acceptodds
Under review as a conference paper at ICLR 2027

Causal Knowledge Distillation for Robust Remote Sensing Semantic Segmentation

Abstract

Knowledge distillation (KD) is widely used to compress large remote sensing semantic segmentation models for deployment on resource-constrained satellites. However, variations across satellite sensors and environmental conditions in real-world deployment can induce significant domain shifts between the source domain and unseen deployment domains. Existing KD methods primarily optimize source-domain accuracy while often overlooking robustness under domain shifts. To improve domain generalization in remote sensing semantic segmentation, we revisit KD from a causal perspective and attribute its limited robustness to two sources: domain-dependent noncausal knowledge transferred by the teacher, as well as noncausal shortcuts acquired by the student from source-domain supervision. Specifically, we propose Causal Knowledge Distillation (CKD), comprising Teacher Causal Knowledge Transfer (TCKT) and Student Causal Knowledge Acquisition (SCKA). TCKT learns prompts to refine teacher representations by suppressing domain-sensitive responses, thereby reducing the transfer of noncausal knowledge during distillation. SCKA applies causal frequency intervention to student inputs to mitigate the acquisition of domain-related shortcuts and encourage stable causal representation learning. In addition, CKD can be readily integrated with different distillation objectives without modifying the student architecture. Extensive experiments across multiple domain shifts demonstrate that integrating CKD with existing KD methods consistently improves robustness.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Knowledge Distillation, Semantic Segmentation, Causal Learning, Remote Sensing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_KP4kOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

To address the issues where traditional knowledge distillation in remote sensing semantic segmentation may transfer domain-specific biases from the teacher and lead the student to learn unstable shortcuts from source-domain supervision, the authors propose CKD. This framework comprises two components: TCKT and SCKA. The former utilizes synthetically generated weather-perturbed images to learn frequency-filtering weights—refining teacher features through a combination of task supervision and domain adversarial training—while the latter applies frequency-based interventions to the student's input, compelling the student to align with the refined teacher representations despite variations in visual appearance. Both components are employed solely during training and do not alter the student's deployment architecture. In my view, while this work clearly models two sources of bias in robust distillation, the interpretation of "causal knowledge" and the specific contributions of the custom-designed modules still require more direct validation.

Critical strengths
  1. The analysis of the problem offers valuable insights; by considering how to balance the distortions introduced during teacher-to-student knowledge transfer against the shortcuts students take in independent learning, it explains why merely improving the teacher's knowledge may still be insufficient to produce students with robust understanding.

2.The authors not only conducted experiments involving eight types of degradation and various backbones but also performed direct transfer tests from Potsdam to Vaihingen to demonstrate the model's generalization capability.

Critical weaknesses
  1. The paper does not provide explicit quality screening criteria, statistics on misalignment, or assessments of label consistency regarding the validity of pixel-level labels in the generated images.

  2. The paper combines frequency-domain modulation, domain adversarial learning, input perturbation, and knowledge distillation; however, the experiments primarily compare different distillation losses without adequately comparing approaches such as simple learnable frequency-domain masks or fixed frequency-band filtering.

Questions for authors

The main issue with the author's work is that the effectiveness of the modules and the claimed causal relationships have not been demonstrated through ablation studies and primary experiments.

2 comments
AuthorsOpenReview ↗

Response to Reviewer KP4k

Thank you for your insightful feedback and helpful suggestions.

W1. Label Consistency. To preserve pixel-level label consistency, we explicitly constrain the generation prompts to modify only image appearances while preserving the original semantic content and spatial structures. We further employ a two-stage screening procedure. First, we use the teacher model to evaluate generated images and discard samples with mIoU below 50% against the original labels. Second, we manually inspect the remaining images to remove samples exhibiting structural misalignment or semantic inconsistencies with the original pixel-level annotations.

W2. Frequency Filtering Comparisons. We further conduct ablations by replacing TCKT with fixed frequency-band filtering or simple learnable frequency masks, both without domain-adversarial learning, while retaining SCKA. As shown in the table, our method consistently outperforms both alternatives across all four distillation baselines. These results suggest that the improvements stem not merely from frequency filtering, but also from suppressing non-causal, domain-specific information through adversarial learning.

MethodKDCWDViTKDCustomKD
Fixed-band filtering66.2468.3266.7167.25
Simple learnable mask67.8568.8769.6367.81
Ours71.6172.4171.4371.35

Q1. Component Effectiveness and Causal Motivation. To address this concern, in addition to the frequency-filtering ablations discussed above, we conduct further experiments, including an undistilled student trained with SCKA and four distillation baselines (KD, CWD, ViTKD, and CustomKD) trained with the same Qwen-generated perturbation images as CKD (+Data Aug.):

MethodBackboneSource DomainDomain Shift Avg.
TeacherViT-L80.9375.63
StudentViT-B75.7368.08
Student+SCKAViT-B74.1366.72
KD | +Data Aug. | +CKDViT-B76.06 | 77.59 | 77.5265.64 | 64.72 | 71.61
CWD | +Data Aug. | +CKDViT-B78.17 | 77.52 | 78.3568.04 | 68.85 | 72.41
ViTKD | +Data Aug. | +CKDViT-B76.12 | 75.62 | 77.3663.12 | 64.57 | 71.43
CustomKD | +Data Aug. | +CKDViT-B76.71 | 77.52 | 77.0466.02 | 65.23 | 71.35

We further extend the cross-dataset evaluation by training on Potsdam and testing on Vaihingen to assess CKD's generalization performance:

MethodBackboneVaihingen
TeacherViT-L72.26
StudentViT-B55.08
KD | +CKDViT-B55.16 | 57.73
CWD | +CKDViT-B52.27 | 60.16
ViTKD | +CKDViT-B53.62 | 58.06
CustomKD | +CKDViT-B54.39 | 56.60
GKD | +CKDViT-B63.94 | 66.80

Together with the frequency-filtering ablations, these results support the effectiveness of CKD and its proposed modules, indicating that the improvements are not solely attributable to data augmentation. Instead, TCKT facilitates the transfer of causal knowledge from the teacher, while SCKA promotes its acquisition by the student, jointly enabling robust knowledge distillation under domain shifts.

Thank you for your response. I acknowledge the effectiveness of the proposed method. I am maintaining my current score because, while I consider this a solid piece of work, the novelty does not particularly impress me. Furthermore, the response to W1 fails to provide specific statistics regarding your process. Consequently, I have not raised my score, though I have increased my confidence rating. Good luck.

Reviewer_SmA5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a knowledge distillation method that can suppress the impact of domain shift on remote sensing semantic segmentation. This paper revisits knowledge distillation from the perspective of causal learning. Furthermore, it leverages frequency-domain characteristics to suppress domain-dependent factors in both teacher and student representations. Finally, the effectiveness of the proposed method is demonstrated through comparative experiments, ablation studies, and qualitative analysis.

Critical strengths
  1. This paper offers a causal insight into distillation.
  2. On the one hand, this paper offers a causal learning perspective to explain how domain shift affects knowledge distillation, providing insightful implications for future research. On the other hand, the experimental results demonstrate that the proposed method can effectively improve the domain generalization ability of the student model.
Critical weaknesses
  1. The SCKA module is not closely related to knowledge distillation.
  2. The paper proposes two modules, but SCKA seems more akin to a frequency-domain data augmentation method than a mechanism that directly addresses how domain shift affects the knowledge transfer process, unlike TCKT. Since the paper is positioned as a knowledge distillation method, the connection between SCKA and the distillation process is not sufficiently clear or well justified, which weakens the alignment between the proposed module and the paper's central claim.
  3. The formalization of SCKA module is incomplete.
  4. As a key component of the SCKA module, the construction of the image-dependent spectral weighting WXW_X is not adequately described in either the main text or the supplementary material. This makes the specific manner in which the SCKA module introduces adaptive frequency intervention ambiguous.
Questions for authors

Please see the weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer SmA5

We appreciate your thoughtful comments and careful evaluation.

W1/W2. SCKA and Knowledge Distillation. SCKA is not merely frequency-domain data augmentation but is closely integrated with the distillation process to mitigate the student's acquisition of spurious source-domain correlations. Specifically, the student learns from frequency-intervened images while aligning its representations with TCKT-refined teacher features from corresponding clean images. This cross-view distillation encourages the acquisition of domain-invariant, task-relevant knowledge. While TCKT suppresses non-causal domain information transferred from the teacher, SCKA prevents the student from relying on source-domain-specific information during knowledge acquisition. Furthermore, our additional experiments show that SCKA alone does not improve generalization (66.72 vs. 68.08), highlighting the importance of its integration with knowledge distillation rather than augmentation alone.

MethodBackboneSource DomainDomain Shift Avg.
StudentViT-B75.7368.08
Student+SCKAViT-B74.1366.72

W3/W4. WXW_X Construction. The image-dependent spectral weighting WXW_X is constructed from image-level gradient statistics and frequency locations to adaptively control perturbation strength while preserving task-relevant semantic information. Specifically,

\begin{cases} 0.10, & r(u,v)\le r_1(X),\\ 0.50, & r_1(X)<r(u,v)\le r_2(X),\\ 0.85, & r_2(X)<r(u,v)\le r_3(X),\\ 1.00, & r(u,v)>r_3(X). \end{cases}$$ Here, $r(u,v)$ denotes the normalized frequency radius, and $r_i(X)$ are image-adaptive thresholds determined by Sobel gradient statistics, following the adaptive frequency partitioning principle of CVGL (Ouyang and Zhu, NeurIPS 2025). We will clarify these details in the revision.
Reviewer_eyKjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Causal Knowledge Distillation (CKD). The teacher part of it filters out the teacher features that are relevant to the semantic task but insensitive to domain perturbations through frequency cues. The student part performs content-adaptive frequency intervention on the input images to enhance robustness during domain variations. The paper mainly completed the evaluation and testing of various perturbations on the Potsdam dataset.

Critical strengths
  1. The paper points out that not all knowledge in the teacher model can be transferred, and teachers may also pass on the sensitivity related to the source domain to students. Separating the processing of teacher-end knowledge filtering from student-end robust learning has clear practical significance.
  2. The method has good modular characteristics and shows relatively stable effectiveness under various configurations. CKD can combine multiple distillation targets such as KD, CWD, and ViTKD, and has achieved improvements on both DINOv3-Mask2Former and ResNet-DeepLabV3 architectures.
Critical weaknesses
  1. The description of innovation seems unclear. For instance, FreeKD has utilized the frequency cues provided by task supervision and pixel-level frequency masks to filter the teacher's knowledge. And TCKT shares obvious similarities with this design. The paper does not explicitly discuss or directly compare FreeKD.
  2. The main experiment of this paper verifies the robustness of synthetic perturbations, which is not true domain generalization. The authors also mentioned CrossEarth in the Related Work section, but did not conduct similar cross-domain experiments. The eight test conditions adopted in this paper all come from artificially corroding the same batch of Potsdam test images. They do not cover changes in geographical regions, sensors, platforms, spatial resolution, or category distribution, and therefore are insufficient to support a strong conclusion of domain generalization.Using only the Potsdam dataset for the cross-domain experiments seems to lack persuasiveness.
  3. The only truly cross-dataset experiment is also incomplete. The Potsdam to Vaihingen experiment is only presented in the appendix, and GKD and GKD+CKD are not reported; however, GKD is precisely the strongest and most relevant baseline in the main experiment. This omission undermines the credibility of the method's comparison under real dataset deviations
  4. The current experiments are unable to distinguish between the method mechanism and the benefits brought by the additional data augmentation. CKD used the weather perturbation images generated by Qwen, but the traditional distillation baseline did not obtain the same augmented data. The paper did not test the non-distilled students trained using SCKA or the same generated images, so some of the improvement might mainly come from the additional data.
Questions for authors

I have no further questions. I hope the author can provide their responses to the "Critical Weaknesses".

3 comments
AuthorsOpenReview ↗

Response to Reviewer eyKj

Thank you for your constructive feedback and valuable suggestions. We hope our revisions and additional experiments address your concerns and support a more favorable assessment of our work.

W1. Comparison with FreeKD. Our CKD central objective addresses a causal problem in robust distillation: teacher features contain both task-relevant causal knowledge and domain-dependent noncausal responses, and conventional distillation transfers both to the student. TCKT uses frequency prompts as an implementation vehicle, while its core design is adversarial learning that suppresses domain-sensitive teacher responses before transfer. SCKA further intervenes on the student side to reduce acquisition of domain-related shortcuts. We have cited related frequency-domain methods but omitted FreeKD. We will cite and discuss FreeKD explicitly, clarify the distinction, and add a direct comparison.

MethodBackboneSource DomainDomain Shift Avg.
TeacherViT-L80.9375.63
StudentViT-B75.7368.08
FreeKDViT-B78.2664.69
FreeKD+CKDViT-B76.7772.07

These results show that CKD is complementary to frequency-based distillation objectives and substantially improves robustness under domain shifts.

W2. Cross-Domain Evaluation. To further evaluate CKD's cross-dataset generalization, we have moved the Potsdam-to-Vaihingen evaluation, where models are trained on Potsdam and tested on Vaihingen, to the main manuscript. The results show that CKD consistently improves the performance of distillation baselines under cross-dataset shifts.

MethodBackboneVaihingen
TeacherViT-L72.26
StudentViT-B55.08
KD | +CKDViT-B55.16 | 57.73
CWD | +CKDViT-B52.27 | 60.16
ViTKD | +CKDViT-B53.62 | 58.06
CustomKD | +CKDViT-B54.39 | 56.60
GKD | +CKDViT-B63.94 | 66.80

We additionally evaluate CrossEarth, a domain generalization method, under eight domain shifts. Despite its domain generalization design, CrossEarth still exhibits performance degradation under these shifts (81.26 vs. 78.56). To further validate the effectiveness of TCKT in suppressing the transfer of non-causal, domain-specific information, we integrate it into CrossEarth, improving the average performance from 78.56 to 80.34 (+1.78).

MethodBackboneSource DomainDomain Shift Avg.
CrossEarthViT-L81.2678.56
CrossEarth+TCKTViT-L81.4180.34

W3. Additional GKD Comparisons. We have added GKD and GKD+CKD results to the Potsdam-to-Vaihingen evaluation and moved the comparison to the main manuscript.

MethodBackboneVaihingen
TeacherViT-L72.26
StudentViT-B55.08
KD | +CKDViT-B55.16 | 57.73
CWD | +CKDViT-B52.27 | 60.16
ViTKD | +CKDViT-B53.62 | 58.06
CustomKD | +CKDViT-B54.39 | 56.60
GKD | +CKDViT-B63.94 | 66.80

W4. Contributions of CKD. To disentangle CKD's contributions from data augmentation, we train four distillation baselines (KD, CWD, ViTKD, and CustomKD) using the same Qwen-generated images as CKD (+Data Aug.) and additionally evaluate an undistilled student with SCKA. As shown in the table, these perturbations serve merely as data augmentation for the baselines, yielding limited improvements or even performance degradation due to potential discrepancies with unseen domain shifts. In contrast, TCKT leverages these perturbations to suppress non-causal, domain-specific information in teacher features through domain-adversarial learning, facilitating the transfer of domain-invariant knowledge. Meanwhile, SCKA perturbs student input frequencies to discourage learning source-domain-specific information and promote the acquisition of causal knowledge during distillation. Therefore, CKD's core contribution lies in causal knowledge transfer and acquisition rather than data augmentation alone.

MethodBackboneSource DomainDomain Shift Avg.
TeacherViT-L80.9375.63
StudentViT-B75.7368.08
Student+SCKAViT-B74.1366.72
KD | +Data Aug. | +CKDViT-B76.06 | 77.59 | 77.5265.64 | 64.72 | 71.61
CWD | +Data Aug. | +CKDViT-B78.17 | 77.52 | 78.3568.04 | 68.85 | 72.41
ViTKD | +Data Aug. | +CKDViT-B76.12 | 75.62 | 77.3663.12 | 64.57 | 71.43
CustomKD | +Data Aug. | +CKDViT-B76.71 | 77.52 | 77.0466.02 | 65.23 | 71.35
Reviewer_eyKjOpenReview ↗

Official comment of eyKj

Thanks to the authors for the thorough response. The authors have addressed some of my concerns with additional experiments such as evaluating CrossEarth and Comparison with FreeKD. I will take these into further consideration.

Thank you for your positive assessment of our work and for recognizing the merits of our proposed approach. We greatly value your encouraging comments and constructive suggestions throughout the review process. We truly appreciate the time and careful consideration you have dedicated to evaluating our paper.

Comments on the paper

1 comment
Reviewer_SmA5OpenReview ↗

I will keep my rating as "Weak acceptance".

Thank you for the author's reply. I will keep my rating as "Weak acceptance".

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.