MambaSNN: An Unsupervised Hyperspectral Band Selection Method Based on State-Space Model and Spiking Neural Networks

Hyperspectral image (HSI) analysis critically relies on band selection (BS) to reduce dimensionality, improve classification accuracy, and minimize computational costs. Despite recent advances in unsupervised deep learning approaches, there exist some issues, including capturing long-range spectral dependencies, which can lead to substantial computational overhead when processing high-dimensional data. To address these challenges, we propose MambaSNN, a novel unsupervised BS method that integrates a state-space spectral module (SpeMamba) with a spiking-neural-network reconstruction module (SNNRecNet). SpeMamba captures intrinsic long-range spectral dependencies via recursive state transitions. This enables efficient and discriminative feature extraction with linear computational complexity. Meanwhile, SNNRecNet embeds leaky integrate-and-fire (LIF) spiking neurons into a convolutional reconstruction network, encouraging sparse spike-based encodings and alleviating the impact of noisy information, thereby preserving more discriminative spectral spatial information for BS. By jointly optimizing these modules, MambaSNN can effectively reduce spectral redundancy while maintaining classification performance. Comprehensive experiments have been conducted to validate the effectiveness of the proposed MambaSNN across five benchmark HSI datasets. Comparative results demonstrate that MambaSNN consistently outperforms state-of-the-art BS methods in both accuracy and runtime efficiency. These results validate the efficacy and practicality of the proposed approach for unsupervised HSI classification in high-dimensional, resource-constrained applications.