acceptodds
Under review as a conference paper at ICLR 2027

Uncertainty-Guided Causal Disentanglement Network for Remote Sensing Change Detection

Abstract

Existing remote sensing change detection (RSCD) methods typically rely on intertwined feature representation learning, jointly encoding semantic backgrounds, real changes, and environmental perturbations. This entangled representation couples change indicators with invariant background semantics, rendering models highly susceptible to spurious correlations induced by variations in illumination, seasonal phenology, and sensor discrepancies. Furthermore, existing paradigms lack the capacity to model pixel-level prediction reliability. To address these limitations, this paper proposes an Uncertainty-Guided Causal Disentangled Network for RSCD. We first formalize the RSCD process via a Structural Causal Model (SCM), establishing a causal generative graph governing invariant backgrounds, authentic semantic changes, and confounding noise, and theoretically demonstrating that explicit feature disentanglement effectively blocks the backdoor path through which confounders bias change inference. Building upon this formulation, we design a Pearson correlation-guided feature disentanglement that distills shared invariant representations via cross-temporal correlations and explicitly segregates change components from background factors. To enhance the discriminative capacity of the disentangled representations, we further introduce a maximum-entropy clustering-based disentangled prototypical contrastive learning strategy. Unlike conventional prototype learning that forces all features to rigidly converge toward a single prototype, this strategy guides initial prototypes toward core prototypes, thereby better preserving the multimodal topology and diversity of real-world change manifolds. Finally, to bolster feature confidence, a progressive evidence-accumulating uncertainty estimation module is incorporated. By parameterizing a Dirichlet distribution over the disentangled representations, this module not only quantifies predictive uncertainty but also optimizes epistemic reliability via an evidence-based loss. Extensive experiments demonstrate that the proposed framework achieves substantial improvements across both qualitative and quantitative benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Remote sensing change detection, Causal disentanglement, Disentangled prototypical contrastive learning, dirichlet distribution

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_34sAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper solves the problem in distinguishing real land-cover changes from variations in remote sensing images. UCDNet uses cross-temporal co-relation to estimate shared features and subtracts these from the temporal representations to obtain the change features. A contrastive objective encourages discrimination between change and no-change, while a prediction head models class probabilities through a Dirichlet distribution and produces per-pixel uncertainty estimates. The authors motivate the feature decomposition through a structural causal model. Experiments cover LEVIR-CD, SYSU-CD, SECOND, and CDD-CD, with additional module ablations, cross-domain evaluation, complexity analysis, and feature-separability measurements.

Critical strengths

According to me the critical strengths are listed below :

  1. Coherent Design Framework: Shared-feature suppression, prototype-based discrimination, and evidential prediction address some of the important aspects of change detection. The uncertainty maps shown in Figure 6 appear to highlight ambiguous regions and boundaries. Such outputs could be useful for identifying predictions that require further inspection, provided their relationship to prediction errors is quantitatively validated. Thus I feel it is interpretable.

  2. The evaluation covers several complementary aspects of the method. Experiments span four datasets and eight comparison methods, alongside component ablations, cross-domain testing, complexity measurements, and feature-separability analysis. Reporting transfer limitations and computational cost is helpful because it gives a broader picture than in-domain accuracy alone.

Critical weaknesses
  1. The causal interpretation is not established by the feature-subtraction argument. Equation 4 marginalizes over the proposed confounder, whereas the implemented method subtracts an estimated shared feature and predicts from the residual. These operations are not equivalent in general. The derivation should state which assumptions connect the learned shared feature to the proposed confounder and justify dropping the dependence on ΦI\Phi_I. For example, a conditional-independence or sufficiency condition would need to be stated and justified; it does not follow from subtraction alone. The first-order approximation and the cancellation of cross-terms also need a clearer validity argument, including the distribution underlying the expectations and any centering assumptions. The 0.80-point F1 drop when the disentanglement module is removed shows that the module is useful for prediction, but it does not distinguish causal deconfounding from ordinary nuisance suppression or feature regularization. A controlled intervention experiment and an ablation using simple feature subtraction would provide more direct evidence.

  2. The experiments do not establish the reliability of the gains or isolate the claimed robustness mechanism. The improvements over the strongest baselines are 0.25 percentage points on LEVIR, 0.57 on SECOND, and 0.48 on CDD. Without repeated runs across seeds or uncertainty estimates, it is difficult to judge whether these differences are reproducible. This does not mean the gains are insignificant, but their statistical reliability remains unclear. The cross-domain results also do not show a consistent advantage: on SECOND, Table 5 reports 54.09 for UCDNet versus 57.66 for DMINet. Since temporal invariance and cross-domain invariance are different properties, this does not refute the causal model, but it leaves the claimed robustness benefit insufficiently supported. The authors should specify which variations the method is intended to remove and test them under controlled conditions. Reporting change-class IoU alongside mIoU would also make performance on the target class easier to interpret.

  3. The uncertainty component needs quantitative validation beyond qualitative maps. Uncertainty concentrated near object boundaries is plausible, but it does not show that the predicted probabilities are calibrated or that the uncertainty reliably identifies errors. This matters because uncertainty estimation is presented as a central contribution. The evaluation should include calibration measures, such as reliability diagrams, ECE, Brier score, or negative log-likelihood, together with an error-ranking evaluation such as risk–coverage or sparsification curves. A comparison against predictive entropy from a standard classification head with the same backbone would help establish the benefit of the evidential formulation. The bounded activation exp⁡(tanh⁡(⋅))\exp(\tanh(\cdot)) also deserves an ablation and an explanation of how its per-stage evidence range interacts with the subsequent evidence fusion.

  4. The contribution beyond UAFDcd needs to be made explicit. The published description of UAFDcd already includes a causal backdoor argument, correlation-based extraction of shared temporal features, differential feature separation, prototype contrastive learning, and Dirichlet-based uncertainty estimation [1]. Causal framing alone is therefore not enough to distinguish UCDNet. This overlap does not mean the methods are equivalent, but it calls for a precise comparison. The paper should identify the differences in the decomposition mechanism, prototype construction, contrastive objective, and uncertainty formulation. If maximum-entropy core-prototype selection is a main addition, comparisons against centroid, medoid, and other prototype-selection strategies would help establish its contribution. Controlled comparisons with matched backbones and training settings would be more informative than aggregate performance numbers alone.

Reference:

[1] B. Wang, S. Qiao, Z. Zhao, P. Qin, and J. Zeng. UAFDcd: Uncertainty-Aware Feature Decoupling Network for Remote Sensing Change Detection. IEEE Transactions on Geoscience and Remote Sensing, vol. 64, 2026. DOI: 10.1109/TGRS.2026.3701174.

Questions for authors

These are my questions for all the authors:

  1. The decomposition Fu=FT−FpF_u = F_T - F_p in Equations 1–2 relies on assumptions about how the features separate. How are these assumptions connected to the learned FpF_p? Please also specify the expectations, centering assumptions, and approximation regime used in Equations 1–2.

  2. Does YY in the causal graph denote the physical change label or the model's prediction? Do the other nodes represent physical scene factors or learned features? Please clarify how the arrows and interventions should be interpreted.

  3. Does "invariant" mean shared across timestamps, insensitive to appearance perturbations, or transferable across datasets? Which of these properties does the analysis establish? Could controlled appearance changes, with the physical change mask held fixed, be used to test the claimed mechanism?

  4. Could you report repeated-run results for Table 1, including the variability of the differences from the strongest baselines? Please clarify whether F1 is computed for the change class or macro-averaged, how scores are aggregated, and report change-class IoU alongside mIoU.

  5. When GFDG_{FD} is removed in Table 2, which features are supplied to LDCL\mathcal{L}_{DCL}, and what replaces the shared/private decomposition? Please also confirm the precision reported in the no-DCL row and explain the corresponding precision–recall trade-off.

  6. Why is the maximum-entropy candidate preferable to a centroid or medoid as the core prototype? What benefit does this selection provide independently of the rest of the contrastive-learning design?

  7. Why use exp⁡(tanh⁡(⋅))\exp(\tanh(\cdot)) for evidence generation? How does its bounded per-stage range affect the final fused evidence and uncertainty? Could you provide quantitative calibration and error-ranking results against a standard predictive-entropy baseline?

  8. What are the specific methodological differences from UAFDcd? Please also clarify the number of evidence stages, the cluster counts, whether λ\lambda is fixed across datasets, and whether code and dataset-specific configurations will be released.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 34sA (Part 1/4)

Q1:

Response:

Thank you for your insightful question. In Section 3.1 of the revised manuscript, we have explicitly clarified the mathematical link between the feature-separation assumptions and the learned feature FpF_p, along with its regime of validity:

  • First-Order Approximation & Regime of Validity: In Eq. (1), we adopt a first-order Taylor expansion as a local linear approximation of the physical factor mapping in the high-level deep semantic embedding space. This approximation holds within the regime where, on the high-level feature manifold of the bitemporal encoder, cross-temporal appearance perturbation displacements (ΔΦN\Delta \Phi_N) and authentic land-cover change displacements (ΔΦC\Delta \Phi_C) are approximately locally orthogonal to the shared background semantic manifold (ΦI\Phi_I). Under this condition, cross-correlations vanish in second-order statistics, enabling the learned cross-temporal shared feature FpF_p to effectively approximate the invariant background confounder ΦI\Phi_I.
  • Expectation & Centering Assumptions: Eq. (2) states E[ΔΦC1⊤ΔΦC2]=0\mathbb{E}[\Delta\Phi_{C_1}^\top \Delta\Phi_{C_2}] = 0 and E[ΔΦN1⊤ΔΦN2]=0\mathbb{E}[\Delta\Phi_{N_1}^\top \Delta\Phi_{N_2}] = 0, which rests on the premise that cross-temporal changes and random environmental perturbations are conditionally independent given the background and follow symmetric zero-mean distributions (E[ΔΦ]=0\mathbb{E}[\Delta\Phi] = 0) enforced by channel-wise feature centering.

Q2:

Response:

Thank you for pointing out the need to further clarify node semantics.

  • Clarification of Node Semantics:
    • I,C,NI, C, N represent high-dimensional physical generative latent variables (invariant background semantics, physical land-cover changes, and appearance perturbations such as illumination and sensor variations, respectively).
    • X={T1,T2}X = \{T_1, T_2\} represents the observed bitemporal remote sensing image pair.
    • YY represents the ground-truth physical surface change label (Ground Truth Change Mask), while the model prediction is denoted as Y^\hat{Y}.
  • Explanation of Causal Arrows & Intervention: The causal arrows I→XI \to X and C→XC \to X describe the data-generating process whereby surface physical attributes jointly generate the observed imagery. Without intervention, the invariant background II creates spurious correlations between observed features and change labels via the open backdoor path (X←I→YX \leftarrow I \to Y). The feature-level intervention do(X)\text{do}(X) in Fig. 2(a) explicitly estimates and strips the common-mode feature FpF_p (approximating ΦI\Phi_I) via the disentanglement module GFD\mathcal{G}_{FD}, thereby blocking the background confounding path in feature space and preserving pure change residuals FuF_u.
AuthorsOpenReview ↗

Response to Reviewer 34sA (Part 2/4)

Q3:

Response:

Thank you for this critical suggestion.

  • Conceptual Definition: In this work, "invariance" specifically refers to the insensitivity of temporal common-mode representations to irrelevant appearance perturbations (such as illumination gradients, seasonal phenology, and sensor gain variations), rather than cross-region or cross-resolution domain invariance.
  • Controlled Appearance Perturbation Intervention Experiment: Following your suggestion, we constructed a controlled intervention benchmark on the LEVIR-CD test set: keeping the ground-truth change mask strictly fixed, we applied four representative non-causal appearance perturbations to the post-temporal imagery and compared UCDNet against a parameter-matched Non-Causal Baseline (replacing GFD\mathcal{G}_{FD} with stacked standard convolutional residual blocks of identical parameter budget). The results are summarized below:
Intervention ConditionNon-Causal Baseline F1 / IoU1\text{IoU}_1 (%)UCDNet F1 / IoU1\text{IoU}_1 (%)ΔF1\Delta\text{F1} (Base vs. Ours)ΔIoU1\Delta\text{IoU}_1 (Base vs. Ours)
Clean (Original)89.15 / 80.4289.95 / 81.730.00% vs. 0.00%0.00% vs. 0.00%
Overexposure (γ=0.8\gamma = 0.8)88.40 / 79.2189.26 / 80.61-0.75% vs. -0.69%-1.21% vs. -1.12%
Shadow / Dark (γ=1.4\gamma = 1.4)87.18 / 77.2888.50 / 79.37-1.97% vs. -1.45%-3.14% vs. -2.36%
Seasonal Shift (ΔH=45,S×0.8\Delta H = 45, S \times 0.8)84.47 / 73.1285.76 / 75.07-4.68% vs. -4.19%-7.30% vs. -6.66%
Sensor Noise (σ=0.15\sigma = 0.15)87.50 / 78.4988.55 / 79.45-1.65% vs. -1.40%-1.93% vs. -2.28%

The results confirm that across all controlled appearance disturbances, UCDNet maintains superior absolute accuracy and exhibits substantially lower performance degradation under strong illumination, shadow, and seasonal shifts than the parameter-matched non-causal baseline, directly validating the causal disentanglement mechanism.

  • Visualizing Controlled Nuisance Interventions: To visually illustrate the severity and physical realism of the synthetic perturbations, while keeping the pre-temporal image T1T_1 and ground-truth mask YY strictly fixed, we apply four non-causal appearance perturbations exclusively to T2T_2 (visualized in Fig. 5 of the revised manuscript):
    • Overexposure: Simulates intense midday solar illumination and atmospheric scattering, markedly elevating overall brightness and washing out low-contrast surface textures.
    • Shadow / Dark: Simulates cloud occlusion and dusk conditions, severely compressing radiometric dynamic range in dark regions.
    • Seasonal Shift: Perturbs hue and saturation in the HSV space to replicate phenological vegetation color changes and spectral drift across seasons, altering surface colors without modifying physical building structures.
    • Sensor Noise: Injects high-frequency additive Gaussian white noise to mimic sensor hardware degradation and low-SNR transmission interference.

Q4:

Response:

Thank you for your thorough requirements.

  • Metric Definition & Aggregation Protocol: All F1, Precision, and Recall scores in the main text are strictly computed for the positive change class (C=1C=1) rather than macro-averaged across background and change classes; mIoU is the mean of change-class IoU (IoU1\text{IoU}_1) and unchanged-class IoU (IoU0\text{IoU}_0).
  • Repeated Runs & Statistical Significance Analysis: We re-executed both UCDNet and the strongest baseline (UAFDcd [1]) across 5 identical random seeds on LEVIR-CD, reporting sample mean and standard deviation (Mean±SD\text{Mean} \pm \text{SD}), net gain (Δ\Delta), two-tailed Welch's t-test pp-values, and Change-Class IoU (IoU1\text{IoU}_1):
MethodPrecision (%)Recall (%)F1-Score (%)OA (%)mIoU (%)Change IoU (IoU1\text{IoU}_1) (%)
UAFDcd [1]90.64 ± 1.0687.83 ± 0.9189.20 ± 0.1698.92 ± 0.0289.69 ± 0.1480.51 ± 0.26
UCDNet (Ours)91.43 ± 0.3688.19 ± 0.2289.78 ± 0.1598.98 ± 0.0290.19 ± 0.1381.45 ± 0.25
Net Gain (Δ\Delta)+0.79%+0.36%+0.57%+0.06%+0.50%+0.94%
pp-valuep=0.178p = 0.178p=0.432p = 0.432p=0.00036∗∗∗p = 0.00036^{***}p=0.0017∗∗p = 0.0017^{**}p=0.00038∗∗∗p = 0.00038^{***}p=0.00036∗∗∗p = 0.00036^{***}

(Note: * denotes p<0.05p < 0.05, ** denotes p<0.01p < 0.01, and *** denotes p<0.001p < 0.001.)

  • Key Conclusions: UCDNet consistently outperforms UAFDcd across all 6 metrics. All comprehensive segmentation indicators pass strict hypothesis testing: Change-Class IoU1\text{IoU}_1 increases by +0.94% (p=0.00036p = 0.00036), F1-Score by +0.57% (p=0.00036p = 0.00036), mIoU by +0.50% (p=0.00038p = 0.00038), and OA reaches 98.98% (p=0.0017p = 0.0017), confirming that the gains are statistically significant and reproducible rather than random fluctuations.

AuthorsOpenReview ↗

Response to Reviewer 34sA (Part 3/4)

Q5:

Response:

  • Feature Replacement Strategy When Removing GFD\mathcal{G}_{FD}: When the Pearson covariance-based disentanglement module GFD\mathcal{G}_{FD} is removed (w/o GFD in Table 2), to ensure strict alignment in parameter capacity and input structure for LDCL\mathcal{L}_{DCL} (consistent with the non-causal baseline in our controlled intervention experiments), we replace GFD\mathcal{G}_{FD} with parameter-matched stacked standard convolutional residual blocks. This baseline directly maps concatenated bitemporal features Concat(FT1,FT2)\mathrm{Concat}(F_{T_1}, F_{T_2}) via standard residual convolutions to estimate Fp=F(Concat(FT1,FT2))F_p = \mathcal{F}(\mathrm{Concat}(F_{T_1}, F_{T_2})) without any cross-temporal channel covariance modeling. Subsequently, the exact same residual subtraction (Fu1=FT1−FpF_{u_1} = F_{T_1} - F_p and Fu2=FT2−FpF_{u_2} = F_{T_2} - F_p) is performed, and Fp,Fu1,Fu2F_p, F_{u_1}, F_{u_2} are fed into LDCL\mathcal{L}_{DCL} for supervision.
  • Typographical Erratum & Precision–Recall Trade-off: The Precision value 99.59% in the w/o LDCL row of Table 2 in the initial manuscript was indeed a typographical transcription error (the actual recorded experimental value is 91.88%), which we sincerely apologize for and have corrected in the revised manuscript. Without contrastive margin expansion (w/o LDCL), the model struggles to push apart ambiguous negative samples in feature space, causing decision boundaries to become overly conservative (predicting changes only in high-confidence core regions), which manifests as high Precision (91.88%) paired with a sharp drop in Recall (86.66%) and lower overall F1.

Q6:

Response:

Thank you for this thoughtful question. We address this from both geometric manifold motivation and standalone ablation experiments:

  • Geometric Motivation: In RSCD, changed regions (e.g., farmland-to-factory, water-to-road) exhibit strongly multimodal heterogeneous distributions in feature space. Conventional centroids or medoids compute geometric centers that frequently fall into low-density transition zones between distinct change sub-modes ("false average samples"); pulling features toward such artificial centers over-shrinks decision boundaries and increases false negatives (lower Recall). Maximum-entropy core prototypes select real candidate samples within each sub-cluster that maximize mutual information and coverage uniformity relative to the overall distribution, avoiding the "mean trap" while suppressing boundary outliers.
  • Standalone Ablation of Prototype Selection: Freezing all other modules and hyperparameters of UCDNet and varying only the prototype selection strategy on LEVIR-CD yields:
Prototype Selection StrategyPrecision (%)Recall (%)F1-Score (%)mIoU (%)IoU1\text{IoU}_1 (%)OA (%)
Cluster Centroid92.26985.23988.61589.19579.55798.884
Cluster Medoid93.06984.64788.65889.23779.62798.897
Max-Entropy Core Prototype (Ours)91.93288.04589.94690.34081.72998.997

Simply switching the prototype selection rule improves Recall by +2.81% over Cluster Centroid and boosts F1-Score by +1.33% (Change IoU1\text{IoU}_1 by +2.17%), confirming its standalone contribution.


AuthorsOpenReview ↗

Response to Reviewer 34sA (Part 4/4)

Q7:

Response:

Thank you for this constructive question.

  • Bounded Evidence Mechanism & Multi-Stage Fusion: In Evidential Deep Learning (EDL), unbounded activations (such as Softplus or unbounded exponential) are prone to exponential explosion when feature activations are strong, causing total Dirichlet strength SS to diverge and inducing severe overconfidence. By adopting bounded exponential mapping exp⁡(tanh⁡(⋅))\exp(\tanh(\cdot)), single-stage evidence increments are strictly constrained within a stable bounded interval. This stage-wise bounded design ensures balanced weighting between shallow high-resolution edge textures and deep high-level semantics during cross-stage evidence accumulation, preventing large-norm deep responses from overwhelming shallow boundary uncertainty signals.
  • Quantitative Calibration & Error-Detection Comparison: We added quantitative evaluations of Expected Calibration Error (ECE), Brier Score, Negative Log-Likelihood (NLL), and Error Detection AUROC on the LEVIR-CD test set, comparing our method against a matched-backbone standard classification head (Shannon predictive entropy) and an unbounded Softplus evidence variant:
MethodECE ↓\downarrowBrier Score ↓\downarrowNLL ↓\downarrowError Detection AUROC ↑\uparrow
Standard Head (Softmax Predictive Entropy)0.01320.01590.07810.9320
UCDNet (Unbounded Softplus Evidence)0.03870.01580.08460.9763
UCDNet (Ours, Multi-Stage Bounded Evidence)0.01310.01520.06920.9731
  • Gains over Standard Predictive Entropy: Incorporating Dirichlet evidential modeling lowers Brier Score from 0.0159 to 0.0152 and NLL from 0.0781 to 0.0692, while boosting Error Detection AUROC from 0.9320 to 0.9731, demonstrating superior capability in identifying ambiguous boundary errors.
  • Necessity of Bounded Activation: While unbounded Softplus achieves a slightly higher AUROC (0.9763), the lack of an upper bound causes severe overconfidence, inflating ECE nearly threefold (from 0.0131 to 0.0387) and worsening NLL to 0.0846. Our bounded design achieves optimal probability calibration (ECE 0.0131) alongside top-tier error-detection discrimination (0.9731).

Q8:

Response:

Thank you for your question. We clarify the core methodological differences, hyperparameter consistency, and open-source plan below:

  • Methodological Comparison:
DimensionUAFDcd (IEEE TGRS 2026)UCDNet (Ours)Core Difference & Motivation
Disentanglement MechanismGated separation in difference spaceCross-temporal channel Pearson covariance adaptive projection matrix rrExplicitly models cross-channel second-order linear correlation filtering with stronger mathematical rigor
Prototypical ContrastConventional single-center static prototype alignmentDynamic K-Means + Max-Entropy Core Prototype selectionPreserves complex multimodal manifold topology of real changes and prevents single-centroid representation collapse
Uncertainty ModelingSingle-stage discriminative head Dirichlet outputCross-stage progressive multi-level evidence accumulation with bounded normalizationSynergistically fuses high-frequency spatial textures and high-level semantic evidence to calibrate boundary uncertainty
  • Hyperparameters & Open-Source Release: The number of stages (K=3K = 3), cluster counts (Kpri=3,Kpub=4K_{\text{pri}} = 3, K_{\text{pub}} = 4), and five loss weights (λ=[5,5,1.5,0.5,0.5]\lambda = [5, 5, 1.5, 0.5, 0.5]) are strictly fixed across all four datasets without dataset-specific tuning. Source code, pretrained weights, reproduction scripts, and full configuration files have been released via an anonymous GitHub repository.
Reviewer_wK6xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes UCDNet for remote-sensing change detection. It combines a structural causal model, Pearson correlation-guided feature disentanglement, maximum-entropy prototypical contrastive learning, and progressive evidential uncertainty estimation. On LEVIR-CD, SYSU-CD, SECOND, and CDD, UCDNet is competitive or best on several in-domain metrics, and the paper reports module ablations, cross-domain tests, uncertainty visualizations, and complexity. The method is plausible and the empirical package is reasonably complete, but the causal interpretation and uncertainty claims are stronger than the tests currently justify.

Critical strengths
  1. The paper targets a real failure mode in change detection: illumination, season, and sensor shifts can create spurious changes. The attempt to separate shared background from private change features is well aligned with this problem.
  2. The model is decomposed into identifiable components and the ablation table shows that removing the feature decoupler, contrastive objective, evidence loss, or reconstruction-related term reduces performance on LEVIR-CD.
  3. The evaluation spans four standard datasets and includes cross-domain generalization, visual examples, feature-space visualization, and an explicit FLOPs/parameter comparison. On LEVIR-CD, the reported F1 and mIoU are 89.95 and 90.34.
  4. The evidential output is potentially useful in practice because uncertainty is concentrated near ambiguous change boundaries rather than being presented only as a single confidence score.
Critical weaknesses
  1. The causal claims are not experimentally identified. The SCM is a useful conceptual decomposition, but the model is trained with ordinary supervised change labels and the reported gains could be explained by feature regularization and prototype learning. Please add interventions or synthetic nuisance shifts where the causal and non-causal factors are known, and compare against a non-causal model with the same parameter budget.
  2. The uncertainty module is evaluated mainly through visual alignment with errors. The paper should report calibration metrics such as ECE, Brier score, risk-coverage curves, AUROC for error detection, and performance under selective prediction. Without these measurements, it is difficult to assess whether the Dirichlet evidence is calibrated.
  3. Cross-domain results are substantially lower than in-domain results, and UCDNet is not uniformly best: DMINet performs better on SECOND. This is an important limitation for the stated motivation. The paper should analyze which nuisance shifts remain unresolved and report per-category or per-scene breakdowns rather than presenting a broad causal generalization claim.
  4. The loss has several components and the regularization weights are tuned on SYSU-CD, then reused elsewhere. The paper should state whether this tuning uses the test set indirectly, report sensitivity across seeds, and compare with a simple single-loss or independently tuned baseline.
  5. The claimed Pearson-based separation relies on a first-order statistical relationship between cross-temporal features and invariant background. The assumptions behind the derivation, especially conditional independence and vanishing cross terms, are strong for real remote-sensing scenes and should be tested or clearly framed as an approximation.
Questions for authors
  1. Can the authors add synthetic or real nuisance interventions with known causal factors and compare against a matched non-causal model?
  2. Please report ECE, Brier score, risk-coverage, error-detection AUROC, and selective prediction results for the evidential uncertainty.
  3. Why does DMINet outperform UCDNet on SECOND, and which domain shifts account for the remaining performance drop?
  4. How were the five loss weights selected, and were test labels or test-set statistics used at any stage?
  5. Which conditional-independence assumptions are required for the Pearson derivation, and how are they supported on real data?
2 comments
AuthorsOpenReview ↗

Response to Reviewer wK6x (Part 1/2)

Q1:

Response:

Keeping the ground-truth change mask fixed on the LEVIR-CD test set, we applied four representative non-causal appearance perturbations and compared UCDNet against a parameter-matched Non-Causal Baseline (replacing GFD\mathcal{G}_{FD} with stacked standard convolutional residual blocks of identical parameter count):

Intervention ConditionNon-Causal Baseline F1 / IoU1\text{IoU}_1 (%)UCDNet F1 / IoU1\text{IoU}_1 (%)ΔF1\Delta\text{F1} (Base vs. Ours)ΔIoU1\Delta\text{IoU}_1 (Base vs. Ours)
Clean (Original)89.15 / 80.4289.95 / 81.730.00% vs. 0.00%0.00% vs. 0.00%
Overexposure (γ=0.8\gamma = 0.8)88.40 / 79.2189.26 / 80.61-0.75% vs. -0.69%-1.21% vs. -1.12%
Shadow / Dark (γ=1.4\gamma = 1.4)87.18 / 77.2888.50 / 79.37-1.97% vs. -1.45%-3.14% vs. -2.36%
Seasonal Shift (ΔH=45,S×0.8\Delta H = 45, S \times 0.8)84.47 / 73.1285.76 / 75.07-4.68% vs. -4.19%-7.30% vs. -6.66%
Sensor Noise (σ=0.15\sigma = 0.15)87.50 / 78.4988.55 / 79.45-1.65% vs. -1.40%-1.93% vs. -2.28%

Under strictly aligned model capacity and training settings, UCDNet exhibits consistently smaller performance degradation under intense illumination, shadow, and seasonal color shifts, proving that GFD\mathcal{G}_{FD} suppresses non-causal appearance confounding rather than relying on parameter scaling or generic regularization.


Q2:

Response:

Thank you for your suggestion. We have supplemented Section 5.3.1 and Fig. 8 in the revised manuscript with complete quantitative calibration metrics and selective prediction (Risk–Coverage) evaluations on the LEVIR-CD test set, benchmarked against a matched-backbone standard predictive-entropy baseline and an unbounded Softplus evidence variant:

MethodECE ↓\downarrowBrier Score ↓\downarrowNLL ↓\downarrowError Detection AUROC ↑\uparrow
Standard Head (Softmax Predictive Entropy)0.01320.01590.07810.9320
UCDNet (Unbounded Softplus Evidence)0.03870.01580.08460.9763
UCDNet (Ours, Multi-Stage Bounded Evidence)0.01310.01520.06920.9731
  • Gains over Standard Predictive Entropy: Dirichlet evidential modeling improves probability calibration (Brier Score 0.0159→0.01520.0159 \to 0.0152; NLL 0.0781→0.06920.0781 \to 0.0692) and boosts Error Detection AUROC from 0.9320 to 0.9731.
  • Necessity of Bounded Activation: Unbounded Softplus induces severe overconfidence, inflating ECE nearly threefold (0.0131→0.03870.0131 \to 0.0387) and worsening NLL (0.0846), whereas our bounded activation prevents evidence explosion across multi-stage accumulation.
  • Risk–Coverage Curve: Fig. 8 confirms that rejecting high-uncertainty pixels monotonically lowers empirical error rates.

Q3:

Response:

Thank you for highlighting this key analytical point. We have calibrated the generalization scope in Section 1 and Section 5.2 of the revised manuscript and analyzed the structural and domain-shift factors on SECOND:

  • Clarification of Generalization Scope: Our causal disentanglement primarily targets cross-temporal appearance shifts (e.g., illumination and seasonal variations) within paired observations. While stripping these appearance confounders yields emergent cross-domain resilience—enabling UCDNet to outperform all baselines on CDD-CD and SYSU-CD (Table 5)—it is not designed to resolve severe cross-sensor spectral discrepancies or resolution gaps. We have revised Sections 1 and 5.2.4 accordingly to avoid overclaiming universal domain invariance.
  • Structural Factors on SECOND: SECOND contains fine-grained semantic change categories with substantial "spectrally similar but semantically distinct" transitions (e.g., concrete ground transitioning to low-rise buildings of similar material) as well as slender linear structures. DMINet emphasizes shallow dual-branch local convolutional texture interactions; conversely, UCDNet's GFD\mathcal{G}_{FD} relies on global channel-wise Pearson covariance statistics, which can smooth subtle local semantic variations into the common-mode background when local changes overlap heavily with background channel distributions, slightly lowering fine-grained recall.
  • Primary Domain-Shift Factors: The cross-dataset performance drop stems from: (1) spatial resolution gaps causing object scale mismatch; (2) spectral response discrepancies across different satellite sensors; and (3) cross-regional land-cover class prior shifts and phenological heterogeneity.
AuthorsOpenReview ↗

Response to Reviewer wK6x (Part 2/2)


Q4:

Response:

Thank you for your rigorous scrutiny of experimental protocol.

  • Strict Blind-Test Protocol: The five loss weights λ=[5,5,1.5,0.5,0.5]\lambda = [5, 5, 1.5, 0.5, 0.5] were determined exclusively via grid search on the independent validation set of SYSU-CD. Throughout hyperparameter exploration and model selection, all test-set images, labels, and statistics across every dataset remained strictly isolated (blind test) and were never accessed.
  • Cross-Dataset Parameter Freezing: Once determined on the SYSU-CD validation set, this exact hyperparameter configuration was frozen and directly transferred to LEVIR-CD, SECOND, and CDD-CD without any target-dataset re-tuning or fine-tuning, demonstrating strong hyperparameter robustness.

Q5:

Response:

Thank you for your in-depth question. We have explicitly supplemented the conditional independence assumptions and their empirical/architectural support in Section 3.1 of the revised manuscript:

  • Explicit Conditional Independence Assumptions: Vanishing cross-terms in expectation in Eqs. (1)–(2) rely on two conditional independence and symmetry conditions:
    1. Conditional Independence of Temporal Changes: C1⊥C2∣IC_1 \perp C_2 \mid I (given the same physical background substrate II, incremental land-cover changes across time steps are statistically independent in generative residuals).
    2. Zero-Mean Orthogonality of Perturbations: N1⊥N2∣IN_1 \perp N_2 \mid I, and environmental disturbances (illumination shifts, sensor thermal noise) follow symmetric zero-mean distributions relative to background mean ( E[ΔΦN]=0\mathbb{E}[\Delta\Phi_N] = 0 ).
  • Support in Real Remote Sensing Data & Architecture:
    • In the specific implementation of the GFD\mathcal{G}_{FD} module (Eq. 5), we explicitly subtract the channel-wise mean from the bitemporal features ( Qcentered=Q−QˉQ_{\text{centered}} = Q - \bar{Q} and Kcentered=K−KˉK_{\text{centered}} = K - \bar{K} ). This step effectively filters out overall brightness offsets and global tonal baselines in remote sensing imagery, enforcing a zero-mean feature distribution so that uncorrelated environmental disturbances naturally cancel each other out and approach zero when computing cross-temporal covariance.
    • Role as an Inductive Bias Approximation: These assumptions serve as a principled theoretical inductive bias guiding orthogonal disentanglement in deep embedding space rather than a strict physical linearity assertion. Both the high feature separability on real test sets (Silhouette Coefficient 0.9009, Fisher Discriminant Ratio 4.7372) and superior degradation resistance under controlled interventions empirically validate this approximation in feature space.
Reviewer_u6NLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UCDNet for remote sensing change detection, combining Pearson correlation-based feature disentanglement, prototype contrastive learning, and evidential uncertainty estimation. Experiments on four benchmarks demonstrate competitive detection performance and improved feature separability.

Critical strengths

The integration of feature disentanglement and uncertainty estimation addresses important challenges in remote sensing change detection.

Extensive benchmark comparisons, cross-domain experiments, and component ablations support the effectiveness of the framework.

Critical weaknesses

The causal interpretation is insufficiently justified. Correlation-based feature separation does not necessarily establish causal disentanglement or eliminate confounding.

The uncertainty evaluation relies mainly on qualitative visualizations, without quantitative calibration or uncertainty-quality metrics.

Questions for authors

What evidence supports the claimed causal disentanglement beyond improved feature separability?

Can the authors provide quantitative uncertainty calibration results, such as ECE or risk-coverage curves?

1 comment
AuthorsOpenReview ↗

Response to Reviewer u6NL

Q1:

Response:

Thank you for this pivotal question. Beyond static feature-space separability metrics (Silhouette Coefficient reaching 0.9009 and Fisher Discriminant Ratio reaching 4.7372), we added new empirical evidence supporting causal deconfounding in Section 5.3.1 of the revised manuscript:

  • Controlled Appearance Perturbation Intervention Experiment (Fixed Physical Labels): By definition of the SCM, a true causal extractor must respond to changes in physical node CC while remaining invariant to interventions on non-causal environmental node NN. Locking the ground-truth change mask on the LEVIR-CD test set, we applied four non-causal appearance perturbations and compared UCDNet against a parameter-matched Non-Causal Baseline (replacing GFD\mathcal{G}_{FD} with stacked standard convolutional residual blocks of equal parameter budget):
Intervention ConditionNon-Causal Baseline F1 / IoU1\text{IoU}_1 (%)UCDNet F1 / IoU1\text{IoU}_1 (%)ΔF1\Delta\text{F1} (Base vs. Ours)ΔIoU1\Delta\text{IoU}_1 (Base vs. Ours)
Clean (Original)89.15 / 80.4289.95 / 81.730.00% vs. 0.00%0.00% vs. 0.00%
Overexposure (γ=0.8\gamma = 0.8)88.40 / 79.2189.26 / 80.61-0.75% vs. -0.69%-1.21% vs. -1.12%
Shadow / Dark (γ=1.4\gamma = 1.4)87.18 / 77.2888.50 / 79.37-1.97% vs. -1.45%-3.14% vs. -2.36%
Seasonal Shift (ΔH=45,S×0.8\Delta H = 45, S \times 0.8)84.47 / 73.1285.76 / 75.07-4.68% vs. -4.19%-7.30% vs. -6.66%
Sensor Noise (σ=0.15\sigma = 0.15)87.50 / 78.4988.55 / 79.45-1.65% vs. -1.40%-1.93% vs. -2.28%
  • Isolation of Causal Mechanism: Under strictly matched model capacity and training hyperparameters, UCDNet exhibits smaller performance drops than the non-causal baseline across strong illumination, shadow, and seasonal shifts, proving that Pearson covariance disentanglement (GFD\mathcal{G}_{FD}) genuinely suppresses non-causal appearance confounding rather than relying on parameter scaling or generic regularization.
  • Visualizing Controlled Nuisance Interventions: To visually illustrate the severity and physical realism of the synthetic perturbations, while keeping the pre-temporal image T1T_1 and ground-truth mask YY strictly fixed, we apply four non-causal appearance perturbations exclusively to T2T_2 (visualized in Fig. 5 of the revised manuscript):
    • Overexposure: Simulates intense midday solar illumination and atmospheric scattering, markedly elevating overall brightness and washing out low-contrast surface textures.
    • Shadow / Dark: Simulates cloud occlusion and dusk conditions, severely compressing radiometric dynamic range in dark regions.
    • Seasonal Shift: Perturbs hue and saturation in the HSV space to replicate phenological vegetation color changes and spectral drift across seasons, altering surface colors without modifying physical building structures.
    • Sensor Noise: Injects high-frequency additive Gaussian white noise to mimic sensor hardware degradation and low-SNR transmission interference.

Q2:

Response:

Thank you for your valuable suggestion. We have supplemented Section 5.3.1 and Fig. 8 in the revised manuscript with comprehensive quantitative calibration metrics and Risk–Coverage curves on the LEVIR-CD test set, benchmarked against a matched-backbone standard classification head (predictive entropy baseline) and an unbounded Softplus evidence variant:

MethodECE ↓\downarrowBrier Score ↓\downarrowNLL ↓\downarrowError Detection AUROC ↑\uparrow
Standard Head (Softmax Predictive Entropy)0.01320.01590.07810.9320
UCDNet (Unbounded Softplus Evidence)0.03870.01580.08460.9763
UCDNet (Ours, Multi-Stage Bounded Evidence)0.01310.01520.06920.9731
  • Gains over Predictive Entropy: Dirichlet evidential modeling reduces Brier Score (0.0159→0.01520.0159 \to 0.0152) and NLL (0.0781→0.06920.0781 \to 0.0692) while elevating Error Detection AUROC from 0.9320 to 0.9731.
  • Necessity of Bounded Activation: Unbounded Softplus suffers from severe overconfidence, inflating ECE nearly threefold (0.0131→0.03870.0131 \to 0.0387) and degrading NLL (0.0846), whereas our bounded evidence formulation balances top error-detection discrimination (0.9731) with optimal calibration (ECE 0.0131).
  • Risk–Coverage Curve: Fig. 8 confirms that progressively filtering high-uncertainty pixels across 80%–100% coverage monotonically lowers empirical error rates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.