acceptodds
Under review as a conference paper at ICLR 2027

Terrain-Informed Landslide Learning for Spatially Explicit Landslide Prediction

Abstract

Landslide prediction from remote sensing data is commonly formulated as a pixel-level segmentation problem, where accurate delineation requires the integration of spectral, terrain, and spatial information across multiple scales. Existing deep learning models can capture these representations effectively, but improvements in prediction accuracy are often accompanied by increased model complexity and computational cost. To address this accuracy–complexity trade-off, we propose TILL (Terrain-Informed Landslide Learning), a terrain-aware framework for spatially explicit landslide prediction. TILL integrates Band-Group Attention and a Terrain-Gated Encoder to jointly model spectral groups and terrain characteristics, followed by Multi-Scale Feature Aggregation and decoding to preserve spatially detailed representations. A Recurrent Refinement Unit subsequently refines the initial prediction using the learned features and prediction state. This design combines terrain-aware encoding, multi-scale representation, and recurrent prediction refinement within a unified segmentation pipeline. TILL produces pixel-level landslide probability maps with an accuracy of 0.85 using 33M parameters and 38G FLOPs. TILL is trained and evaluated on the Landslide4Sense dataset, which is specially designed for landslide detection. Compared with the average performance of the evaluated methods, TILL has a better accuracy-complexity tradeoff, with 9.68% improvement on accuracy, 7.69% fewer parameters, and 1.94% fewer FLOPs. The results show that TILL provides a considerable balance between landslide prediction accuracy and model complexity for spatially explicit delineation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Deep Neural Network, Landslide Prediction, Multi-modal

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

This paper investigates landslide segmentation based on aerial imagery. It proposes to integrate digital elevation model to add terrain information and improve segmentation, alongside a band attention strategy to improve the analysis from multispectral data. It evaluates its contributions through the prism of accuracy and computational efficiency on the Landslide4Sense dataset.

Critical strengths
  • Integrating digital elevation model to perform landslide segmentation is a good idea.
Critical weaknesses

A. Writing

While not critical in itself, the writing is overall quite subpar. A lot of concepts are only superficially described, e.g. the Band-Group Attention and the Terrain-Gated Encoder which are supposed to be big contributions. I am not confident I would be able to reimplement the TAG based on the description from 3.1.2, for example. There are also many weird details in the writing, such as out-of-place equations, very detailed trivial stuff (Equations 2, 5, 6, 10, 11).

B. Unconvincing experimental results

Overall the experimental results have failed to convince me that they properly support the claims. For example, the paper claims that the proposed method is on the optimal Pareto front for the accuracy/compute trade-off (measured in FLOPs), but it is obviously not: as shown in Figures 3 and 4, simply removing the final module decreases mIoU by 2 points but improves FLOPs count by 15%.

Compared to previous works, the results are also reported on a single benchmark (LandSlide4Sense), while common practices consist in also evaluating on the Bijie Landslide and Luding Landslide datasets. This is problematic considering that all the ablations and the hyperparameters tuning have been performed on the Landslide4Sense dataset, which makes the reported result optimistic and we cannot be sure the tweaked architecture with all its bells and whistles will properly generalize on other benchmarks.

Last but not least, the relative improvements metrics make no sense (Equations 10 and 11): they are heavily dependent on the choice of baselines. Choosing underperforming baselines (which is the case as I will explain below) result in overstated relative improvements that bear no relation with the actual SotA in landslide segmentation.

C. Related work on Band Group Attention

The paper claims to introduce band group attention but it is not clear how novel this is. There have been plenty of papers contributing band-level attention, especially for hyperspectral imagery: SiamBAG, Attend in Bands, BAE-Net, etc.

The proposed implementation (Eq (3)) is overall a straightforward attention but on grouped channels, which does not seem particularly new. In addition, this contribution is never evaluated with a dedicated ablation study anyway.

D. Related work and comparison to prior landslide segmentation models

Continuing of the previous trend of missing related work, there is no comparison or discussion to previous terrain-aware models. Yet, integrating the DEM into segmentation models is a relatively hot topic, especially for floods (e.g. Tan et al., Adhikari et al.) and landslide (see in particular Dar et al., ECCV 2026, Zhang et al., JSTARS 2026, Zhao et al., PLOS 2025, Liu et al., TGRS 2025, etc.).

The absence of any discussion of these papers has two broad consequences that severely undermine the contribution of the reviewed manuscript:

  • the proposed "terrain-gated encoder" is never compared or positioned to existing "terrain-aware" modules from the state of the art of computer vision for remote sensing, despite many relevant works existing;
  • the only baselines reported in the experimental results are generic semantic segmentation models, which ignore previous baselines specifically tailored for landslide segmentation including FFS-Net and TerrainGraphtNet.

This is made worse by the fact that the reported metrics on LandSlide4Sense (OA and mIoU) are inconsistent with the state of the art. For example, the paper claims 76.2% mIoU (which is on par, but slightly below, the 76.39% reported by TerrainGraphNet), but also reports an amazing ~85.5 F1 score which blows TGN's 70.3 out of the water, meanwhile reporting ~85% accuracy, a complete downfall against TGN's ~98.6.

While I'm not saying the reported numbers are wrong, my point is that the absence of comparison makes it very hard to understand what is going on. Is the method better than the state of the art? It is impossible to know based on the manuscript.

Questions for authors

See my main questions above.

I have some additional minor remarks regarding the presentation:

  • l.43: missing reference ot Landslide4Sense
  • l.103: I've never seen this symbol for the indicator function, it seems an uncommon choice.
  • Equation (2) gives no useful information
  • Figure 4(b): I'm not sure what this is supposed to show.
  • It's unclear how the hyperparameters (e.g. l.396-400) have been tuned.
  • Figure 6 and 4.4: I don't see how this shows spatial uncertainty.
Reviewer_j13kOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a new architecture (TILL) for landslide mapping that integrates terrain information with a better accuracy/complexity tradeoff. The model is evaluated on one dataset and compared to simple baselines.

Critical strengths
  1. The paper is easy to follow.
  2. The architecture is sound and quite detailed, which is good for reproducibility.
Critical weaknesses
  1. Limited novelty, the architecture mainly combines standard components.
  2. Related work on deep learning for landslide mapping is poorly covered.
  3. Only one dataset for evaluation.
  4. Advertised gains (+9.68% accuracy, −7.69% parameters, −1.94% FLOPs) are measured against the average of the four baselines, not against any actual model. This is misleading.
  5. No comparison against efficient encoders (e.g., MobileNet).
  6. Experimental protocol for the baselines is not detailed enough.
  7. Figure 6 is not informative.
  8. Fig. 7 has several errors: the landslide mask is labelled “Band 4 (665 nm)”, Blue (B2) is missing, B8a (absent from Landslide4Sense) is included, and the “spectral depth” arrow runs along a spatial dimension.
  9. Several inconsistencies in the manuscript: the ablation gains stated in §4.2 match neither panel of Fig. 3, the two panels report different F1 for the same configuration (0.838 vs. 0.783), and Fig. 5 shows 100 training epochs while Table 1 specifies 60.
Questions for authors
  1. How were the baselines configured and trained: input bands, backbone, loss?
  2. How do Fig. 3(a) and Fig. 3(b) relate, and which gives the correct ablation numbers?
Reviewer_x8RtOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

TILL is a landslide segmentation architecture that integrates multispectral imagery with slope and elevation data. It incorporates components such as band-group attention, multi-scale feature aggregation, terrain-aware gating, iterative mask refinement, and deep supervision. In experiments using the Landslide4Sense dataset, the accuracy, parameter count, and FLOPs of four different segmentation models were compared, and the F1 score and mIoU for each individual component were analyzed. The results demonstrate an accuracy of 0.85, an F1 score of 0.856, and an mIoU of 0.762, suggesting that the TILL architecture achieves an excellent balance between accuracy and computational complexity.

Critical strengths
  1. In this work, treating spectral channels and terrain channels differently is a sound design choice. Rather than handling all input channels uniformly, the architecture incorporates these domain-specific considerations through explicit band weighting and a process of terrain-conditioned refinement.

  2. This paper aims to evaluate both segmentation quality and computational cost, presenting an architectural evolution that spans several components. A comprehensive review of the experimental results can help identify which terrain-aware mechanisms are truly useful in real-world environments.

Critical weaknesses
  1. The quantitative evidence presented lacks the internal consistency required to support the core contributions claimed in the paper. Figure 3(a) shows an improvement in the F1 score from 0.712 to 0.856, while Figure 3(b) indicates an improvement from 0.621 to 0.783, achieved through modifications to several components. The text presents both instances as evidence for the same cumulative improvement. Furthermore, there is a clear discrepancy between the magnitude of improvement shown in the graphs and the explanations provided for them. Additionally, while Appendix B.4 and Table 1 specify 60 epochs and 24,180 training steps, the learning curves in Figure 5 extend to 100 epochs. Although it is possible that these represent different experimental runs or are merely illustrative graphs, the text provides no clear distinction or explanation. The lack of a consistent correspondence between the experimental setup/execution and the reported figures makes it difficult to assess the validity of the empirical evidence. This is not merely a matter of how results are presented; it concerns the fundamental soundness of the research.

  2. The description of the comparison protocol is insufficient to substantiate the "trade-off" claimed in the paper. Figure 2 compares models based on accuracy, whereas the analysis of key components utilizes F1 scores or mIoU as metrics. There is no unified benchmark table based on a consistent protocol for evaluating segmentation performance. Although Appendix A discusses the use of training and validation data, it lacks clear information regarding the composition and size of a held-out test set, used to derive the final results, as well as the thresholds established during the validation process. The model presented in Figure 2 does not merely offer a lower cost; it occupies a "non-dominated" position; a reasonable compromise, in terms of the trade-off between accuracy and cost. However, in the absence of F1 scores or mIoU results based on a common protocol, and given the lack of standardization regarding input data, training conditions, and cost-estimation criteria, the purported advantages for foreground segmentation remain uncertain. To clearly understand the achieved performance and its associated costs, a summary table is required, one that outlines the data-splitting strategy, training and evaluation specifications, and the segmentation performance versus cost for each methodology.

Questions for authors
  1. Figure 3(a) shows F1 scores ranging from 0.712 to 0.856, while Figure 3(b) shows a range of 0.621 to 0.783; furthermore, the component labels differ between the two. Could you clarify which specific runs and evaluation splits correspond to each configuration and increment, and ensure that the figures align with their accompanying descriptions? (Weakness 1)

  2. While Appendix B.4 and Table 1 specify 60 epochs, Figure 5 displays results up to 100 epochs. What training budget was used to generate the results in Figures 2–3? Additionally, are the results shown in Figure 5 from a separate experimental run, or are they merely illustrative? (Weakness 1)

  3. Although the training and validation processes are described in Appendix A, the composition of the dataset used for final evaluation remains unclear. Please provide the number of samples and the partitioning units for the training, validation, and test sets, and explain how the formal data splits were utilized. Furthermore, could you specify which evaluation dataset was used for each of the results presented in Figures 2–3? (Weakness 2)

  4. Figures 2–3 do not indicate the selection protocol employed. On what data were the checkpoints, component configurations, and mask thresholds selected? Additionally, were these selections finalized prior to the held-out evaluation? (Weakness 2)

  5. While Figure 2 compares accuracy, Figure 3 presents F1 and mIoU scores only for the TILL-variant models. Could you provide the foreground F1 scores and clearly defined mIoU values ​​for all baseline models under the same evaluation conditions? (Weakness 2)

  6. Figure 2 indicates that costs are not consistently low, but rather involve a trade-off. To allow for a proper comparison between quality and cost, could you specify the input channels, resolution, initialization method, training budget, and the criteria used for calculating FLOPs for each methodology? (Weakness 2)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.