A Spiking and Memory-Enhanced State-Space Model for Hyperspectral Image Classification

Hyperspectral imaging (HSI) captures hundreds of contiguous spectral bands and enables fine-grained material discrimination. We present SpikeHopMamba, a hybrid architecture that combines neuromorphic spike encoding, a linearized Hopfield energy module, and selective state-space (SSSM) blocks to produce compact, multimodal patch tokens for HSI classification. SpikeHopMamba fuses static spectral tokens, sequentially encoded spiking features, and patch-level associative-energy cues, then models long-range spatial dependencies via efficient SSSM blocks. We train the model end-to-end with cross-entropy (and label smoothing) and report results averaged over five independent runs. Experiments on Hanchuan (HC), Salinas (SA), and OHID1–9 show that SpikeHopMamba attains competitive accuracy while preserving computational efficiency [overall accuracy (OA) up to 99.60% and κ up to 99.53%].