Mamba-FCS: Semantic Change Detection

• Mamba-FCS is a semantic change detection framework that integrates a Visual State Space Model with joint spatio-frequency fusion for bi-temporal remote sensing imagery.
• It employs a Siamese multi-task architecture, coupling binary change detection with semantic decoding through change-guided attention to enhance edge precision.
• The framework leverages FFT-based frequency enhancement and a Separated Kappa loss to achieve state-of-the-art performance on datasets like SECOND and Landsat-SCD.

Mamba-FCS is a semantic change detection framework for bi-temporal remote sensing imagery that is built on a Visual State Space Model backbone and augments it with a Joint Spatio-Frequency Feature Fusion block, a Change-Guided Attention mechanism, and a Separated Kappa loss. Its target setting is semantic change detection rather than binary differencing alone: given a pre-change image $I^{T_1}$ and a post-change image $I^{T_2}$, the framework seeks to determine both where change occurred and what semantic transition occurred. The model is presented as a Siamese multi-task architecture that couples Binary Change Detection (BCD) and Semantic Change Detection (SCD), and the reported results on SECOND and Landsat-SCD are 88.62% Overall Accuracy, 65.78% $F_{\text{scd}}$, and 25.50% SeK on SECOND, and 96.25% Overall Accuracy, 89.27% $F_{\text{scd}}$, and 60.26% SeK on Landsat-SCD (Wijenayake et al., 11 Aug 2025).

1. Problem formulation and motivation

The framework addresses Semantic Change Detection (SCD) in high-resolution and medium-resolution remote sensing scenes. In this setting, Binary Change Detection localizes changed versus unchanged pixels, whereas Semantic Change Detection predicts semantic labels for both timestamps and thereby infers semantic transitions such as building $\rightarrow$ ground. The paper characterizes these two tasks as naturally intertwined: BCD localizes changed regions, while SCD requires accurate semantic understanding within those regions (Wijenayake et al., 11 Aug 2025).

The stated difficulty of SCD arises from several concurrent factors. Long-range spatial context is needed in high-resolution scenes; illumination, phenology, and sensor or view changes create spectral inconsistencies; and class imbalance is severe because unchanged pixels dominate while rare transitions occupy very few pixels. The paper also identifies an architectural limitation in prior systems: BCD and SCD decoders are often isolated, which prevents mutual guidance. In addition, prior Mamba-based SCD methods are described as purely spatial, leaving frequency cues underused (Wijenayake et al., 11 Aug 2025).

This problem framing explains the three central design choices of Mamba-FCS. The VMamba backbone is used for efficient long-range spatial modeling; a spatio-frequency fusion block is introduced to incorporate frequency-domain evidence; and BCD supervision is fed back into semantic decoding through Change-Guided Attention. A plausible implication is that the framework is designed to improve boundary fidelity and rare-transition recognition without abandoning the computational advantages associated with state-space models.

2. Architecture and backbone organization

Mamba-FCS is a Siamese multi-task framework with a shared VMamba encoder, one BCD decoder, and two SCD decoders, one for $T_1$ and one for $T_2$. The inputs are

$$I^{T_1}, I^{T_2} \in \mathbb{R}^{3 \times H \times W},$$

and the shared encoder produces four hierarchical feature stages for each timestamp: $$X_1^{T_1}, X_2^{T_1}, X_3^{T_1}, X_4^{T_1} = \mathcal{F}_{encoder}(I^{T_1}),$$

$$X_1^{T_2}, X_2^{T_2}, X_3^{T_2}, X_4^{T_2} = \mathcal{F}_{encoder}(I^{T_2}),$$

with

$I^{T_2}$0

These features are consumed by $I^{T_2}$1 and by the two timestamp-specific semantic decoders $I^{T_2}$2 and $I^{T_2}$3 (Wijenayake et al., 11 Aug 2025).

The encoder is based on VMamba. A patch partition layer is followed by four hierarchical stages with spatial resolutions $I^{T_2}$4, $I^{T_2}$5, $I^{T_2}$6, and $I^{T_2}$7. The configuration reported for VMamba-Base is

$I^{T_2}$8

and

$I^{T_2}$9

Each Visual State-Space block contains a 2D Selective Scan module that scans the feature map in four directions—top-left $F_{\text{scd}}$0 bottom-right, bottom-right $F_{\text{scd}}$1 top-left, top-right $F_{\text{scd}}$2 bottom-left, and bottom-left $F_{\text{scd}}$3 top-right—to preserve a global receptive field while retaining linear complexity (Wijenayake et al., 11 Aug 2025).

The decoder organization is top-down. In the BCD branch, each stage fuses $F_{\text{scd}}$4 and $F_{\text{scd}}$5, passes the result through a VSS block, and applies a CBAM-based upsampling unit. The semantic branches are architecture-identical but weight-independent, which allows each decoder to specialize to its timestamp. Their outputs are

$F_{\text{scd}}$6

where $F_{\text{scd}}$7 is the number of semantic classes (Wijenayake et al., 11 Aug 2025).

3. Joint spatio-frequency feature fusion

A defining component of the model is the Joint Spatio-Frequency Feature Fusion block, denoted $F_{\text{scd}}$8, which is applied at each decoder stage. It combines five sources of information: spatial features from $F_{\text{scd}}$9, spatial features from $F_{\text{scd}}$0, frequency-domain features from $F_{\text{scd}}$1, frequency-domain features from $F_{\text{scd}}$2, and explicit spatial difference features. The stated purpose is to sharpen edges, suppress illumination artifacts, and enhance subtle changes (Wijenayake et al., 11 Aug 2025).

For each stage $F_{\text{scd}}$3, frequency features are obtained by applying channel-wise 2D FFT followed by log-amplitude compression: $F_{\text{scd}}$4

$F_{\text{scd}}$5

The log-amplitude form $F_{\text{scd}}$6 is used to emphasize high-frequency components, compress dynamic range, and make the representation robust to subtle structural changes. The paper explicitly associates this branch with improved handling of edges, textures, minor boundaries, and changes under shadows or illumination shifts (Wijenayake et al., 11 Aug 2025).

The spatial differencing branch is

$F_{\text{scd}}$7

The full fusion input is then constructed as

$F_{\text{scd}}$8

A $F_{\text{scd}}$9 convolution reduces this concatenated tensor to

$\rightarrow$0

after which a CBAM module applies channel attention and spatial attention to yield the final fused representation $\rightarrow$1 (Wijenayake et al., 11 Aug 2025).

The ablation evidence assigns a particularly important role to the FFT2 branch. On SECOND, removing the FFT2 branch reduces performance from OA 88.62, $\rightarrow$2 65.78, mIoU 74.07, SeK 25.50 to OA 87.86, $\rightarrow$3 64.52, mIoU 73.16, SeK 24.03. The paper interprets this as evidence that frequency-domain cues improve edge precision and suppress hallucinated changes (Wijenayake et al., 11 Aug 2025).

4. Change-guided decoding and SeK-oriented optimization

The mechanism that connects the binary and semantic tasks is Change-Guided Attention (CGA). The BCD decoder generates intermediate change maps $\rightarrow$4, and these maps modulate the semantic features at each stage: $\rightarrow$5 where $\rightarrow$6 is sigmoid and $\rightarrow$7 is element-wise multiplication. At the coarsest stage, $\rightarrow$8 is passed through a VSS block and CBAM-based upsampling; at finer stages, $\rightarrow$9 is added to the upsampled features from the previous stage and processed again. In effect, the semantic decoders are constrained to focus on regions with high change probability (Wijenayake et al., 11 Aug 2025).

The reported ablation indicates that removing CGA lowers SECOND performance to OA 88.08, $T_1$0 63.61, mIoU 73.93, SeK 24.07. The largest drop occurs in $T_1$1, which the paper interprets as evidence that semantic prediction is less precise when BCD does not guide SCD. This suggests that decoder interaction is a central part of the method rather than an auxiliary refinement (Wijenayake et al., 11 Aug 2025).

The optimization objective includes a Separated Kappa term designed for class-imbalanced semantic change detection. SeK is defined as

$T_1$2

with

$T_1$3

$T_1$4

where $T_1$5 is the confusion matrix and class 1 is no-change. The metric is computed for both timestamps, averaged, clipped at zero, and converted into a loss: $T_1$6

$T_1$7

$T_1$8

The total loss is

$T_1$9

with $T_2$0 and $T_2$1 (Wijenayake et al., 11 Aug 2025).

5. Datasets, training protocol, and empirical performance

The framework is evaluated on two public datasets. SECOND contains 4,662 image pairs at 0.53 m/pixel, with image size $T_2$2, urban scenes from Hangzhou, Chengdu, and Shanghai, and six classes: non-vegetated surfaces, trees, low vegetation, water bodies, buildings, and playgrounds. The split is 2,968 train and 1,694 test. Landsat-SCD contains 2,425 image pairs at 30 m/pixel and size $T_2$3, from Tumushuke, Xinjiang, China, with four land-cover classes—farmland, desert, buildings, and water bodies—and ten semantic change types. Its split is 1,455 train, 485 validation, and 485 test (Wijenayake et al., 11 Aug 2025).

Implementation is in PyTorch. Optimization uses AdamW with learning rate $T_2$4, weight decay $T_2$5, and batch size 4. Training runs for 30,000 iterations on SECOND and 50,000 iterations on Landsat-SCD. The reported augmentations are random rotations, horizontal flips, vertical flips, random saturation, contrast adjustment, and brightness adjustment (Wijenayake et al., 11 Aug 2025).

The evaluation metrics are OA, mIoU, SeK, and $T_2$6. OA is explicitly noted as potentially misleading under class imbalance, which motivates the emphasis on mIoU and SeK (Wijenayake et al., 11 Aug 2025).

Dataset	Mamba-FCS results	Best competing method
SECOND	OA 88.62%, $T_2$7 65.78%, mIoU 74.07%, SeK 25.50%	ChangeMamba: 88.12 / 64.03 / 73.68 / 24.11
Landsat-SCD	OA 96.25%, $T_2$8 89.27%, mIoU 88.81%, SeK 60.26%	ChangeMamba: 96.08 / 86.61 / 86.91 / 53.66

On SECOND, the gains over ChangeMamba are +0.50 OA, +1.75 $T_2$9, +0.39 mIoU, and +1.39 SeK. On Landsat-SCD, the gains are especially large in $$I^{T_1}, I^{T_2} \in \mathbb{R}^{3 \times H \times W},$$0, mIoU, and SeK: +2.66, +1.90, and +6.60, respectively. The paper characterizes these as state-of-the-art results (Wijenayake et al., 11 Aug 2025).

The ablation study on SECOND attributes measurable gains to all four novel components evaluated in isolation. Removing SeK lowers results to OA 88.43, $$I^{T_1}, I^{T_2} \in \mathbb{R}^{3 \times H \times W},$$1 65.09, mIoU 73.67, SeK 24.71; removing the difference branch yields OA 88.34, $$I^{T_1}, I^{T_2} \in \mathbb{R}^{3 \times H \times W},$$2 65.11, mIoU 73.77, SeK 24.83. The paper summarizes the importance ordering as roughly FFT2 branch, CGA, SeK loss, and difference branch, while also emphasizing that all are beneficial (Wijenayake et al., 11 Aug 2025).

6. Position within the broader Mamba literature

Within the broader Mamba literature, Mamba-FCS is a dense prediction architecture for bi-temporal remote sensing, rather than a generic Mamba backbone or an autoregressive inference system. Its use of VMamba and 2D Selective Scan places it in the family of vision-oriented state-space models, but its distinctive contributions are frequency-aware fusion, explicit BCD-to-SCD guidance, and SeK-aligned optimization (Wijenayake et al., 11 Aug 2025).

This specialization distinguishes it from other Mamba applications. BMACE is a bidirectional Mamba-based network for automatic chord estimation that operates on 10-second audio excerpts converted to CQT features and uses two Mamba blocks with opposite temporal masking directions for sequence modeling (Yuan et al., 5 Jan 2026). HS-Mamba addresses hyperspectral image classification through a full-field interaction strategy that combines a dual-channel spatial-spectral encoder on non-overlapping patches with a lightweight global inline attention branch on the whole image (Peng et al., 22 Apr 2025). A separate white blood cell classification paper presents a Mamba-based ensemble framework, but explicitly notes that the acronym “Mamba-FCS” does not appear in that work (Clifton et al., 15 Apr 2025). SpecMamba, by contrast, is an FPGA accelerator for Mamba inference with speculative decoding and addresses hidden-state backtracking, tree verification, and hardware workload mismatch in autoregressive settings rather than image segmentation or change detection (Zhong et al., 24 Sep 2025).

A common misconception is that “Mamba-FCS” might denote any Mamba-based feature or classification system. The literature summarized here indicates a narrower usage: in the cited work, Mamba-FCS refers specifically to a semantic change detection framework for remote sensing that couples spatio-frequency fusion, change-guided semantic decoding, and SeK-based optimization (Wijenayake et al., 11 Aug 2025). This suggests that the importance of the method lies not merely in replacing attention with a state-space backbone, but in integrating that backbone with SCD-specific inductive structure: bi-temporal fusion, decoder interaction, and a loss aligned to changed-region semantic agreement.