acceptodds
Under review as a conference paper at ICLR 2027

Not All Pixels Flow Alike: Spatial-Clock Flow Matching for Remote Sensing Change Detection

Abstract

Remote sensing change detection (RSCD) aims to identify genuine scene changes from bitemporal imagery, yet visual discrepancies caused by illumination, blur, noise, and residual misregistration are often entangled with true changes. Existing methods mainly learn discriminative mappings from bitemporal observations to change masks. However, implicitly coupling nuisance suppression and change structure recovery makes it difficult to distinguish appearance variations from genuine changes, especially under weak changes and severe nuisance interference. To explicitly model this correction process, we formulate RSCD as continuous evidence-to-geometry transport using Flow Matching (FM). Nevertheless, applying FM to RSCD requires explicit modeling of nuisance-consistent target coupling and geometry-guided spatial conditioning. We therefore propose TempoFlow-CD, which reformulates FM through target coupling and temporal parameterization. Nuisance-Invariant Geometry Flow (NIGF) aligns nuisance-varied observations with shared mask–distance geometry endpoints, while Change-Adaptive Spatial Clock (CASC) enables content-aware transport progression through adaptive local phases and velocity refinement. Extensive experiments on six RSCD benchmarks demonstrate that TempoFlow-CD achieves superior F1 and IoU performance with improved boundary localization and robustness to radiometric perturbations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Flow Matching, Remote Sensing Change Detection, Spatially Asynchronous Dynamics, Nuisance Invariance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_RiuQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper formulates RSCD as continuous evidence-to-geometry transport using Flow Matching (FM) and proposes TempoFlow-CD, which reformulates FM through target coupling and temporal parameterization. NIGF couples nuisance-varied sources through a shared mask–distance target and endpoint consistency, while CASC combines geometrysupervised local phase conditioning with boundary-guided velocity refinement on a shared integration grid. Experiments are conducted on six RSCD benchmarks demonstrate improved detection accuracy and boundary localization, with additional evaluations supporting robustness to radiometric disturbances.

Critical strengths

(1) The paper is easy to follow and the motivation of this paper is clear. It claims that visual discrepancies caused by illumination, blur, noise, and residual misregistration are often entangled with true changes.

(2) The experiment is conducted on many different datasets show the good performance of the proposed methods.

Critical weaknesses

(1) Although the motivation of the proposed method is clear, the author did not analyze the visual discrepancies caused by illumination, blur, noise, and residual misregistration are existed in the different datasets. For example, in the LEVIR-CD dataset, how many visual discrepancies are existed in the true scenario. It is better to describe different scenarios that visual discrepancies effect the true changes.

(2) The figure 2 is confused to describe the proposed TempoFlow-CD. In this figure, the inputs pass through VRWKV Encoder, NIGF, and CASC. However, the output is not clear in this figure. The motivation of the NIGF and CASC is not clear in the figure.

(3) For the overall objective function, the author did not analyze the hyperparameters of λ in this equation. Also, the ablation study does not discuss this in the experiment. Without this, we can not determine which term is more important in the current design.

(4) More visualization results should focus on different visual discrepancies caused by illumination, blur, noise, and residual misregistration in the experiment part. Only figure 4 demonstrates several changes in different datasets. More detailed analysis and discussion should be added to support the motivation.

(5) More state-of-the-art change detection methods should be evaluated and discussed in the experiment. Actually, it is better to classify the compared method with different types, such as flow matching-based methods, hybrid-based methods, et al.

Questions for authors

Please see the weakness.

Reviewer_shByOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TempoFlow-CD for remote sensing change detection (RSCD), which formulates change detection as a continuous transport process from bitemporal image evidence to structured change information. A central challenge in RSCD is distinguishing nuisance-induced appearance variations from genuine changes while recovering accurate change geometry. The paper further identifies two limitations in applying Flow Matching: the lack of cross-view consistency and geometry-dependent spatial phase conditioning. To tackle these problems, the proposed framework aims to progressively transform image-derived features into geometry-aware representations through continuous flow dynamics, while explicitly incorporating nuisance consistency and spatially adaptive conditioning. The proposed framework consists of two main components. First, the authors propose Nuisance-Invariant Geometry Flow (NIGF), which aligns nuisance-perturbed observations with a shared mask–distance geometry target and enforces cross-view endpoint consistency. They further introduce Change-Adaptive Spatial Clock (CASC), which incorporates geometry-supervised local phase conditioning and boundary-guided velocity refinement to enable spatially adaptive transport.

Critical strengths

The proposed method achieves strong performance across multiple benchmark datasets. The subsequent experimental analyses and ablation studies demonstrate the effectiveness of the proposed modules.

Critical weaknesses
  1. Although the formulation is mathematically consistent with Flow Matching, the proposed transport process is constructed between paired image-derived features and ground-truth-derived geometry embeddings, with intermediate states obtained through linear interpolation. Consequently, the method appears closely related to time-conditioned supervised geometry regression. While the paper provides relevant ablations, it remains insufficiently demonstrated what distinctive benefits the continuous transport formulation offers beyond direct regression or iterative feature refinement.
  2. CASC supervises spatial phase offsets using a handcrafted target derived from normalized interior distances. However, the paper does not sufficiently justify this predefined relationship between geometry and transport phase. Although the ablation results show that offset supervision improves performance, they do not establish why this particular rule is preferable to alternative or fully learned phase assignments, leaving the specific phase assignment rule insufficiently justified.
Questions for authors
  1. Could the authors clarify the specific benefits of Flow Matching over direct regression or iterative refinement, and provide evidence that the intermediate transport states represent meaningful progressive geometry recovery?
  2. Why is the spatial clock supervised by a handcrafted distance-based target rather than learned fully adaptively? What justifies the specific phase assignments, and what advantages do they offer over fully learned offsets?
  3. Figure 6 mainly provides a schematic illustration of the transport process without explicitly showing the geometry represented by intermediate states. Could the authors visualize intermediate predictions along actual ODE trajectories to clarify whether meaningful change geometry progressively emerges during inference?
Reviewer_Srv8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TempoFlow-CD, a flow-matching framework for remote sensing change detection that models the continuous transformation from bitemporal image features to change geometry representations. The proposed NIGF aligns appearance-perturbed observations with shared mask–distance geometry targets, while CASC introduces spatially adaptive phase conditioning and boundary-guided velocity refinement.

Critical strengths
  1. The paper presents a well-structured framework that integrates nuisance-consistent learning and geometry-aware spatial conditioning into flow matching for remote sensing change detection.
  2. The experiments cover multiple change detection benchmarks, with extensive ablation studies demonstrating the effectiveness of the proposed components and improvements in detection accuracy and boundary localization.
Critical weaknesses
  1. ChangeFlow has already formulated remote sensing change detection as a latent rectified-flow problem, and its semantic change detection extension further learns transport from image-derived latents to label latents. The conceptual distinction between this method and these existing flow-based approaches is not sufficiently highlighted.
  2. NIGF combines appearance perturbations, shared geometry targets, and cross-view endpoint consistency, which are closely related to conventional data augmentation and consistency learning. Similar nuisance-invariant predictions could also be learned by applying these techniques to existing flow-based change detection methods, such as ChangeFlow.
  3. Figure 2 is difficult to follow due to its dense layout and numerous interconnected components.
Questions for authors
  1. What are the key methodological differences between TempoFlow-CD and existing flow-based change detection methods such as ChangeFlow?
  2. What additional advantages does NIGF provide beyond conventional data augmentation and consistency learning?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.