acceptodds
Under review as a conference paper at ICLR 2027

Breaking Data Silos with Scene-Degradation-Aware Learning for Generalizable Remote Sensing Cloud Removal

Abstract

Cloud removal models often struggle to generalize across diverse land-surface scenes and cloud degradations. While combining multiple datasets can broaden the range of training conditions, their uneven and redundant distributions limit the benefit of direct data aggregation; meanwhile, the increased diversity introduces substantially different restoration patterns that are difficult to accommodate with a single shared-parameter model. To address this challenge, we approach it from two complementary perspectives. First, Scene-Degradation-Aware Data Distillation (SDA-DD) derives a scene-degradation-aware sampling distribution over the pooled cloudy-clear pairs, encouraging the selected training data to be more evenly distributed across diverse restoration conditions. Second, we develop OmniCR with a Local-Routing Mixture-of-Experts (LR-MoE) architecture that routes different windows to experts with complementary restoration preferences, allowing spatially varying regions within the same image to adopt different restoration behaviors. Soft-Scene Optimization further emphasizes scene-degradation relations with restoration losses, preventing easy-restored conditions from dominating training. Experiments on multiple CR benchmarks demonstrate consistent performance gains across diverse scenes and cloud-degradation conditions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Remote Sensing, Cloud Removal, Image Restoration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gu6tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies cloud removal across different land surfaces and cloud conditions. It combines 300,948 cloudy–clear image pairs from 14 datasets.

SDA-DD selects a smaller training set. It uses features from clear images to describe the land surface. It uses feature differences between cloudy and clear images to describe degradation. The selection method gives more weight to rare conditions.

OmniCR uses different experts for different image regions. During training, it gives more weight to scene and cloud conditions with large errors. The paper tests the method on ten datasets.

Critical strengths
  1. The problem has practical value. Existing datasets contain different scenes and cloud types. Sections 4.1 and Appendix C explain why direct data pooling can create an uneven training set.

  2. The selected data help several models. In Table 2, SDA-DD improves MPRNet, Restormer, and OmniCR over random selection. The gains are 2.08, 2.35, and 2.87 dB, respectively. All three models use 20% of the data.

Critical weaknesses
  1. The tests do not show how much each new part helps. Table 2 compares different feature choices. It does not compare selection with and without inverse-density weights while keeping the features fixed. Thus, it does not isolate the value of these weights. In Table 3, local routing adds 0.05 dB. ReLU routing adds another 0.18 dB. These results do not explain the much larger gain of the complete system. A comparison with equal model size or compute would help. Appendix E measures coverage with the same score used to select samples. This test does not show whether the method beats a simpler balancing method. The cited [UniRain] also combines data selection, experts, and task weights. OmniCR uses different features and local routing. Section 2.1 should explain why these differences matter.
  2. Important result and cost details are missing. Table 1 reports four individual datasets and an average over ten. The paper should report the other six results. It should state whether the average gives equal weight to datasets or images. Appendix A explains how the method reduces memory use. It does not give the number of reference samples or the retained rank for the Nyström approximation. It also does not explain how the approximation keeps similarity scores nonnegative. Table 5 reports computation for one model pass. The comparison needs the full sampling cost and measured inference time. The authors should also give selection time, baseline training schedules, and variation across repeated runs.
Questions for authors

See weakness.

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

This paper proposes SDA-DD for sample selection based on joint scene–degradation representations and develops OmniCR by combining local expert routing with soft-scene difficulty reweighting. The aim is to improve the use of multisource remote sensing data and make a unified cloud removal model more adaptable across conditions.

Critical strengths
  1. The sampling strategy is effective across models. In Table 2, under the same 20% data budget, SDA-DD improves the PSNR of MPRNet, Restormer, and OmniCR over random sampling by 2.08, 2.35, and 2.87 dB, respectively, supporting its effectiveness across architectures.
Critical weaknesses
  1. The distinction from the closest prior work is insufficiently substantiated. UniRain (CVPR 2026) already combines data selection, MoE, and dynamic reweighting. Section 2.1 discusses only its retrieval and sampling components, without adequately comparing the architectures and optimization strategies or providing targeted experimental comparisons.
  2. Differences in the training schedule may confound the data-budget comparisons. The paper uses 100K training steps, a global batch size of 8, and a 10-epoch warm-up. If an epoch denotes one complete pass through the current subset, Soft-Scene would never be activated during training in the full-data and 50% settings. In contrast, it would be activated at approximately step 75,240 in the 20% setting. Thus, some comparisons in Figure 5 may vary in both data quantity and the loss-weighting schedule. The actual implementation requires clarification.
Questions for authors
  1. How is an epoch defined? Please specify the exact steps at which Soft-Scene is activated and how often difficulty updates occur under the 100%, 50%, and 20% data budgets.
  2. What is the most important new mechanism introduced by this work relative to UniRain?
  3. With the same features, a 20% data budget, and a fixed training schedule, does SDA-DD still outperform source-balanced random sampling and joint-kernel coverage selection without inverse-density weighting?
2 comments

The response did not impress me any further, so I will maintain the original score.

Thank you for reviewing our response. Your comments have helped us clarify our technical contributions and strengthen our experimental evaluation. We sincerely appreciate the time and effort you have devoted to reviewing our work.

Reviewer_42XMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the challenges of imbalanced data distributions, sample redundancy, and limited generalization of existing remote sensing cloud removal models. The authors propose a framework combining data distillation and adaptive restoration. First, Scene-Degradation-Aware Data Distillation (SDA-DD) jointly models scene content and cloud degradation features to select a more balanced training subset. Second, OmniCR employs a Local-Routing Mixture-of-Experts (LR-MoE) architecture to adaptively restore different image regions, together with a Soft-Scene Optimization strategy that dynamically adjusts training weights based on different scene-degradation relationships. The experiments use training data collected from 14 public cloud removal datasets and evaluate the method on the test sets of 10 datasets. Comparative experiments and ablation studies report improvements in PSNR and SSIM.

Critical strengths

SDA-DD is effective across different restoration models, rather than being limited to OmniCR. With only 20% of the training data, it achieves higher PSNR than random data selection on MPRNet, Restormer, and OmniCR, demonstrating its cross-model applicability and data efficiency.、

OmniCR handles different terrain and cloud degradation conditions within a single restoration framework. It achieves strong results on multiple public datasets and performs well across different terrain and cloud categories, showing its potential for diverse remote sensing cloud removal tasks.

Critical weaknesses

The technical novelty relative to prior unified image restoration methods is not sufficiently clear. UniRain (CVPR 2026) already combines data distillation, an MoE architecture, and dynamic reweighting. Although the proposed components differ in their specific designs, the paper does not adequately explain why these differences constitute substantial technical advances over existing approaches.

Questions for authors

The evaluation uses held-out test samples from data sources represented in the training candidate pool. Does the claimed generalization also extend to entirely unseen data sources? If so, are there experimental results to support this claim?

UniRain (CVPR 2026) also combines data distillation, an MoE architecture, and dynamic reweighting for unified image restoration. Given the similarity in the overall design, could the authors clarify the specific technical advances introduced in this paper and explain why they constitute substantial novel contributions beyond the existing framework?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.