Temporally Ordered Region-Token Mamba with Logit-Space Diffusion for Remote Sensing Change Detection

1. Introduction

• BMD-CD combines structured temporal interaction, feature disentanglement, and targeted prediction refinement in one framework.The paper evaluates it on four standard benchmarks, S2Looking, and zero-shot transfer to ValaisCD and B-FLAIR-test.Source: §1 · p. 2
• Bitemporal Ordered Mamba organizes region tokens into temporal partitions and uses bidirectional propagation for linear-complexity cross-temporal interaction.All t1 tokens precede all t2 tokens.Source: §1 · p. 2
• Orthogonal Feature Disentanglement forms shared- and change-oriented features, while logit-space diffusion refines ambiguous boundaries.Unchanged-region consistency is used to suppress pseudo-change responses.Source: §1 · p. 2

2.1. Robust Bitemporal Representation

• DDPM-CD aggregates multilevel features across diffusion timesteps, whereas Orthogonal Feature Disentanglement explicitly separates shared- and change-oriented outputs.The cited description notes that DDPM-CD’s representations are robust to environmental perturbations but do not explicitly organize change-sensitive and change-invariant information into separate branches.Source: §2.1 · p. 2

2.2. Boundary Preservation and Prediction Refinement

• RemoteVAR uses coarse-to-fine autoregressive prediction, while BMD-CD applies a small number of DDIM steps directly to coarse change logits for boundary refinement.RemoteVAR’s diffusion steps are described as feature extraction; a lightweight classification head produces its final change maps.Source: §2.2 · p. 2

2.3. Efficient Global and Cross-Temporal Modeling

• Global self-attention has quadratic token cost, while selective state-space models provide linear sequence complexity for long-range propagation.BMD-CD explicitly partitions region tokens by time before bidirectional propagation, making temporal sequence organization explicit while retaining linear complexity.Source: §2.3 · pp. 2–3

3.1. Overall Architecture

• The four-stage architecture combines shared Swin-T feature extraction, LightMerge for shallow levels, BOMO–OFD for deep levels, hierarchical decoding, and five-index DDIM refinement.The top-down decoder produces coarse logits h and diffusion condition features C before the conditional decoder refines h in logit space.Source: §3.1 · p. 3

3.2. Multi-hierarchy Patch Encoding and Shallow Interaction

• Shallow levels l ∈ {1, 2} retain local detail, while deeper levels l ∈ {3, 4} enter the BOMO–OFD branch.The backbone features are described as having progressively larger receptive fields and coarser resolutions at deeper levels.Source: §3.2 · p. 3

3.2.1. LightMerge for Shallow Feature Interaction

• LightMerge combines absolute-difference and element-wise co-activation cues with convolutional blocks and channel reweighting.A mean–max spatial attention map suppresses background responses; the resulting features retain local and boundary information for the top-down decoder.Source: §3.2.1 · p. 3

3.3. BOMO–OFD Branch for Deep Feature Interaction

• The branch applies bidirectional state-space propagation to the ordered sequence and reconstructs spatial features before temporal refinement and OFD.Its region-token design reduces the state count from 2HlWl spatial states to 128 region states.Source: §3.3 · p. 3, §3.3.1 · p. 3

3.3.1. Ordered Region Tokens and State Propagation

• BOMO pools each deep feature map to 64 region tokens, orders all first-image tokens before second-image tokens, and scans the compact 128-token sequence bidirectionally.Corresponding token indices represent the same spatial regions; the ordered arrangement lets forward states propagate context from the first observation into the second.Source: §3.3.1 · p. 3

3.3.2. Spatial Reconstruction and Temporal Refinement

• The reconstructed BOMO context is combined with projected feature differences, with a spatial- and channel-adaptive gate balancing contextual and localized signals.The context correction starts from the direct difference because γl is initialized to zero.Source: §3.3.2 · p. 4

3.3.3. Orthogonal Feature Disentanglement

• OFD uses learned pairwise orthogonal rotations to form change-oriented and complementary outputs, passing only the change-oriented output to the decoder.Training encourages diverse rotations and applies unchanged-region consistency to shallow bitemporal features; the loss promotes diverse rotations rather than directly enforcing zero correlation between outputs.Source: §3.3.3 · p. 4

3.4. Hierarchical Fusion and Coarse Logit Projection

• The decoder receives LightMerge outputs P1 and P2 and change-oriented OFD outputs P3 and P4, projected to 128 channels.Sample-dependent scale weights are computed from global descriptors of the resolution-aligned features.Source: §3.4 · p. 4
• Fusion proceeds from the deepest level upward, concatenating each weighted lateral feature with the upsampled decoder state.The scale weights are obtained by applying a softmax to an MLP over the global descriptors.Source: §3.4 · pp. 4–5
• The output projections produce full-resolution coarse logits h and an eight-channel condition map C for diffusion.The decoder head produces a 64-channel full-resolution feature before the projections.Source: §3.4 · p. 5

3.5. Conditional Diffusion Decoder in Logit Space

• The decoder predicts noise conditioned on the corrupted clean-logit target, timestep, coarse logits, and multiscale condition, then fuses the recovered estimate with coarse logits through a learned residual gate.The coarse logits provide the initial prediction, while the condition map supplies hierarchical decoder information.Source: §3.5 · p. 5
• Training applies BCE and Dice losses to the final refined logits and adds a noise-regression loss for the diffusion branch.This makes segmentation gradients supervise the final refinement path as well as the coarse prediction.Source: §3.5 · p. 5
• Inference uses five scheduled denoiser evaluations at indices {50, 38, 25, 12, 0}, with the deterministic DDIM trajectory initialized from the coarse logits.The terminal t = 0 evaluation is output-neutral; reported FLOPs and latency include it, making the reported cost conservative relative to a four-call implementation.Source: §3.5 · p. 5

4.1. Datasets

• BMD-CD is evaluated on LEVIR-CD, DSIFN-CD, WHU-CD, CDD, and S2Looking, with zero-shot transfer from S2Looking to ValaisCD and B-FLAIR-test.Transfer evaluation uses a threshold selected on S2Looking validation and fixed for target datasets.Source: §4.1 · p. 5, Table 3 · p. 6, §4.3 · p. 7

4.2. Implementation Details

• Experiments use a Swin-T encoder pretrained on ImageNet-1K, train for 125 epochs with batch size 8, and select K = 64 on validation.All experiments use one NVIDIA RTX 4090 GPU; diffusion uses 100 forward steps and a warm-started five-index DDIM inference schedule.Source: §4.2 · p. 5

4.3. Comparison with State-of-the-Art

• BMD-CD improves LEVIR-CD F1 by 1.4 over PeftCD, while WHU-CD and CDD results are within 0.005 F1 of the strongest baselines; boundary-F1 improvements are consistent.The comparison also reports competitive or best performance among entries with comparable results, with the clearest region-level gains on LEVIR-CD and DSIFN-CD.Source: Table 1 · p. 6, Table 2 · p. 6, §4.3 · p. 7

4.4. Ablation and Analysis

• Component ablations find diffusion gives the largest single-component boundary gain, including +8.5 B-F1 on WHU-CD, while BOMO improves recall and region-level F1.The full combination of BOMO, OFD, and logit-space refinement achieves the best overall result among evaluated variants.Source: §4.4 · pp. 7–8, Table 4 · p. 7

4.5. Ordering & Diffusion Analysis

• Grouped [a; b] ordering, bidirectional scanning, K = 64, and five-index DDIM perform best in the sensitivity analysis; interleaving loses 1.27 F1 against grouped ordering.The five-index schedule saturates the refinement gain; the grouped ordering processes all t1 tokens before all t2 tokens.Source: Table 5 · p. 8, §4.5 · p. 8

4.6. Efficiency Analysis

• The full model requires 32.09 GFLOPs and 47 ms per 256 × 256 pair, versus 13.20 GFLOPs and 22 ms without diffusion; LEVIR-CD F1 rises from 0.915 to 0.937.The added cost is associated with five-index DDIM refinement and primarily benefits boundary quality; the trade-off varies by dataset.Source: Table 6 · p. 8, §4.6 · p. 8

• BMD-CD combines temporally ordered region-token state-space modeling with logit-space diffusion refinement for remote-sensing change detection.Source: §5 · p. 8
• Experiments across standard benchmarks, S2Looking, and cross-dataset evaluations demonstrate competitive accuracy and a favorable accuracy–efficiency trade-off.Source: §5 · p. 8
• Ablation results confirm complementary contributions from temporal ordering, orthogonal feature disentanglement, and logit-space refinement.Source: §5 · p. 8