Answers are grounded in this paper and cite their sources.
Hyperspectral image (HSI) classification requires learning rich spatial-spectral representations over hundreds of contiguous spectral bands. A key challenge is to capture both fine local spatial–spectral structures and long-range contextual dependencies efficiently. CNN-based methods are effective in modeling local patterns but are often limited by their receptive fields when global context is required. Transformer-based methods can capture long-range dependencies, yet their attention mechanism typically incurs quadratic complexity. These limitations motivate a linear-time model that jointly captures local details and long-range context. In this paper, we propose HCxLSTM, a dual-branch framework built on extended Long Short-Term Memory (xLSTM) for hyperspectral image classification. Our design employs Paired Blocks to process patch tokens in alternating directions, enabling a multi-directional scanning that preserves comprehensive spatial cues. One branch leverages mLSTM (matrix-based memory) for richer feature capacity and parallelizable state updates, while the other utilizes sLSTM ( scalar-based memory) for fine-grained gating and frequent revisions. We further introduce a CrossmLSTM module that fuses upper-branch “queries” with lower-branch “key–value” features in a linear-time fashion. Through this two-branch design, HCxLSTM captures both long-range contextual relationships and localized spatial–spectral details. Extensive experiments on four HSI datasets demonstrate that our method outperforms state-of-the-art approaches in accuracy and efficiency.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: