DC-Mamba: Diverse Mamba-Based Models

• DC-Mamba is a family of architectures that leverage Mamba-based state-space modeling augmented with auxiliary modules to address domain-specific challenges.
• In time series forecasting, a dual-channel encoder processes temporal and variable tokens separately, leading to significant reductions in MSE and MAE.
• Variants in remote sensing and reinforcement learning integrate modules like deformable alignment and sub-goal prompting to balance efficiency and precision.

Recent arXiv usage shows that DC-Mamba refers to several distinct Mamba-derived architectures rather than a single universally standardized model. The label appears in multivariate time series forecasting as a dual-channel model that couples Bi-Mamba with a linear Transformer (Fan et al., 6 Jul 2025), in remote sensing change detection as an “align-then-enhance” extension to ChangeMamba (Sun et al., 19 Sep 2025), in reinforcement learning as Decision Mamba and its hierarchical or hybrid variants that replace or complement Decision Transformer-style backbones (Correia et al., 2024, Huang et al., 2024), and in semi-supervised medical image segmentation as Diversity-enhanced Collaborative Mamba (Li et al., 19 Aug 2025). A related remote-sensing architecture, CDMamba, further clarifies how Mamba is adapted to dense change detection through explicit global-local feature integration (Zhang et al., 2024). This usage pattern suggests that DC-Mamba is best understood as a family of domain-specific Mamba adaptations organized around complementary modeling channels, guidance signals, or diversity mechanisms.

1. Nomenclature and scope

The term is not used uniformly across the cited literature. In one case it denotes a dual-channel forecasting architecture; in another, a remote-sensing change-detection framework; in others, closely related abbreviations denote Decision Mamba or Diversity-enhanced Collaborative Mamba. The common denominator is not a single architecture, but the use of Mamba or SSM-based sequence modeling as a core computational primitive, typically supplemented by auxiliary mechanisms that compensate for known limitations of pure selective scanning, such as weak explicit global interaction, missing local detail, geometric misalignment, or insufficient training diversity (Fan et al., 6 Jul 2025, Sun et al., 19 Sep 2025, Correia et al., 2024, Huang et al., 2024, Li et al., 19 Aug 2025, Zhang et al., 2024).

Variant name	Application domain	Core idea
DC-Mamber	Multivariate time series forecasting	Dual-channel Bi-Mamba + linear Transformer
DC-Mamba	Remote sensing change detection	BTDA + SSCA on ChangeMamba
Decision Mamba / HDM	Offline RL / imitation learning	Replace DT/HDT Transformer stacks with Mamba
DM-H	In-context RL	Mamba long-memory sub-goals + Transformer local prediction
DCMamba	Semi-supervised medical segmentation	Diversity from data, network, and feature levels
CDMamba	Remote sensing binary change detection	Global-local ConvMamba + adaptive temporal fusion

A recurrent feature across these uses is architectural specialization. Mamba is rarely deployed as an isolated drop-in replacement; instead, it is paired with tokenization choices, auxiliary fusion modules, hierarchical control structures, or contrastive objectives. A plausible implication is that the most productive uses of Mamba in applied vision and sequence modeling arise when its linear-time state-space dynamics are matched to a task-specific inductive bias rather than treated as a complete substitute for attention or convolution.

2. Dual-channel forecasting in multivariate time series

In multivariate time series forecasting, DC-Mamber is defined as a dual-channel architecture designed to exploit the complementary strengths of Mamba and Transformer-style modeling (Fan et al., 6 Jul 2025). The problem formulation distinguishes two prevailing strategies. Channel-independent modeling treats each variable’s full temporal history as a token and is suited to intra-variable dynamics; channel-mixing modeling treats the multivariate snapshot at each timestamp as a token and is suited to global temporal dependency modeling. The architecture explicitly assigns these two views to different encoders instead of forcing one representation to satisfy both roles.

Given an input history $\mathbf{X}\in\mathbb{R}^{B\times L\times V}$, the model normalizes the input and constructs two parallel representations: a temporal token representation $\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$ and a variable token representation $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$. These are produced by two separate embedding layers. The temporal branch preserves the time axis as the token axis, adds positional encoding, and maps the variable dimension $V$ into latent dimension $D$. The variable branch transposes the input so that each variable becomes a token containing its full temporal trajectory, then projects the look-back length $L$ into $D$. The paper states that the use of two dedicated embeddings avoids the “information coupling” problem of a single shared representation.

The V-Encoder operates on $\mathbf{X}_{var}$ and is built on Bi-Mamba. Its layer structure is residual Bi-Mamba, LayerNorm, and an MLP block implemented with 1D convolutions and ReLU. The bidirectional formulation scans both forward and backward,

$$\overrightarrow{Y}=\overrightarrow{\text{Mamba}(X_{var})},\qquad \overleftarrow{Y}=\overleftarrow{\text{Mamba}(X_{var})},\qquad Y_{\text{Bi-Mamba}}=\overrightarrow{Y}+\overleftarrow{Y},$$

with the stated goal of improving contextual aggregation for variable-wise temporal structure. The underlying SSM is summarized in continuous time as

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

followed by zero-order hold discretization and selective SSM computation.

The T-Encoder operates on $\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$0 and is responsible for channel-mixing global temporal dependency modeling. It uses a linear-attention layer, residual connection, LayerNorm, and an MLP block with 1D convolutions. To avoid the $\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$1 cost of full self-attention, the paper adopts a Linformer-style low-rank projection:

$\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$2

and computes attention with projected keys and values, reducing complexity to $\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$3. The architectural claim is that linear attention is well matched to timestamp tokens containing cross-variable snapshots, whereas Bi-Mamba is well matched to variable tokens containing local temporal evolution.

After encoding, the temporal feature map $\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$4 is projected to $\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$5 so that it can be concatenated with the variable feature map $\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$6. A Feature-fusion layer applies

$\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$7

where $\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$8, and a projector maps the fused representation to $\mathbf{X}_{tem}\in\mathbb{R}^{B\times L\times D}$9 before permutation to $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$0.

The model is evaluated on eight public datasets—ETTm1, Weather, ECL, Solar-Energy, PEMS03, PEMS04, PEMS07, and PEMS08—using MSE and MAE. The reported training setup is PyTorch 3.10.9 on an NVIDIA GeForce RTX 3090 with CUDA 11.4, L2 loss, Adam optimizer, and 10 epochs with early stopping. Baselines include S-Mamba, iTransformer, RLinear, PatchTST, Crossformer, TiDE, TimesNet, DLinear, SCINet, FEDformer, and Stationary. The paper reports a 4.2% reduction in MSE and 4.9% reduction in MAE on average across datasets, with especially large gains on traffic datasets. Against S-Mamba, the reported traffic improvements include PEMS03 MSE 0.136 $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$1 0.106 and MAE 0.241 $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$2 0.214, PEMS04 MSE 0.096 $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$3 0.083 and MAE 0.204 $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$4 0.188, PEMS07 MSE 0.088 $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$5 0.081 and MAE 0.188 $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$6 0.177, and PEMS08 MSE 0.156 $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$7 0.147 and MAE 0.243 $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$8 0.209.

The ablation results are central to the method’s interpretation. Removing either encoder degrades performance; swapping encoder roles hurts performance; and using only one tokenization strategy is inferior to using both. The paper also claims linear-time complexity in each branch and notes hyperparameter sensitivity to learning rate, $\mathbf{X}_{var}\in\mathbb{R}^{B\times V\times D}$9, and $V$0. The stated limitations are mostly architectural: manual dual tokenization and encoder assignment, relatively simple feature fusion, and extra parameters and engineering complexity relative to a single-stream model.

3. “Align-then-enhance” remote sensing change detection

In remote sensing change detection, DC-Mamba is a lightweight, plug-and-play enhancement to the ChangeMamba backbone that addresses two persistent problems: geometric misalignment between bi-temporal images and weak separability of subtle change signals from noise (Sun et al., 19 Sep 2025). The framework is explicitly organized as “align-then-enhance”. It keeps the ChangeMamba backbone and task head unchanged, and inserts two modules: Bi-Temporal Deformable Alignment (BTDA) and Scale-Sparse Change Amplifier (SSCA).

BTDA operates on high-level semantic features from the two temporal branches rather than on raw images. A lightweight convolutional head over concatenated bi-temporal features predicts a 2D offset field $V$1 and a scalar gate $V$2. The offset is bounded by

$V$3

and regularized through

$V$4

The aligned post-change feature is then obtained by gated deformable sampling,

$V$5

The gate can shrink toward $V$6 in already aligned regions, allowing the module to preserve useful semantics and avoid over-warping.

SSCA is applied before final classification and is designed to amplify reliable change evidence while suppressing noise. It aggregates four cues,

$V$7

where the absolute feature difference provides direct change evidence, the gradient terms emphasize edges and texture changes, and DSS captures structural differences robust to illumination variation. These cues produce a spatial gate map $V$8. Channel recalibration is computed through global average pooling and a two-layer MLP,

$V$9

and the final residual enhancement is

$D$0

The reported experimental setup uses a Tesla V100-SXM2-16GB, CUDA 12.6, Python 3.10, images resized to 1280 × 630, a dataset of 400 images total, a 9:1 train/val split plus a 50-image independent test set, and 1800 iterations. The paper reports the following final metrics for DC-Mamba: Precision 0.6871, Recall 0.5174, OA 0.7876, F1 0.5903, IoU 0.4187, and Kappa 0.4507. Relative to ChangeMamba, F1 improves from 0.5730 to 0.5903, IoU from 0.4015 to 0.4187, and precision from 0.5817 to 0.6871. It also reports that OA decreases slightly, while emphasizing that IoU and Kappa are more informative under class imbalance.

The ablation clarifies the role of each module. Adding BTDA to ChangeMamba increases precision from 0.5817 to 0.6218 while recall remains nearly unchanged at 0.5549, indicating reduced pseudo-changes from misregistration. Adding SSCA on top raises precision to 0.6871 but lowers recall to 0.5174, which the paper interprets as a deliberate trade-off favoring cleaner boundaries, suppression of noise, and more selective emphasis on high-confidence change pixels. The architectural contribution is therefore not a deeper backbone, but an explicit decomposition of RSCD into semantic alignment followed by sparse enhancement.

4. Decision Mamba in imitation learning and reinforcement learning

In reinforcement learning, the closely related designation Decision Mamba refers to replacing or augmenting Decision Transformer-style sequence models with Mamba-based SSMs (Correia et al., 2024, Huang et al., 2024). The motivating critique is consistent across both papers. Decision Transformer conditions action prediction on sequences such as $D$1, depends heavily on returns-to-go during training and deployment, requires a causal mask and positional encoding, and inherits quadratic attention cost. Mamba is proposed as an alternative because it provides linear-time sequence processing and a sequential state-evolution mechanism that can, according to the authors, substitute for explicit RTG guidance in some settings.

The first formulation introduces Decision Mamba (DM) and Hierarchical Decision Mamba (HDM). DM uses linear embedding layers for state, action, and optionally RTG streams, concatenates the embeddings into a sequential input, and passes them through repeated Mamba layers. The final output is projected to action space, followed by tanh and rescaling to the task action range. DM is explicitly evaluated both with RTG and without RTG, and the paper states that it does not need causal masking or positional encoding. HDM retains the hierarchical decomposition of HDT: a high-level planner predicts sub-goals, and a low-level controller predicts actions conditioned on state, action, and sub-goal histories. The sub-goal augmentation uses the highest-valued future state according to

$D$2

Both high-level and low-level models are trained with L2 loss.

This work frames offline imitation learning on a dataset of trajectories $D$3 with $D$4. It evaluates on seven D4RL environment families—Ant, Antmaze, HalfCheetah, Hopper, Kitchen, Maze2d, and Walker2d—using maximum accumulated return, reported as the best average return over validation episodes during training. The stated setup uses 1 million epochs, batch size 16, learning rate 1e-4, 6 layers, embedding size 128, and context length 20. The reported summary is that DM outperforms DT in 15 out of 27 settings, HDM outperforms HDT in 16 out of 27 settings, and HDM is the best overall model. A notable behavioral result is that DM without RTG still works well, whereas DT is framed as strongly dependent on RTG specification.

A second RL paper tests a plain Mamba replacement and then proposes Decision Mamba-Hybrid (DM-H). Here the argument is more restrictive: a plain Mamba backbone improves efficiency but may be slightly worse in effectiveness than transformer-based decision models, because RL trajectories exhibit fine-grained sequential dependencies that attention handles well. DM-H therefore uses Mamba to process multiple historical trajectories and produce a sub-goal token every $D$5 steps, then prompts a Transformer/GPT-style decoder with this sub-goal and a short local context. The across-episodic context is written as $D$6, and the local prompted sequence as

$D$7

To force the Transformer to use the sub-goal, the paper introduces valuable sub-goals from offline data, selecting high-reward near-future states by dividing future accumulated reward by distance.

DM-H is evaluated on Grid World, Tmaze, and D4RL benchmarks. The reported result is state-of-the-art in long and short-term tasks, including a total average of 87.53 on the reported D4RL table, above DT: 77.69, AD (Transformer): 82.84, and AD (Mamba): 76.97. The efficiency claim is especially prominent: 28× faster online testing than transformer-based baselines in long-term tasks. Across both RL papers, a stable conclusion is that Mamba is attractive for long-horizon memory and efficiency, but the strongest variants often remain hybrid or hierarchical rather than purely Mamba-based.

5. Diversity-enhanced Collaborative Mamba in semi-supervised medical image segmentation

In semi-supervised medical image segmentation, DCMamba denotes Diversity-enhanced Collaborative Mamba, a framework built around the principle that useful supervision can be extracted from diversity at the data, network, and feature levels (Li et al., 19 Aug 2025). The method uses two Mamba-based U-Net branches, $D$8 and $D$9, and is explicitly motivated by the observation that the scan direction in visual Mamba affects context accumulation and can therefore induce complementary segmentation behaviors.

The first diversity source is patch-level weak-strong mixing augmentation. For labeled and unlabeled sets

$L$0

each image is divided into $L$1 patches,

$L$2

and weak and strong augmentations $L$3 and $L$4 are applied patchwise before random mixing into two views $L$5 and $L$6. The stated rationale is that patch-level mixing matches Mamba’s patchwise sequence modeling more naturally than CutMix or MixUp while preserving anatomical structure.

The second diversity source is network diversity through distinct scan directions. Branch $L$7 uses a horizontal + vertical scan module, whereas branch $L$8 uses a diagonal + anti-diagonal scan module. The paper notes that each network scans in four routes—top-left to bottom-right, bottom-right to top-left, top-right to bottom-left, and bottom-left to top-right—and that the two branch types emphasize different structural dependencies. Cross-supervised pseudo-labeling on unlabeled data uses

$L$9

and the unsupervised loss is the sum of Dice and cross-entropy terms in which each branch is trained against the other branch’s pseudo-labels.

The third diversity source is uncertainty-weighted contrastive learning. Each branch produces four route-specific features $D$0, and uncertainty is estimated through feature variance across routes,

$D$1

The fused feature is

$D$2

and same-image features from different branches are pulled together in embedding space through a contrastive objective,

$D$3

The full training objective is

$D$4

with Gaussian warm-up

$D$5

The framework is evaluated on six public datasets. The 2D tasks are Synapse, ISIC17, ISIC18, and Kvasir-SEG; the 3D tasks are 3D Synapse, AbdomenCT-1K, and FLARE22. The backbone is Mamba-UNet initialized from VMamba-Tiny pre-trained on ImageNet-1k. The reported optimizer is SGD with momentum 0.9, weight decay 1e-4, base learning rate 0.01, and 20k iterations for most datasets, or 10k for AbdomenCT-1K. The paper emphasizes large gains over Semi-Mamba-UNet, including +6.69% Dice on Synapse with 20% labeled data and +4.52% Dice in the 10% labeled setting. Detailed Synapse results are 67.27 ± 0.81 Dice and 9.28 ± 0.74 ASD at 20% labeled, versus 60.58 ± 3.61 Dice and 10.99 ± 5.90 ASD for Semi-Mamba-UNet; and 60.41 ± 0.84 Dice and 8.96 ± 0.70 ASD at 10% labeled, versus 55.89 ± 2.43 Dice and 12.88 ± 0.54 ASD.

The ablation shows that diversity is cumulative: on Synapse, the baseline without all three components gives 58.00 Dice, adding only diverse augmentation gives 61.07, only diverse scan gives 61.62, only diverse feature gives 59.09, and the full model gives 67.63. A separate ablation reports 85.01 Dice for plain contrastive loss and 86.42 Dice for the uncertainty-weighted diverse-feature variant. The framework therefore treats disagreement not as noise to be minimized immediately, but as a structured source of supervision.

6. Related change-detection variants and recurrent design patterns

A closely related remote-sensing model, CDMamba, helps situate the broader DC-Mamba family by making explicit a design tension that also appears in other variants: Mamba supplies efficient global context, but dense prediction requires additional mechanisms for local detail, temporal guidance, or both (Zhang et al., 2024). CDMamba is a Siamese bi-temporal encoder-decoder for binary change detection that introduces a Scaled Residual ConvMamba (SRCM) block and an Adaptive Global Local Guided Fusion (AGLGF) block. SRCM combines global SSM-based sequence modeling with local convolutional detail and a scaled residual connection,

$D$6

AGLGF performs global-guided and local-guided bi-temporal fusion, then adaptively combines them through learned gating scores before absolute subtraction and decoding.

The reported CDMamba results are F1 93.76 / IoU 88.26 on WHU-CD, F1 90.75 / IoU 83.07 on LEVIR-CD, and F1 83.01 / IoU 70.95 on LEVIR+-CD. The paper states improvements over compared Mamba-based methods of +0.97 / +1.21 in F1/IoU on WHU-CD, +0.98 / +0.59 on LEVIR-CD, and +2.10 / +2.24 on LEVIR+-CD. Its ablation likewise isolates complementary contributions: a baseline Mamba-like model gives 87.31 F1 / 77.48 IoU on WHU-CD, SRCM alone gives 92.83 / 86.61, AGLGF alone gives 91.32 / 84.03, and full CDMamba gives 93.76 / 88.26.

Taken together, the broader DC-Mamba literature exhibits several recurring patterns. First, pure Mamba is often judged insufficient for the target task in isolation: forecasting models add a linear Transformer branch, RL models introduce hierarchical or hybrid prompting, RSCD models add alignment or enhancement modules, and dense vision models incorporate convolution or contrastive learning. Second, tokenization and view selection are treated as first-class design choices. The dual-channel forecasting model separates temporal tokens from variable tokens; DCMamba turns scan direction into a supervisory signal; DM-H compresses across-episodic memory into sub-goal tokens. Third, reported gains frequently reflect targeted error correction rather than uniform improvement across all metrics. BTDA primarily improves precision; SSCA sharpens predictions at some recall cost; plain Decision Mamba may improve efficiency while trailing transformers in some settings; dual-channel forecasting adds parameters and engineering complexity even while improving accuracy.

These patterns also address several potential misconceptions. DC-Mamba is not a single benchmark-standard architecture, not always a pure Mamba replacement, and not uniformly attention-free or convolution-free. In many instances, the central contribution lies in how Mamba is paired with another mechanism—linear attention, deformable alignment, sub-goal prompting, diverse scanning, or global-local fusion—so that linear-time state-space modeling is used where it is strongest and compensated where it is weakest.