acceptodds
Withdrawn from ICLR 2027

Rethinking Landslide Mapping as Coarse Prediction Under Weakly Defined Supervision

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent deep learning work usually treats landslide mapping as full-resolution semantic segmentation and seeks better performance through stronger backbones, multimodal fusion, or boundary refinement. Yet landslide supervision is often weakly defined: masks can be incomplete, visually interpreted, boundary-ambiguous, and misaligned with the region-level decisions that matter in practice. We revisit the task from a simpler angle: change only the output resolution. Without adding a new architecture, decoder, or inference stage, we vary the spatial granularity of labels and predictions and study how this alone changes performance. Across three datasets, moderate coarse prediction can improve region-level detection, although the gain is dataset-dependent and not monotonic with resolution. To separate useful coarse prediction from trivial task easing, we evaluate direct and pooled fine-resolution predictions on common grids, alongside label-faithfulness and anti-triviality diagnostics. The results point to a dataset-dependent reliable resolution sweet spot and suggest that output granularity is a practical design variable for landslide mapping rather than a minor implementation choice.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
landslide mapping, remote sensing, weak supervision, output resolution, coarse prediction, label uncertainty, semantic segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_9P9hOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper revisits the widely adopted full-resolution semantic segmentation paradigm for landslide mapping and proposes treating output resolution as an important task design variable. The authors argue that existing landslide datasets may suffer from ambiguous boundaries, incomplete annotations, and spatial inconsistencies, making pixel-level supervision not always the most appropriate choice. To address these issues, the paper employs max pooling to transform model predictions and ground-truth labels into different spatial resolutions and uses a BCE-Dice loss for training, without introducing a new network architecture. Experimental results demonstrate that appropriately reducing output resolution can improve region-level landslide detection performance on certain datasets, although the optimal resolution varies across datasets. However, the methodological novelty is relatively limited, and the authors should further clarify how the proposed approach differs from existing studies. Furthermore, the paper lacks an in-depth explanation of the underlying mechanisms that determine the optimal output resolution.

Critical strengths

The research question is of practical significance. This paper aims to address spatial uncertainty in landslide mapping supervision and the potential mismatch between pixel-level predictions and practical region-level decision-making needs. Rather than focusing solely on improving network architectures, the study offers a valuable and practically relevant perspective on landslide mapping.

Critical weaknesses
  1. The methodological novelty is limited, and the distinction from existing studies needs to be further clarified. The core method involves applying max pooling to the model's output logits and supervision labels, followed by loss computation at a coarser spatial resolution. This operation is relatively common and is closely related to multiple instance learning, region-level supervision, and pooling-based weakly supervised segmentation. Although the authors emphasize that their contribution lies in reformulating the landslide mapping task rather than proposing a new network architecture, the fundamental differences from these existing approaches have not been sufficiently explained. I suggest including representative baselines from task reformulation or weakly supervised segmentation methods. The authors should also clearly distinguish between the technical novelty of the proposed method and the empirical findings specific to landslide mapping. For a machine learning conference such as ICLR, it is particularly important to clarify what broader insights these findings offer to the machine learning community.

  2. The causal relationship between weak annotations and the benefits of coarse-grained prediction remains unclear. The central motivation of this paper is that coarser output resolutions can alleviate supervision issues arising from incomplete annotations and ambiguous boundaries. However, the current experiments do not provide sufficient evidence to establish this mechanism.

  3. The underlying mechanisms determining the optimal output resolution are not sufficiently explored. One of the main conclusions of this paper is that different datasets exhibit different reliable resolution sweet spots. However, the current analysis relies primarily on empirical sweeps over output resolutions, without providing a quantitative explanation of the relationship between the optimal resolution and dataset characteristics. Establishing a more interpretable principle for selecting the appropriate output resolution would substantially enhance the scientific value of this work.

  4. The practical value of the proposed approach has not been sufficiently validated. The paper emphasizes that coarse-grained predictions better align with the practical needs of region-level landslide screening, but it lacks more concrete downstream application evaluations. I suggest including additional application-oriented evaluation metrics, such as object-level recall and small-landslide detection rates.

  5. The experimental results should be supported by repeated runs. I recommend repeating the main experiments with at least five different random seeds and reporting the mean and standard deviation to better assess the robustness and statistical variability of the results.

Questions for authors

Please refer to the Critical Weaknesses section for details.

Reviewer_5iSLOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper explores the impact of the spatial resolution of segmentation labels in landslides mapping. The authors propose to pool the model predictions into coarser grids and train without any other changes to the architecture. The approach is evaluated on 3 datasets and compared against pooling of the full resolution predictions. The results show that coarser predictions can improve region-level detection, and that the best resolution is dataset-dependent.

Critical strengths
  1. The method is simple and can be used on any segmentation model.
  2. The evaluation framework is reasonable.
Critical weaknesses
  1. The proposed approach has limited novelty and is very much related to multiple-instance learning, which is not discussed in the manuscript.
  2. Missed landslides in the labels can only be recovered if they get aggregated in the same cell with another landslide using the proposed max pooling.
  3. The experiments are basically a sweep over a hyperparameter with dataset dependent results. They do not provide insights on how the optimal r relates to the landslides size, GSD or label noise, nor how it can be predicted for new datasets.
  4. The result of the comparison between coarse training and pooling could also be explained by the fact that the former is trained on the same cell-level task it is evaluated on.
  5. In Figure 4, the pooled r=1 baseline (orange) changes smoothly with resolution, but the proposed coarse-trained models (blue) oscillates and does not show a clear trend on Sen12. Moreover, it is not clear whether the differences significantly exceed run to run variance.
  6. A more complete comparison with label relaxation could be added to the experiments.
  7. The scope of the paper is narrow and more appropriate for a remote sensing venue.
Questions for authors
  1. Which version of the Sen12Landslides labels is used? It would be interesting to compare coarse training and pooling on the original vs. the refined labels.
  2. Is the optimal r related to the landslide size, GSD or label noise? Otherwise, it is unclear how r should be chosen for a new dataset.
Reviewer_KSxeOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies coarse-resolution supervision for landslide mapping when pixel-level masks are incomplete or have ambiguous boundaries. It compares directly trained coarse predictions with pooled fine-resolution predictions on three datasets and introduces diagnostics for spatial fidelity and trivial outputs. The task-level perspective is relevant, but the evidence for improved robustness under weak supervision remains limited.

Critical strengths
  1. The paper addresses a practical issue in landslide mapping: pixel-level annotations can be incomplete and boundary-ambiguous.
  2. Comparisons across three datasets and common evaluation grids provide a useful starting point for studying coarse prediction.
  3. The effort to diagnose label faithfulness and degenerate predictions goes beyond reporting coarse-grid F1 or IoU alone.
Critical weaknesses
  1. Weak-label robustness. Max-projected targets (Eq. 3) cannot recover landslide events entirely missing from annotations. The perturbation experiments in Sec. 5.6 would be more convincing with controlled event-level omission tests.
  2. Source of improvements. Coarse training pools dense logits (Eq. 2), while post-hoc pooling already explains part of the gains. A native low-resolution prediction head, evaluated with the same targets and grids, would help isolate the effect of coarse supervision.
  3. Reliability assessment. Table 2 reports NCPR = 0 alongside IoUS = 0.8499 at 128 × 128 resolution. This suggests that a high skill score alone does not guarantee a spatially meaningful prediction.
  4. Evaluation and stability. The scale- and threshold-selection procedures should be stated clearly, and key results should be reported across more random seeds. On GVLM, strict IoU is essentially unchanged (0.715 vs. 0.716; Table 1).
  5. Writing and presentation. The manuscript contains awkward negative constructions and repetitive contrastive phrasing, which makes parts of the writing feel mechanical. More direct academic prose, cleaner figure layouts, larger fonts, and more legible labels would improve readability.
Questions for authors
  1. Can the authors test event-level omissions and other controlled annotation errors while keeping validation and test labels unchanged?
  2. Can a native low-resolution prediction head be compared against pooled dense logits to isolate the effect of coarse supervision?
  3. How should the high IoUS but zero NCPR result at 128 × 128 resolution be interpreted, and what safeguards prevent scale selection from favoring spatially trivial predictions?

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

Related papers

Discussion (0)

Sign in to comment.