acceptodds
Withdrawn from ICLR 2027

What Caused Changes? Causal Physics Disentanglement Network for Remote Sensing Change Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

To address sensitivity to imaging environment variation and limited physical interpretability in traditional remote sensing change detection (RSCD), we propose causal physics disentanglement network combining RS imaging physics with causal inference. Physical parameter decomposition separates intrinsic ground-object information from environmental interference. Moreover, surface reflectance difference serves as core causal evidence for change discrimination, while non-causal backdoor paths induced by illumination and atmospheric variation are explicitly blocked. Based on main RS energy transfer paths (i.e., direct solar irradiance, diffuse sky irradiance, and atmospheric path radiance), spectral radiance model of ground objects is firstly constructed.Secondly, causal inference clarifies relationships between physical variables and RSCD. Specifically, surface reflectance variation serves as core physical evidence of real ground-object change. Meanwhile, total irradiance and atmospheric transmittance are treated as non-causal imaging factors that may induce pseudo-change response. Building on this, bi-temporal physical parameter decomposition and reconstruction network is developed to disentangle input image features into total irradiance, surface reflectance, and atmospheric transmittance, etc. Moreover, orthogonality, reconstruction, and spatial smoothness constraints are jointly imposed to ensure decomposed physical parameters plausibility. Finally, reflectance guided dual-level consistency constraint is introduced. Bi-temporal reflectance difference serves as primary causal pathway for RSCD. Specifically, it guides semantic branch at feature and prediction levels toward real ground-object change. Meanwhile, non-causal backdoor paths from total irradiance and atmospheric transmittance through observed imagery are explicitly blocked. Experimental results demonstrate that proposed model outperforms existing state-of-the-art methods in both qualitative and quantitative metrics on multiple public RSCD datasets, fully verifying its effectiveness and adaptability to complex scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Remote Sensing Change Detection, ground-object spectral radiance model, Causal Inference, Causal Physical Decoupling

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_SxV3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CPDcd, a change detection network meant to tell real ground change apart from differences in lighting and atmosphere between the two dates. Starting from a simple model of how sunlight reaches the sensor, it splits the features of each image into three maps meant to stand for illumination, atmospheric transmittance and surface reflectance, and predicts change mainly from how the reflectance map differs between the dates. Extra losses keep the three maps apart, keep the illumination and atmosphere maps smooth, ask the maps to rebuild the input image, and discourage the illumination and atmosphere maps from predicting change on their own. A causal graph is used to argue that this removes spurious cues from lighting and haze. The method reports F1 of 90.33 on LEVIR-CD, 69.38 on SECOND and 80.11 on SYSU-CD, ahead of eight retrained baselines.

Critical strengths

The problem is real. Pseudo-change from illumination, haze and seasonal radiometry causes a large share of false alarms in bi-temporal imagery, and putting an explicit multiplicative image model into the loss is a cheap and readable way to go after it. If the three maps behaved as named, EE and TT would also be useful diagnostics in their own right. It counts in my assessment only as framing, though.

Critical weaknesses
  1. The three maps are not shown to be physical, and the robustness claim is never tested. The inputs are 8-bit RGB products, not calibrated radiance, and the maps are linear projections of encoder features with no sun or atmosphere input. The reflectance map is trained through the change labels, and in Figure 8 it looks like a building mask. Path radiance, the term through which haze mostly shows up in visible bands, is dropped after Eq. 4. Why drop the one term that carries haze? Section 4.4 does not help. Narrow histograms (Fig. 4) are what the smoothness loss produces anyway, a transmittance of 0.2 to 0.33 would mean heavy haze, the ENVI curves in Figure 6 cover about 1.5 to 2.75 µm instead of visible light, and all 36 values in Figure 5 are multiples of 1/255, as if read from an 8-bit image. Most important, there is no test under lighting or haze change (perturbed test pairs, a seasonal subset, cross-dataset transfer) and no component ablation, since Table 2 only sweeps loss weights. Whether the physics branch adds anything over the encoder and difference U-Net is therefore unknown. Retinex-style decomposition (Wei et al., BMVC 2018), the closest prior idea, is not discussed.

  2. The comparison does not support the state-of-the-art claim. BIT and ChangeFormer get 80.01 and 84.40 F1 on LEVIR-CD in Table 1, against 89.31 and 90.40 in their own papers on the standard 256$\times$256 split, so if that split was used, CPDcd at 90.33 is below the published ChangeFormer number. Table 1 is also internally inconsistent. ChangeFormer on SECOND lists F1 66.79 with P 74.61 and Re 50.58, which give 60.29. The λ\lambda setting comes from Table 2, whose selected row reproduces the Table 1 LEVIR-CD test numbers to the second decimal, so Table 2 reports test results and nothing indicates that λ\lambda was chosen on a validation split. The top three settings are within 0.02 F1 of each other anyway.

Questions for authors
  1. Which encoder, crop size and train, val and test split did you use for each dataset, and how was SECOND binarized?
  2. Which output is the final change map? The conclusion mentions adaptive fusion of a reflectance branch and a semantic branch, but Section 3 never defines it. Fig. 2(c) draws Y^\hat{Y} from the U-Net row and labels that row DPD_P and PPP_P, the reverse of Eq. 14, and Eqs. 16 to 18 never supervise PFP_F with YY directly.
  3. Focal loss is unbounded as the true-class probability goes to zero, so the two negative terms in Eq. 16 can be driven to minus infinity by making PEP_E and PTP_T confidently wrong. How is Lpc\mathcal{L}_{pc} kept bounded? A prediction that is confidently wrong also still carries the change information.
  4. Where do the Figure 5 values come from, and which box is which class?
  5. Can you add a stress test, for example gamma, brightness and haze applied to one date of the LEVIR-CD test pairs, with false alarm rates for CPDcd and for the same backbone without the decomposition?
Reviewer_MepAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CPDcd, a causal physics disentanglement framework for remote sensing change detection. It incorporates radiative transfer modeling to separate irradiance, atmospheric transmittance, and surface reflectance. A reflectance-guided consistency mechanism is introduced to suppress environmental interference. Experiments on LEVIR-CD, SECOND, and SYSUCD demonstrate competitive performance.

Critical strengths
  1. Integrating radiative transfer modeling with change detection is meaningful. The proposed decomposition provides an interesting perspective on separating environmental variations from genuine land-cover changes.
  2. The method achieves promising results across three benchmark datasets. The qualitative visualizations also demonstrate improved change localization and boundary discrimination.
Critical weaknesses
  1. The proposed causal intervention is mainly implemented through prediction consistency constraints. It remains unclear whether these constraints genuinely block spurious causal pathways. Moreover, the physical decomposition lacks sufficient evidence of parameter identifiability.
  2. Equation (16) subtracts classification losses for irradiance and transmittance branches, potentially encouraging incorrect predictions rather than causal independence. The experiments also lack component-removal ablations and controlled evaluations of environmental robustness.
Questions for authors

Q1. Could the authors clarify how the proposed consistency constraints implement causal intervention rather than conventional feature regularization?

Q2. How does Eq. (16) prevent unstable optimization or encourage causal independence when classification losses are subtracted?

Q3. How is the identifiability of irradiance, reflectance, and transmittance guaranteed from observed images?

Q4. Could the authors provide separate ablation studies for the decomposition network, orthogonality constraint, and causal consistency mechanism?

Q5. Could the authors evaluate robustness under controlled illumination and atmospheric variations to support the claimed environmental invariance?

Q6. Could the authors provide quantitative validation of the estimated physical parameters against real reflectance or atmospheric measurements?

Q7. Could the authors clarify how the final change prediction is supervised and how the reflectance and semantic branches are fused?

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

The paper proposes CPDcd, a "causal physics disentanglement network" for remote sensing change detection (RSCD). It first builds a spectral radiance model that decomposes a bi-temporal RS image into three physical quantities — total irradiance E, atmospheric transmittance T, and surface reflectance P (Sec. 3, Eq. 1–11). These are framed as a causal graph where reflectance change P→Y is the genuine cause of true change, while E and T act as non-causal factors that create pseudo-changes through observed imagery (the "backdoor" paths, Fig. 2b). A shared encoder disentangles features into E/T/P, regularized by orthogonality, reconstruction and smoothness losses (Eq. 10–17), and a reflectance-guided dual-level (feature- and prediction-level) consistency constraint is used to "block" the non-causal paths. Experiments on LEVIR-CD, SECOND and SYSU (Table 1) report gains over 8 baselines, with ablations on loss weights (Table 2) and qualitative/physical-plausibility visualizations (Figs. 3–8).

Critical strengths

1.Principled physics-grounded formulation.Casting RSCD as decomposition into irradiance/transmittance/reflectance with an explicit radiance model (Eq. 1–11) and using reflectance difference as the change cue is well-motivated and physically interpretable; the resulting decoupled variables are shown to follow plausible distributions (Fig. 4) and consistent reflectance ordering (Fig. 5–6, Sec. 4.4), which is a meaningful interpretability advance over black-box CD. 2.Consistent empirical gains. CPDcd improves the change F1/mIoU over 8 competitive baselines across three datasets (Table 1), and the loss-weight ablation (Table 2) supports the contribution of the orthogonality/reconstruction/alignment terms.

Critical weaknesses

1.The central "causal" claim is asserted, not implemented or identified.The paper's title and core thesis rest on "blocking non-causal backdoor paths" via intervention, yet no causal graph formalization, do-operator, backdoor adjustment, counterfactual or re-weighting appears in the method (Sec. 3, Eq. 1–17 only contain reconstruction/orthogonality/alignment losses). As presented, "causality" is a narrative label over an architectural design rather than a causal-identification procedure, so the claimed disentanglement of the true cause from confounders is unjustified. 2.Physical disentanglement is unverified in identifiability and the evaluation lacks statistical rigor. E/T/P have no ground-truth supervision and plausibility is argued only through aggregate histograms (Fig. 4), leaving scale-ambiguity/identifiability of Eq. 11 unresolved; and Table 1 reports single-run numbers with no variance/significance tests despite small margins over FCdDNet, no efficiency analysis, and no cross-dataset OOD test to support the robustness claims. Either weakness alone is borderline; together they materially reduce confidence in the paper's central contribution.

Questions for authors

1.Please specify exactly how "blocking backdoor paths" is computed. Is there an explicit DAG and any do-operator / backdoor adjustment / re-weighting / counterfactual loss, or is "causal" purely a description of the E/T/P disentangling architecture? If purely architectural, please state the assumptions under which this design provably removes the E/T→X→Y spurious path. 2. Are there ground-truth E, T, P annotations for any pixel? If not, what evidence distinguishes a physically correct decomposition from an arbitrarily rescaled one (Eq. 11 is scale-ambiguous)? Please add an identifiability argument or a synthetic-truth experiment. 3. Please report mean±std over ≥3 seeds and a significance test on Table 1, and add FLOPs / #params / inference time relative to BIT, ChangeMamba and FCdDNet. 4. What is the cross-dataset / cross-sensor generalization performance (train on LEVIR-CD, test on SECOND/SYSU), given the robustness claims? Is code/public release planned? 5. Have you considered comparing against explicit-causal baselines such as Wu et al. (2026) and CARL (Li et al., 2026) cited in your related work?

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

Related papers

Discussion (0)

Sign in to comment.