OLSS-Mamba: Order-Aware Low-Rank Spectral–Spatial Mamba for Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification requires effective spectral–spatial representation learning under limited labeled samples and constrained computational resources. Existing Mamba-based methods generally preserve explicit sequence structures through dense or fixed tokens, but this may retain substantial spectral–spatial redundancy. In contrast, adaptive token compression can reduce redundancy, whereas the resulting tokens often lack a consistent organization for order-sensitive state-space propagation. To address this issue, we propose an order-aware low-rank spectral–spatial (OLSS) Mamba network, termed OLSS-Mamba, for lightweight HSI classification. Its core idea is to couple a low-rank spectral–spatial prototype aggregation module for compact prototype construction with a spectral-anchor-guided prototype sequence Mamba module for ordinal state-space modeling. By incorporating spectral anchors directly into prototype generation, OLSS-Mamba constructs a compact, sample-adaptive, and ordinally organized prototype sequence rather than an unconstrained token set. Bidirectional Mamba is then employed to model the dependences among these prototypes, and the resulting contextual representation is fused with the center-pixel feature for classification. Experimental results on three representative HSI benchmarks show that OLSS-Mamba achieves overall accuracies of 98.10%, 90.17%, and 96.41%, respectively, while maintaining a compact model size and controlled computational cost. The experimental results demonstrate the effectiveness and superiority of the proposed OLSS-Mamba for accurate and efficient HSI classification.