acceptodds
Under review as a conference paper at ICLR 2027

Semantic-Guided Fuzzy Difference Refinement For Remote Sensing Semantic Change Detection

Abstract

Semantic change detection of bi-temporal remote sensing images plays an important role in urban development monitoring, environmental assessment, and land management. Recent deep learning methods have substantially improved change localization and semantic recognition. However, existing methods still struggle to preserve weak temporal evidence and make reliable decisions when real semantic changes exhibit low visual contrast. To address these challenges, we propose the Semantic-Guided Fuzzy Difference Refinement Network, termed SFDRNet, which is organized around two coordinated mechanisms: Temporal Semantic Calibration (TSC) and Uncertainty-Aware Semantic Difference Refinement (USDR). TSC constructs reliable temporal evidence by jointly modeling spatial structure and frequency statistics, and then calibrates the semantic representations of the two dates. USDR injects the calibrated semantic difference into the change representation and applies learnable fuzzy reasoning with complementary local and global evidence to refine ambiguous responses before change classification. In this way, SFDRNet first strengthens weak change evidence and then resolves the uncertainty that remains before prediction. Extensive experiments on two publicly available datasets demonstrate the effectiveness of SFDRNet. Further ablation, contrast-stratified, and uncertainty analyses validate the effectiveness of the two mechanisms in recognizing low contrast changes and refining ambiguous predictions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Semantic Change Detection, Remote Sensing, Low Contrast Change, Fuzzy Reasoning, Uncertainty Refinement

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_eVcGOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed the Semantic-Guided Fuzzy Difference Refinement Network (SFDRNet) to addressed the problem that existing methods struggle to preserve weak temporal evidence and make reliable decisions when real semantic changes exhibit low visual contrast. TSC constructs and calibrates temporal and semantic evidence, while USDR refines ambiguous change responses before prediction.

Critical strengths
  1. The organization of this paper is easy to follow.
  2. The related work summarize many SCD-based methods that are very related to the proposed method.
  3. The supplement shows several visualization results that can support the proposed method.
Critical weaknesses
  1. The motivation of the proposed method is not very clear. The author claimed that existing methods still struggle to preserve weak temporal evidence and make reliable decisions when real semantic changes exhibit low visual contrast. I think this issue is very common in the current SCD task. Many state-of-the-art methods are designed to address this issue in different perspectives such as capturing the tiny change from different temporal images, using some Anchor-based mechanism to capture the true change. The author should discuss the main difference between the proposed method and other different type methods with experiment and theory.

  2. The figure 1 is hard to understand. What is the main difference between the two times images? I do not think it exists small visual change between this two images. In contrast, I think the change between this two images are very obvious. It is better to use a real image-pair to explain the "small visual change". In addition, It seems that the ordinary variation and real weak change are in different distributions. However, why the weak evidence is easily suppressed? I think the motivation in this figure is not very clear to describe the main issue.

  3. The proposed architecture is not very novel compared to existed state-of-the-art methods. In figure 2, it seems that the TSC fuses the high and low level features by FFT and uses semantic decoder to calibrate semantic feature, which is very common in many deep learning based architectures. I think the main contribution of the proposed method is in USDR module. However, it hard to understand in this figure. It seems that the USDR just splits semantic feature channels with global and local branches then using Gaussian Dual- Response to fuse the feature, which is also not novel compared with other methods. It should discuss the different fusing mechanisms to show the main difference of the proposed method.

  4. More state-of-the-art SCD methods should be evaluated and compared in the experiment part, especially some hybrid-based architectures that can capture the local and global features. In addition, more architectures with different types should also be compared and evaluated in the experiment.

  5. It seems that the proposed method is not significantly better than the SCGNet. And it is hard to see the main differences in visualization results between the proposed method with SCGNet in figure 6.

  6. More visualization results or evidences should be illustrated the motivation of the proposed method. It is better to show many "small visual change" examples in the visualization results and give some heat maps or attention maps to show the interpretability of the proposed method.

Questions for authors

please see the weakness.

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

This paper proposes SFDRNet for semantic change detection in bi-temporal remote sensing imagery, with a particular focus on low-contrast semantic changes. The method contains two main components. First, Temporal Semantic Calibration (TSC) combines multi-scale spatial processing and Fourier-amplitude statistics to construct temporal evidence, while using a change-aware gate to control information sharing between the two semantic streams. Second, Uncertainty-Aware Semantic Difference Refinement (USDR) injects the calibrated semantic difference into the change representation and applies local and global Gaussian membership modeling, followed by complementary response aggregation, to generate a residual correction before binary change prediction.

Experiments are conducted on SECOND and Landsat-SCD. SFDRNet reports competitive performance, achieving the highest FscdF_{\mathrm{scd}}, SeK, and mIoU on SECOND and the highest reported results across all four metrics on Landsat-SCD. The paper additionally provides module ablations, a parameter-matched convolutional control, contrast-stratified evaluation, calibration measurements, and an analysis of pixels close to the binary decision boundary.

Critical strengths

(i) The paper identifies a relevant failure mode and organizes the method around a coherent hypothesis. The sequence of strengthening weak temporal evidence with TSC and then refining the remaining ambiguous responses with USDR is intuitively reasonable for semantic change detection, where a missed binary change also removes the corresponding semantic transition.

(ii) This paper provides more mechanism-oriented analysis than a standard aggregate benchmark comparison. In particular, the parameter-matched convolutional control, contrast-stratified evaluation, fixed uncertain subsets defined by an independently trained reference model, calibration metrics, and confident-complement analysis are useful attempts to determine where the improvements arise. The paper is also appropriately cautious in describing the contrast–uncertainty relationship as an association rather than causality.

Critical weaknesses

(i) The empirical advantage is too small to be convincing without repeated runs or statistical uncertainty. On SECOND, the improvement over the strongest competing result is only 0.490.49 points in FscdF_{\mathrm{scd}}, 0.520.52 in SeK, and 0.030.03 in mIoU. On Landsat-SCD, the corresponding advantage over SCGNet is only 0.120.12, 0.350.35, 0.110.11, and 0.040.04 points. The full fuzzy design improves over the parameter-matched convolutional control by only 0.400.40 FscdF_{\mathrm{scd}}, 0.250.25 SeK, 0.090.09 mIoU, and 0.180.18 OA on SECOND. Since no multi-seed mean, standard deviation, or significance analysis is reported, it is unclear whether these differences exceed ordinary training variance. This directly affects the main empirical claim that fuzzy refinement is preferable to a conventional residual module.

(ii) The central mechanisms are insufficiently isolated and the “uncertainty-aware fuzzy reasoning” interpretation is not yet fully justified. TSC bundles multi-scale spatial mapping, FFT-amplitude gating, temporal mixing, and semantic calibration, but is only ablated as one complete block. Therefore, the experiments cannot establish which part is responsible for the low-contrast improvements or whether the frequency component is necessary. Moreover, USDR does not explicitly receive entropy, confidence, or distance to the decision boundary; it is applied to all features. The uncertain-pixel analysis is post hoc, and a generally improved classifier may naturally make most corrections near another model’s decision boundary. The paper does not sufficiently explain why this operation constitutes fuzzy AND/OR rule inference, rather than an RBF-like learned nonlinear feature transformation.

Questions for authors

(i) Could the authors provide a component-level ablation of TSC, separating multi-scale spatial processing, frequency-amplitude gating, temporal mixing, and semantic calibration? This would materially affect whether the claimed low-contrast improvement can be attributed to the proposed spatial-frequency construction rather than to the overall increase in processing capacity.

(ii) Could the authors provide a more formal interpretation or additional control showing that Equation (5) implements meaningful fuzzy inference? For example, how does it compare with an RBF/mixture-of-experts transformation that uses the same Gaussian bases and number of parameters but does not use the proposed positive/complementary terminology?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.