HiMamba2: Hierarchical State Space Modeling With Linear Complexity for Hyperspectral Image Classification

Hyperspectral image (HSI) classification faces persistent challenges arising from high spectral dimensionality, complex spatial-spectral dependencies, and limited labeled samples. Although Transformer-based methods demonstrate strong long-range modeling capability, their quadratic computational complexity O(L²·d) imposes substantial bottlenecks for high-dimensional data. State space model (SSM) approaches achieve linear complexity O(L·d) but predominantly employ single-scale feature extraction and lack effective regularization mechanisms. This paper proposes HiMamba2, a hierarchical SSM architecture incorporating three principal contributions: parallel multi-scale convolutional embedding that captures complementary spectral-spatial representations across multiple receptive field scales; asymmetric weighted residual connections combined with adaptive DropPath regularization and serial Convolutional Block Attention Module (CBAM) dual-attention to enhance training stability and feature discriminability under limited labeled data conditions; an asymmetric dual-stage design with cross-layer fusion that integrates shallow fine-grained features with deep semantic representations while preserving linear complexity O(L·d). Experiments on Houston2013, WHU-Hi-HanChuan, and Pavia University datasets demonstrate overall accuracies of 93.40%, 93.02%, and 97.52%, respectively, outperforming 13 representative methods spanning CNN, attention, Transformer, and Mamba-based architectures. Training requires 262.84 seconds for 116-epoch convergence with 182.95 MB peak GPU memory consumption and 0.665 ms per-sample inference, validating the practical efficiency of the proposed approach. The complete PyTorch implementation of HiMamba2 will be made publicly available at https://github.com/Jason20155/ upon acceptance of this manuscript.