Hyperspectral Mamba: State-Space HSI Models

• Hyperspectral Mamba is an emerging class of models that use selective state-space mechanisms to capture spectral and spatial dependency in hyperspectral data.
• It employs advanced sequence construction techniques, such as piece-wise scanning and dual-branch processing, to optimize performance across various HSI tasks.
• Empirical studies demonstrate state-of-the-art results in classification, denoising, super-resolution, and tracking with efficient computational requirements.

Searching arXiv for papers on hyperspectral Mamba and closely related methods. [Tool call suppressed in this environment: arXiv search used to verify relevant papers and dates.] Hyperspectral Mamba denotes an emerging class of hyperspectral learning models that use Mamba or closely related selective state-space mechanisms to model spectral, spatial, or joint spectral-spatial dependencies in hyperspectral data. In the broad literature, the term functions as an umbrella for hyperspectral image classification, denoising, super-resolution, target detection, and tracking systems built around linear-complexity sequence modeling; in a narrow paper-specific sense, “Hyperspectral Mamba (HSM)” is also the name of a module introduced inside the HyMamba tracker for hyperspectral object tracking (Gao et al., 10 Sep 2025). The category remains unsettled: the 2024 survey on hyperspectral image classification acknowledges Mamba only indirectly through citations such as SpectralMamba and does not define a dedicated HSM taxonomy, provide Mamba equations, or report Mamba experiments (Ahmad et al., 2024).

1. Nomenclature and scope

The literature uses “Hyperspectral Mamba” in two overlapping ways. First, it denotes a research family that adapts Mamba or selective state-space modeling to hyperspectral data. Second, it refers explicitly to the HSM module inside HyMamba, where three directional scanning SSMs are used to learn spatial and spectral information synchronously for hyperspectral object tracking (Gao et al., 10 Sep 2025). This ambiguity is substantive rather than terminological: some papers present full hyperspectral backbones built around Mamba, whereas others insert Mamba into a hybrid pipeline alongside CNNs, Transformers, graph operators, wavelets, or adapters.

A concise task-level view of representative systems is as follows.

Model	Task	Distinctive mechanism
SpectralMamba (Yao et al., 2024)	HSI classification	GSSM + PSS + selective SSM on spectral chunks
SS-Mamba (Huang et al., 2024)	HSI classification	Dual spectral/spatial token streams with center-guided enhancement
MiM (Zhou et al., 2024)	HSI classification	Centralized MCS + T-Mamba + WMF
DualMamba (Sheng et al., 2024)	HSI classification	Parallel Mamba/CNN global-local design
HSIDMamba (Liu et al., 2024)	HSI denoising	Bidirectional continuous scanning with eight directions
HSRMamba (Chen et al., 30 Jan 2025)	HSISR	Local 3D partitioning + global spectral reordering
HSRMamba (Li et al., 16 May 2025)	SHSR	Wavelet decomposition + stripe scanning
SpecMamba (Gong et al., 7 Apr 2026)	Few-shot HTD	DCT Mamba adapter + PGTE + SSPLM
HyMamba / HSM (Gao et al., 10 Sep 2025)	Hyperspectral tracking	SSI with forward, backward, and spectral SSMs

This scope also includes graph-augmented and wavelet-augmented systems such as GraphMamba and WaveMamba, which position Mamba as one component in a broader hyperspectral modeling stack (Yang et al., 2024, Ahmad et al., 2024). A plausible implication is that HSM is less a single architecture than a design space centered on state-space sequence modeling under hyperspectral structural constraints.

2. Mathematical basis and recurrent design principles

Most hyperspectral Mamba papers retain the classical state-space foundation

$$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t), \qquad y(t)=\mathbf{C}h(t),$$

together with zero-order-hold discretization and the recurrent form

$$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$

or closely related equivalents (Huang et al., 2024, Yao et al., 2024). Hyperspectral inputs are typically written as cubes such as

$x \in \mathbb{R}^{H \times W \times B}$

or analogous notation with $V$ or $C$ spectral channels, and the central modeling question is how to convert that cube into a sequence whose ordering preserves the dependencies that matter for the target task (He et al., 2024, Gong et al., 7 Apr 2026).

The distinctive contribution of most HSM systems is therefore not a new state equation, but a new deployment strategy for selective state-space modeling. SpectralMamba uses a standard selective SSM inside a pipeline composed of Gated Spatial-Spectral Merging and Piece-wise Sequential Scanning, thereby treating the spectral axis as the primary sequence while compressing it into chunks (Yao et al., 2024). IGroupSS-Mamba likewise keeps the S6 formulation but alters its use through interval grouping and direction assignment, applying one unidirectional S6 to each non-overlapping interval group rather than scanning all channels in all directions (He et al., 2024). MiM preserves the selective Mamba operator but wraps it in center-oriented scan construction, Gaussian masking, semantic token learning, and multi-scale supervision (Zhou et al., 2024).

A recurring principle is that hyperspectral Mamba systems treat the sequence itself as a learned or engineered object. This is visible in spectral chunking, grouped directional scans, bidirectional continuous scans, stripe scans, and sparse deformable sequencing. The literature therefore suggests that, in hyperspectral settings, Mamba effectiveness depends at least as much on sequence formation and inductive bias as on the internal SSM recurrence.

3. Sequence construction as the central design problem

Sequence construction is the most persistent technical theme in hyperspectral Mamba work. SpectralMamba reduces sequence length by splitting a spectral vector into contiguous pieces through Piece-wise Sequential Scanning, so that the original spectral length $L$ becomes a sequence of $R$ chunk tokens of width $C$, preserving short-range context within each chunk and long-range context across chunks (Yao et al., 2024). SS-Mamba instead generates two explicit token sets—spatial tokens from the whole patch and spectral tokens from a center subcube—then processes them in parallel Mamba branches (Huang et al., 2024).

MiM makes sequence order center-aware. Its centralized Mamba-Cross-Scan constructs four scan types, each split into two sub-sequences that begin at opposite sides of the patch and terminate at the center pixel, so the center becomes the final aggregation point for both directional branches (Zhou et al., 2024). HSIDMamba generalizes scanning in another direction: it serializes features by a bidirectional continuous scanning mechanism with forward and backward continuous scans from eight directions, then fuses the directional outputs with residual and spectral attention modules (Liu et al., 2024).

Later work pushes sequence design further into hyperspectral-specific priors. HSRMamba for single-image hyperspectral super-resolution introduces local 3D spatial-spectral partitioning and global spectral reordering, arguing that naive 1D serialization destroys local adjacency and that Mamba is highly sensitive to input order (Chen et al., 30 Jan 2025). IGroupSS-Mamba partitions channels by interval—channels $1,5,9,\ldots$ in one group, $2,6,10,\ldots$ in another—and assigns left-to-right, right-to-left, top-to-bottom, and bottom-to-top scans to the four groups, reducing redundancy from adjacent-band similarity while preserving directional complementarity (He et al., 2024). SDMamba treats sequence construction explicitly as adaptive selection: spatial and spectral tokens are ranked by angular similarity to an anchor token, sorted, sparsified by a ratio $$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$0, and only the retained ordered subset is sent to Mamba (Xu et al., 13 Apr 2025). The wavelet-based HSRMamba for SHSR replaces global 1D visual scanning with fixed-length stripe scanning, arguing that stripe scanning introduces cross-window connections and better balances global and local modeling than either global 1D scans or disjoint windows (Li et al., 16 May 2025).

Taken together, these designs suggest that “sequence engineering” is a defining trait of HSM. A plausible implication is that hyperspectral Mamba research has converged less on one canonical backbone than on a shared claim: the spectral-spatial cube should not be flattened naively.

4. Architectural families

One major family uses dual-branch spectral-spatial factorization. SS-Mamba generates spatial and spectral token sequences, processes them with two basic Mamba blocks, and couples them through a spectral-spatial feature enhancement module derived from center-region information (Huang et al., 2024). DualMamba follows the same separation but assigns roles asymmetrically: a lightweight spectral/spatial Mamba branch extracts global features, a residual convolution branch extracts local features, and adaptive global-local fusion combines them dynamically (Sheng et al., 2024). WaveMamba extends this factorization with separate spatial and spectral gates followed by Haar wavelet decomposition of both branches before state-space modeling (Ahmad et al., 2024).

A second family uses preconditioning modules before Mamba. SpectralMamba learns a dynamic spatial-spectral mask by depthwise and pointwise convolutions to merge a local patch into one spectrum, then applies selective SSM to chunked spectral tokens (Yao et al., 2024). MiM uses Gaussian Decay Mask to suppress peripheral interference after directional Mamba encoding and then learns compact semantic tokens through STL and STF (Zhou et al., 2024). GraphMamba combines HyperMamba or state-space-style sequence learning with graph-based spatial context, using weighted multi-hop aggregation in one version and token prioritization plus graph affinity and cross-attention in another (Yang et al., 2024, Ahmad et al., 10 Feb 2025).

A third family is frequency-aware or wavelet-aware HSM. WaveMamba uses fixed Haar wavelets to produce $$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$1, $$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$2, $$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$3, and $$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$4 subbands from gated spatial and spectral features before feeding the fused representation into a state-space recurrence (Ahmad et al., 2024). The SHSR-oriented HSRMamba uses discrete wavelet transform to split low-frequency spectral structure from high-frequency spatial detail, then assigns VSSM-based global modeling and local convolutional processing to dedicated branches (Li et al., 16 May 2025). SpecMamba shifts frequency awareness into few-shot target detection by applying DCT along the spectral dimension, grouping coefficients into low-, mid-, and high-frequency bands, and using a Mamba-based adapter on top of frozen Transformer representations (Gong et al., 7 Apr 2026).

A fourth family is task-specialized state propagation. HSIDMamba uses hyperspectral continuous scan blocks for denoising, with bidirectional continuous scanning, residual refinement, and channel attention (Liu et al., 2024). HSRMamba for super-resolution uses contextual spatial-spectral Mamba groups that alternate local structured scans and globally reordered scans (Chen et al., 30 Jan 2025). HyMamba places HSM inside Spectral State Integration so that forward, backward, and spectral hidden states are propagated across network depth and across video frames for tracking (Gao et al., 10 Sep 2025).

5. Empirical record across tasks

In hyperspectral image classification, several HSM systems report state-of-the-art or near-state-of-the-art results under their own protocols. SpectralMamba reports patchwise OA values of 89.52 on Houston2013, 77.90 on Augsburg, 92.48 on Longkou, and 97.66 on Botswana, with markedly low MACs and parameter counts relative to the compared RNN and Transformer baselines (Yao et al., 2024). SS-Mamba reports OA = 91.59 ± 1.85 on Indian Pines, 96.40 ± 2.27 on Pavia University, and 94.30 ± 1.10 on Houston, and its sequence-model ablation finds Mamba stronger than LSTM, GRU, and Transformer within the same spectral-spatial framework (Huang et al., 2024). DualMamba reports OA 99.23%, 99.66%, and 97.47% on Indian Pines, WHU-Hi-Longkou, and Houston 2018 while using 72.94K, 82.86K, and 58.23K parameters, respectively, and the fewest FLOPs among the compared methods (Sheng et al., 2024). IGroupSS-Mamba reports OA 98.71 on Indian Pines, 99.75 on Pavia University, and 99.45 on Houston 2013 with only about 0.057 M parameters and 0.0095 G FLOPs (He et al., 2024). SDMamba reports OA = 99.44 on Indian Pines and 99.14 on Pavia University, and its sparsity ablation reduces FLOPs from 416.23M to 172.41M in the reported comparison (Xu et al., 13 Apr 2025). MiM reports OA = 92.0794, 91.5756, and 92.8917 on Indian Pines, Pavia University, and Houston 2013, respectively, while outperforming vanilla ViM and several Transformer baselines under fixed and disjoint splits (Zhou et al., 2024).

For denoising and super-resolution, the empirical record is similarly task-specific. HSIDMamba reports 41.30 dB / 0.9950 / 0.0459 on the ICVL “Mixture” setting and, on CAVE at $$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$5, 0.68M parameters, 0.77 s runtime, and 35.99 PSNR / 0.9867 SSIM / 0.3221 SAM, outperforming recent Transformer baselines in the reported table (Liu et al., 2024). The contextual HSRMamba reports, on Chikusei at $$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$6, PSNR 40.2781, SSIM 0.9441, SAM 2.3160, CC 0.9557, and ERGAS 5.0131, and on Houston at $$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$7, PSNR 46.9653, SSIM 0.9838, SAM 1.6577, CC 0.9891, and ERGAS 1.8277 (Chen et al., 30 Jan 2025). The wavelet-stripe HSRMamba states that it outperforms existing methods and reduces computational load and model size, but the provided excerpt does not include benchmark numbers beyond that qualitative claim (Li et al., 16 May 2025).

For target detection and tracking, SpecMamba and HyMamba extend HSM beyond pixel classification. SpecMamba reports the highest $$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$8 on all four reported datasets—0.99927 on San Diego I, 0.98915 on San Diego II, 0.99994 on Cri, and 0.99942 on Urban-1—and strong cross-domain performance through DCTMA, PGTE, and SSPLM (Gong et al., 7 Apr 2026). HyMamba reports 73.0\% AUC and 96.3\% DP@20 on HOTC2020, and its ablations show that replacing the simpler MM with HSM inside SSI improves AUC from 0.711 to 0.730 and DP@20 from 0.934 to 0.963 (Gao et al., 10 Sep 2025).

These results are not directly comparable across datasets, sampling protocols, or tasks. What they do show consistently is that hyperspectral Mamba systems are no longer confined to classification and now span restoration, detection, and tracking with task-specific sequence designs.

6. Limitations, ambiguities, and open directions

The most immediate limitation is taxonomic. The survey that frames the evolution of hyperspectral classification does not actually survey Mamba in depth, does not define HSM, and does not provide Mamba experiments or equations, so the category remains only partially institutionalized in the broader HSI literature (Ahmad et al., 2024). This helps explain why “Hyperspectral Mamba” can mean a general family in one context and a tracking submodule in another (Gao et al., 10 Sep 2025).

A second limitation is methodological heterogeneity. Not every “Mamba” paper in this space implements canonical selective state-space modeling. HSIMamba is explicitly Mamba-inspired but does not instantiate the standard selective SSM formulation and instead uses bidirectional Conv1d pathways with learned matrices $$h_t=\overline{\mathbf{A}}h_{t-1}+\overline{\mathbf{B}}x_t,\qquad y_t=\mathbf{C}h_t,$$9 and $x \in \mathbb{R}^{H \times W \times B}$0 plus a spatial block (Yang et al., 2024). The hybrid GraphMamba paper from 2025 is framed as state-space/Mamba-based, but the explicit recurrence written in the paper is GRU-style rather than canonical selective scan (Ahmad et al., 10 Feb 2025). WaveMamba presents a simplified state-space recurrence and omits the usual Mamba-specific selective-scan equations, token-order details, and several implementation specifics (Ahmad et al., 2024). These cases indicate that HSM is partly a branding layer over a broader state-space design space.

A third limitation is reproducibility and specification. Several papers leave out exact scan-order formulas, parameter counts, or efficiency breakdowns. SpecMamba emphasizes parameter-efficient adaptation but does not report FLOPs, latency, or exact trainable parameter counts in the provided description (Gong et al., 7 Apr 2026). HyMamba reports strong tracking accuracy but does not provide parameter count, FLOPs, FPS, or memory usage in the excerpt (Gao et al., 10 Sep 2025). Even strong classification papers can contain partial inconsistencies: WaveMamba’s abstract claims a 4.5\% improvement on Houston and 2.0\% on Pavia, but the main benchmark tables in the provided description do not support those margins relative to the strongest listed baseline (Ahmad et al., 2024). There is also a naming ambiguity: two distinct 2025 super-resolution papers are both titled “HSRMamba,” but they use materially different mechanisms—contextual spatial-spectral partitioning and reordering in one case, wavelet stripe state-space modeling in the other (Chen et al., 30 Jan 2025, Li et al., 16 May 2025).

The technical direction of travel is nevertheless clear. The literature increasingly couples Mamba with frequency-domain structure, wavelets, graph context, physical priors, test-time adaptation, and hybrid global-local fusion rather than using Mamba as a monolithic substitute for attention. This suggests that future HSM systems are likely to remain hybrid. A plausible implication is that the most durable contribution of hyperspectral Mamba research may not be a single canonical backbone, but a set of design rules: respect spectral continuity, engineer scan order explicitly, preserve local 3D structure, and combine linear-complexity state propagation with domain-specific priors when labels are scarce or domain shift is severe.