3D-Selective Scan Module for Change Detection

• The module introduces a 3D-selective scan that extends 2D scanning to capture both spatial and channel dependencies for improved change detection.
• It processes input features along HW, HC, and WC planes to deliver comprehensive representations and refined boundary predictions.
• Integrated in a Siamese encoder-decoder pipeline, it enhances temporal feature interaction and synergizes with convolution and FFT cues.

The 3D-Selective Scan Module, abbreviated 3D-SSM, is a Mamba-style feature operator introduced for remote sensing change detection to extend selective scan from the usual 2D spatial traversal to a three-directional traversal over $HW$, $HC$, and $WC$, so that global information is captured from both the spatial plane and channel perspectives (Huang et al., 24 Jun 2025). In the proposed framework, 3D-SSM functions as the core feature operator in both bi-temporal interaction and decoding, and is paired with a spatiotemporal interaction module (SIM) and a multi-branch feature extraction module (MBFEM) to strengthen subtle change detection, contextual representation, and boundary accuracy.

1. Conceptual basis and motivation

Existing Mamba-based approaches in remote sensing change detection are described as being limited by their inability to capture long-range dependencies between image channels effectively. Methods built on SS2D or similar selective scan designs mainly scan only in the $HW$ spatial plane. That design is effective for long-range spatial dependency modeling, but it leaves a blind spot in that cross-channel dependencies are not deeply modeled as part of the scan itself (Huang et al., 24 Jun 2025).

In remote sensing change detection, this limitation is consequential because channels are not treated merely as feature slots. They encode discriminative semantic cues that can be critical for changes such as buildings, roads, vegetation, or shadows. A purely spatial scan may therefore miss the global relationship that spans both space and channel dimensions. The 3D-SSM is introduced to address this point directly by treating a feature map as a 3D entity in which spatial structure and channel interactions are modeled explicitly.

The stated goals of the module are fourfold: capturing long-range dependencies between distant pixels, modeling inter-channel relationships, producing more complete and boundary-accurate change representations, and preserving Mamba’s linear-complexity advantage relative to Transformer-style global attention. A plausible implication is that the module is intended less as a replacement for the internal SSM recurrence than as a redefinition of what constitutes the scan sequence in vision backbones.

2. Three-directional scan construction and mathematical form

The 3D-SSM takes an input feature map

$F_{in} \in \mathbb{R}^{H \times W \times C}$

and processes it from three perspectives: the $HW$ plane, the $HC$ plane, and the $WC$ plane (Huang et al., 24 Jun 2025). The $HW$ scan corresponds to the standard spatial scan used in 2D selective scan or SS2D. The additional $HC$ and $HC$0 scans expose height-channel and width-channel couplings to the same selective scan mechanism.

For the standard spatial view, the feature map is flattened into

$HC$1

The paper further writes

$HC$2

where $HC$3 is the scalar feature at height $HC$4, width $HC$5, and channel $HC$6. A Vision Mamba-style scanning operator $HC$7 is applied to $HC$8 and its transposed version, and the processed sequence is reshaped back into a 3D tensor.

For the channel-aware scans, the feature is further flattened as

$HC$9

These correspond to the two channel-coupled scan directions and are processed through the same selective scan formulation to produce $WC$0 and $WC$1. The final output is the additive fusion

$WC$2

This additive fusion is the module’s central structural choice. The $WC$3 scan captures spatial context, the $WC$4 scan captures relations along height-channel couplings, and the $WC$5 scan captures relations along width-channel couplings. Their sum yields a more comprehensive feature representation. The paper explicitly frames the novelty not as a new internal SSM cell, but as a rearrangement-and-scan strategy: standard SS2D-style scan for spatial context, plus extra scans after reshaping to expose channel relations. In that sense, 3D-SSM is a multi-view selective scan wrapper around a Mamba/VME block.

3. Role in the spatiotemporal interaction module

Within the Spatiotemporal Interaction Module, 3D-SSM is used to generate global contextual features for each timestamp, while a small 2D convolutional stack provides local detail refinement (Huang et al., 24 Jun 2025). Given encoder outputs from two timestamps,

$WC$6

the module computes

$WC$7

for $WC$8.

The interaction stage then cross-enhances the temporal branches. For example,

$WC$9

$HW$0

and the change-related output is

$HW$1

Here, $HW$2 is softmax, $HW$3 is element-wise multiplication, and $HW$4 is dynamic gating.

The dynamic gate is defined as

$HW$5

$HW$6

where the mean is taken along channels. This makes the fusion adaptive rather than fixed. In operational terms, SIM uses 3D-SSM to build strong global descriptors of each time point, then cross-interacts those descriptors with local features so that the resulting difference feature is more informative than a plain subtraction. This suggests that the contribution of 3D-SSM inside SIM is not only to enlarge receptive field, but also to stabilize bi-temporal comparison by supplying spatial-channel global context before differencing.

4. Decoder integration through MBFEM and the full detection pipeline

The Multi-Branch Feature Extraction Module is the decoder block that progressively refines change features. At decoder stage $HW$7, its input is formed by combining the upsampled previous decoder feature and the SIM feature at the same scale:

$HW$8

where $HW$9 is layer normalization (Huang et al., 24 Jun 2025).

MBFEM then extracts complementary information through multiple branches:

$F_{in} \in \mathbb{R}^{H \times W \times C}$0

$F_{in} \in \mathbb{R}^{H \times W \times C}$1

with $F_{in} \in \mathbb{R}^{H \times W \times C}$2 and $F_{in} \in \mathbb{R}^{H \times W \times C}$3. The learnable parameter $F_{in} \in \mathbb{R}^{H \times W \times C}$4 regulates the residual contribution of the stage input.

The roles of the branches are explicitly differentiated. The FFT branch emphasizes high-frequency information and boundaries, the convolution branch preserves local detail and edge structure, and the 3D-SSM branch supplies long-range spatial-channel context. The FFT branch is specified as

$F_{in} \in \mathbb{R}^{H \times W \times C}$5

where $F_{in} \in \mathbb{R}^{H \times W \times C}$6 and $F_{in} \in \mathbb{R}^{H \times W \times C}$7 are FFT and inverse FFT, $F_{in} \in \mathbb{R}^{H \times W \times C}$8 and $F_{in} \in \mathbb{R}^{H \times W \times C}$9 are patch unfolding and folding, $HW$0 is a learnable frequency modulation matrix, and $HW$1 is GEGLU. The convolution branch is

$HW$2

The complete framework is a Siamese encoder-decoder system. A VMamba-S backbone encoder extracts multi-scale features from the two images, SIM fuses bi-temporal features at each scale, MBFEM progressively decodes the difference features while enriching them with FFT, convolution, and 3D-SSM context, and a final $HW$3 convolution classifier predicts the change mask. The design is therefore not merely “encode each image and subtract.” Instead, 3D-SSM enriches each temporal representation with spatial-channel global context, SIM converts that into a discriminative difference feature, and MBFEM reconstructs a precise change map from coarse to fine scales.

5. Empirical evidence, advantages, and limitations

The experimental evidence reported for the module is largely organized through ablation. When SS2D is replaced with 3D-SSM, performance improves. On WHU-CD, 3D-SSM reaches 95.30 F1 and 99.63 OA, slightly better than SS2D and SS2D+CA. On SYSU-CD, 3D-SSM reaches 84.90 F1 and 92.98 OA, again best among the compared variants. Visualizations are reported to show more accurate boundaries than SS2D (Huang et al., 24 Jun 2025).

The scanning-perspective ablation clarifies why the design is three-directional rather than merely channel-aware. The $HW$4 scan alone performs strongly in precision but more weakly in recall. The $HW$5 or $HW$6 scan alone is weaker. Combining $HW$7 and $HW$8 improves over plain SS2D, and using all three views $HW$9 gives the best overall performance. This is presented as the strongest evidence that the three scanning views are complementary.

A parallel result appears in decoder ablations. Using only one MBFEM branch is weaker than using multiple branches, and the best results come from combining Conv + FFT + 3D-SSM. The reported interpretation is that 3D-SSM is complementary to local convolution and frequency-domain cues rather than redundant. SIM ablations point in the same direction: replacing SIM with simple absolute difference degrades performance, and SIM improves F1 by about 0.37 on WHU-CD and 0.47 on SYSU-CD compared with absolute difference.

The paper summarizes the main advantages of 3D-SSM as global spatial modeling with linear-time selective scan behavior, channel-aware context capture that prior SS2D-style Mamba change-detection methods largely ignore, better boundary quality and finer change localization, and complementarity with convolution and FFT in both SIM and MBFEM. It also reports favourable performance compared to state-of-the-art change detection methods on five benchmark datasets through extensive experiments.

The stated limitations are equally explicit. The method has long inference time relative to CNNs and Transformers, high FLOPs because scanning from three perspectives is computationally heavier, and residual boundary errors still remain in some cases even though results improve. The authors suggest pruning, quantization, and more efficient scanning as future directions. A plausible implication is that the principal open problem is not whether spatial-channel joint scanning is useful, but how to preserve that benefit with lower deployment cost.

6. Terminological scope and related uses of “selective scan”

A common source of confusion is that “3D-selective scan” is used across several research areas for mechanisms that are not equivalent. In the remote sensing sense introduced by 3D-SSM, the “3D” object is the feature tensor $HC$0, and the distinguishing operation is selective scan across spatial and channel planes rather than only across the spatial plane (Huang et al., 24 Jun 2025).

Related terminology in other arXiv papers refers to different constructions. VMambaMorph redesigns the 2D VMamba VSS block for 3D volumetric feature processing in medical image registration, but the paper explicitly does not provide the low-level 3D scan equations, tensor permutations, or explicit voxel traversal order (Wang et al., 2024). MASS is a motion-aligned selective scan for video frame interpolation, yet its scan is one-dimensional along flow-guided trajectories rather than a volumetric 3D traversal (Yoo et al., 26 Jun 2026). MarS3D is a plug-and-play motion-aware framework for multi-scan point-cloud segmentation and is explicitly described as not being a 3D selective scan in the classical state-space sense (Liu et al., 2023).

Outside machine vision, the phrase can become more distant still. Variable slab-selective projection acquisition in MRI is a rotated slab-selective excitation strategy in 3D projection imaging rather than a state-space selective scan operator (Park et al., 2023). The micro-volumetric scanning system in a 3D display is a projector-like focal-plane scanning mechanism that performs depth-wise layer scanning, again without a selective scan recurrence or feature-sequence operator (Wang, 2018).

Accordingly, the precise meaning of 3D-Selective Scan Module in current computer-vision usage depends on the paper’s modeling object and scan definition. In the specific sense established by 3D-SSM, it denotes a Mamba-style operator that extends selective scan from $HC$1 traversal to coordinated traversal over $HC$2, $HC$3, and $HC$4, thereby embedding both long-range spatial dependencies and inter-channel relationships into the change-detection representation.