Different from vanilla diffusion models, the CDD model is equipped with a decoupling autoencoder to predict noises and clear images simultaneously, preserving data manifolds embedded in latent space. To the fuse manifolds of optical and SAR images, we train the model to generate optical images conditioned by SAR images, mapping them into a unified latent space. The learned features extracted from the model are fed to K-means algorithm to produce resulting clusters. To the best of our knowledge, this study could be one of the first diffusion models for multimodal clustering. Extensive comparison experiments on three large-scale optical-SAR pair datasets show the superiority of our method over state-of-the-art (SOTA) methods overall in terms of clustering performance and time consumption. The source code is available at https://github.com/suldier/CDD.

Graph convolution subspace clustering has been widely used in the field of hyperspectral image (HSI) unsupervised classification due to its ability to aggregate neighborhood information. However, existing methods focus on using graph convolution techniques to design feature extraction functions, ignoring the mutual optimization of the graph convolution operator and the self-expression coefficient matrix, leading to suboptimal clustering results. In addition, these methods directly construct graphs on raw data, which may be easily affected by noises and then degrade the clustering performance, as the constructed topology is not credible for the training procedure. To address these issues, we propose a novel method called spatial-spectral adaptive graph convolutional subspace clustering (S ² AGCSC). We employ the reconstruction coefficient matrix to devise a graph convolutional operator with adjacency matrix, which collaboratively computes both the feature representations and coefficient matrix, and the graph-convolutional operator is updated iteratively and adaptively during training. In addition, we harness a combination of spectral and spatial features to introduce additional view information to help learn more robust features and generate more refined superpixels. Experimental validation on three HSI datasets confirms the efficacy of S ² AGCSC.