Clustering-Guided Spatial-Spectral Mamba for Hyperspectral Image Classification

TL;DR

The paper tackles the challenge of inefficient token sequencing in hyperspectral image classification with Mamba-based models. It introduces CSSMamba, a clustering-guided spatial-spectral framework that uses CSpaMamba to create cluster-specific, shorter token sequences and integrates a SpeMamba for spectral information. A Dual Attention Module dynamically sorts tokens while a Learnable Clustering Module adaptively updates cluster memberships, yielding improved spatial-spectral learning and boundary preservation. Experiments on PU, IP, and LN01 demonstrate higher accuracy and clearer boundaries than state-of-the-art CNN, Transformer, and Mamba-based methods, highlighting the method's practical impact for HSI analysis.

Abstract

Although Mamba models greatly improve Hyperspectral Image (HSI) classification, they have critical challenges in terms defining efficient and adaptive token sequences for improve performance. This paper therefore presents CSSMamba (Clustering-guided Spatial-Spectral Mamba) framework to better address the challenges, with the following contributions. First, to achieve efficient and adaptive token sequences for improved Mamba performance, we integrate the clustering mechanism into a spatial Mamba architecture, leading to a cluster-guided spatial Mamba module (CSpaMamba) that reduces the Mamba sequence length and improves Mamba feature learning capability. Second, to improve the learning of both spatial and spectral information, we integrate the CSpaMamba module with a spectral mamba module (SpeMamba), leading to a complete clustering-guided spatial-spectral Mamba framework. Third, to further improve feature learning capability, we introduce an Attention-Driven Token Selection mechanism to optimize Mamba token sequencing. Last, to seamlessly integrate clustering into the Mamba model in a coherent manner, we design a Learnable Clustering Module that learns the cluster memberships in an adaptive manner. Experiments on the Pavia University, Indian Pines, and Liao-Ning 01 datasets demonstrate that CSSMamba achieves higher accuracy and better boundary preservation compared to state-of-the-art CNN, Transformer, and Mamba-based methods.

Figures (3)

• Figure 1: Traditional Mamba token sequencing (top) treats an entire image—despite its highly heterogeneous spatial patterns—as a single long sequence. This results in extremely long global token streams, which suffer from correlation decay and are inefficient at modeling fine-scale structures and subtle spatial variations. In contrast, our Per-Cluster Mamba Token Sequencing approach (bottom) decomposes a large, complex image into multiple smaller, locally homogeneous clusters to better capture subtle and weak patterns and edges, and constructs a dedicated token sequence for each cluster. This leads to significantly shorter sequences with more stationary feature distributions, making them easier to learn and better suited for capturing weak signals, fine details, and class boundaries. Furthermore, unlike conventional Mamba methods that scan tokens in a fixed, dense, and predefined order, our approach sequences pixels within each cluster using a dynamic, learnable, sparse, and adaptive strategy. This flexibility enables more effective modeling of local structures, improves detail preservation, and enhances the network’s ability to learn discriminative features from complex imagery.
• Figure 2: The proposed CSSMamba is a Spatial-Spectral dual-branch architecture that features a spectral branch and a cluster-guided spatial Mamba. The spatial branch leverages a learnable clustering mechanism, which reduces redundant computation by replacing dense grid scanning with dynamic, learnable, sparse and adaptable scanning. The Spatial-Spectral architecture is joined together to produce a feature map that jointly models dense photometric signatures and global semantic structure. A cluster-constrained dual loss guides the latent space to enforce intra-class compactness and inter-class separability, ensuring both spectral fidelity and spatial consistency.
• Figure 3: The Indian Pines (middle), Pavia University (top) and, LN01 (bottom) classification map generated by different methods. (a) Ours (b) ViT (c) ConvNeXt (d) SSTN (e) SSRN (f) SSFTT (g) MammbaHSI (h) SDMamba (i) PCA images of the datasets