acceptodds
Under review as a conference paper at ICLR 2027

FM-ChangeNet++: Learned Nonlinear Feature Transport for Change Detection

Abstract

Most remote sensing change detection methods infer change by directly comparing features extracted from bi-temporal images. Recent flow-matching approaches reinterpret this problem as learning feature transport between pre- and post-temporal representations through a fixed linear interpolation path. While effective, this assumption constrains semantic transitions to follow straight trajectories in feature space, limiting the representation of complex structural changes. We introduce FM-ChangeNet++, a feature transport framework that learns the transformation trajectory itself. Instead of enforcing a fixed linear path, we model feature evolution using an endpoint-preserving residual transport formulation, where a lightweight residual predictor learns smooth nonlinear deviations while guaranteeing that the trajectory starts from the pre-temporal representation and terminates at the post-temporal representation. The velocity field is supervised using the instantaneous derivative of the learned nonlinear path, ensuring that the predicted velocity represents the local tangent direction and transport intensity along the feature trajectory. Its magnitude therefore provides a localized measure of feature evolution that can support semantic change detection. The proposed framework jointly optimizes multi-scale flow estimation, residual path regularization, spatial smoothness, and segmentation supervision within a unified coarse-to-fine architecture. This formulation enables the model to capture complex feature evolution while maintaining endpoint consistency and controlled transport dynamics. Extensive experiments on standard remote sensing benchmarks demonstrate consistent improvements over linear transport formulations and state-of-the-art CD methods. Ablation studies further show that learned residual transport produces more discriminative feature trajectories, more localized velocity representations, and more accurate detection of subtle semantic changes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Earth observation, remote sensing, change detection, flow matching, nonlinear feature transport, representation learning, feature-space transport

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nMjuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces FM-ChangeNet++ for bi-temporal remote sensing change detection. It extends FM-ChangeNet’s linear feature interpolation with a learned nonlinear residual path that preserves both endpoints. Extensive experiments are conducted over LEVIR-CD, WHU-CD, and DSIFN-CD that demonstrate their effectiveness.

Critical strengths
  1. The endpoint-preserving construction and corresponding velocity target are mathematically consistent.

  2. The paper includes component ablations, path and time-sampling studies, cross-domain evaluation, and multiple random seeds.

Critical weaknesses
  1. The novelty is incremental and limited. The main contribution adds a residual path and corresponding supervision to an existing framework; endpoint preservation and derivative computation are relatively straightforward.

  2. Ambiguous ablation protocols. The paper does not adequately explain how the “flow-only” setting trains its segmentation head or fully control for the additional residual module’s capacity.

  3. Limited theoretical support for semantic interpretation. The analysis establishes properties of the constructed path but does not demonstrate that nonlinear velocity reliably separates semantic changes from nuisance variations.

  4. Several implementation details are missing (Dice coefficient lambda, smoothness parameter epsilon, etc), some citations are mismatched (RS-Mamba or RSmamba ? there are different papers.).

Questions for authors
  1. Can you provide some failure cases?

  2. Is it reliable for those spatial mismatch pairs?

Reviewer_YtohOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FM-ChangeNet++, which extends linear flow matching with an endpoint-preserving residual nonlinear feature transport path for remote sensing change detection. The velocity field is supervised using the instantaneous derivative of the learned trajectory. The paper provides theoretical analysis and conducts experiments and ablation studies on LEVIR-CD, WHU-CD, and DSIFN-CD to demonstrate the effectiveness of the proposed method.

Critical strengths

1.Solid theoretical analysis. The paper provides systematic mathematical derivations that offer theoretical support for the proposed nonlinear feature transport framework. 2. Comprehensive experimental evaluation. The method is evaluated on three public benchmarks, demonstrating its effectiveness across different datasets.

Critical weaknesses
  1. Limited methodological novelty. The core contribution mainly involves introducing learnable residual nonlinear interpolation into an existing flow-matching framework, making it more of an incremental extension than a fundamentally new approach. The improvements over FM-ChangeNet are relatively modest and do not sufficiently demonstrate the advantages of the proposed method. 2.Insufficient motivation and justification. Despite extensive mathematical formulations and theoretical derivations, the paper does not adequately explain why change detection requires explicit continuous feature transport or what unique advantages flow matching offers over direct feature comparison and other approaches. In particular, since the constructed nonlinear trajectories do not correspond to the actual temporal evolution of the scene, their semantic interpretation remains unclear. Additionally, the manuscript contains repetitive descriptions of the methodology and theoretical properties. Some figures, particularly Figure 3, are overly dense with small text, which negatively affects readability.
Questions for authors
  1. What are the key advantages of applying flow matching to change detection? Compared with conventional feature differencing, direct feature regression, or other representation learning methods, what are its unique benefits and potential limitations? Can the authors provide further experimental evidence to clarify these differences? 2.Since the learned nonlinear trajectories do not represent the actual temporal evolution of the scene, how can the authors verify that their geometric structure and instantaneous velocity magnitudes genuinely reflect semantic changes? Can more direct evidence or stronger motivation be provided to justify the necessity of learning nonlinear feature trajectories?
Reviewer_7cKdOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes FM-ChangeNet++, a change detection model based on flow matching. Where previous approaches used a linear interpolation path, FM-ChangeNet++ adds a non-linear residual connection. The paper claims an improved detection of changes, justified by the fact that such a non linear approach could model more complex transitions or appearance variations.

Critical strengths
  • The motivation behind modeling temporal evolution as a transport problem and detecting change based on the magnitude of the learned flow sound interesting and novel
  • The design of the trajectory, starting from existing linear methods and ensuring that the endpoints are preserved is coherent. The approach of flow matching, deterministic at inference time vs generative diffusion based methods is also interesting. Overall the methodology is really super interesting!
  • The paper presents extensive ablation studies
Critical weaknesses
  • Incremental novelty: The paper seems to be based on the implementation of FM-ChangeNet (Kazoom et al., Arxiv, 2026) appearing in the results (table1) and mentioned in the introduction but not the related works. The only difference between both models seems to be the non-linear residual correction. While really interesting, this contribution seems really modest.
  • Following the above comment, not only does the contribution seem modest from a methodological point of view, but the results are also really not convincing. In Table 1, for two out of the three datasets the difference between FM-ChangeNet (=linear) and FM-ChangeNet++ (=non linear) are really small. Even more, when looking at the confidence intervals in Table 13 of the supplementary materials, the difference between the two models lies in the confidence intervals, i.e. the difference is not significant.
  • In the same vein, looking at table 12 in the supplementary materials, the dataset-averaged normalized endpoint displacement and curvature proxy are both really small, indicating, according to the caption, that the encoder behaves approximately linearly. Does this mean that the learned non linear path is in fact almost linear? Indicating that letting the model be free to behave non-linearly does not help? If yes, then again the experimental results do not favor a non-linear implementation.
  • In Table 2, the results with fixed velocity target are nearly identical to those with instantaneous derivative. It therefore seems that the type of supervision does not matter much, which is in line with the fact that the non linear model behaves almost linearly, the instantaneous derivative becoming in that case the fixed velocity one.
  • There is no study of residual magnitude. Is it significantly greater than 0?

Additional Non-Methodological Comment: It is to be noted that this work, beside being close to FM-ChangeNet in terms of methodology (the difference being the residual term), the two papers are also similar, both in their general appearance (figure 1 and 2 are very similar, captions nearly identical) and in terms of experiments. Table 1 of the FM-ChangeNet++ paper is the same as in FM-ChangeNet, with one additional line for FM-ChangeNet++, and the second table of FM-ChangeNet can be found in the appendices of FM-ChangeNet++, supporting the arguments regarding the incremental nature of this paper.

Questions for authors
  • How is the instantaneous derivative computed from the data? Is it via autograd? Finite differences?
  • Line 275. Why using different descriptors, i.e. different combinations of the features of both images. Does it have an impact on the performance?
  • Line 211. What is C?
  • In the abstract (line 046), it is stated that the magnitude of the velocity field provides a measure of the evolution that support semantic change detection. Semantic Change Detection is defined as the task of binary CD on which we add a segmentation of the two input images, providing information on the semantics, i.e. the type of changes. How does the velocity relates to that?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary and Request for Score Reconsideration

We sincerely thank all reviewers for their thoughtful feedback and constructive suggestions, which have helped us strengthen our work.

Throughout the rebuttal, we have addressed the main concerns regarding methodological novelty, statistical significance, nonlinear representation learning, ablation protocols, and practical robustness.

Importantly, we clarified several experiments already presented in the Appendix and provided additional evidence demonstrating statistically significant improvements, measurable nonlinear transport, and substantially stronger robustness to spatial misregistration across all three benchmarks. We also clarified implementation details and committed to improving the manuscript's presentation.

As the rebuttal period approaches its conclusion, we would like to kindly follow up, as we have not yet received further questions from all reviewers. We believe our clarifications and additional evidence substantially address the concerns raised, and we would sincerely appreciate hearing from you if any important issues remain.

Otherwise, we would be truly grateful if you could reconsider your initial assessments and consider upgrading your scores in light of the additional evidence and clarifications.

Thank you once again for your valuable time, consideration, and engagement.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.