the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 20 May 2026
Unraveling Spatial Dependencies in Landslide Susceptibility using Directed Acyclic Graphs
Abstract. Data-driven methods for landslide susceptibility assessment (LSA) often suffer from spurious correlations and “black-box” opacity, failing to capture the spatial dependency processes underlying landslide development. To address these limitations, we propose a directed acyclic graph (DAG)-informed interpretable framework by integrating structure-learning algorithms and graph attention models. This approach enables the identification of spatial dependency pathways and quantifies the propagation magnitudes (weights of connected links) of landslide conditioning factors. We applied this framework to the Ili River Basin, Xinjiang, China. A total of 14 robust spatial dependency chains were identified, and the dominant susceptibility-related chains were categorized into four types: (1) Elevation–climate-driven pathways (Elevation → Precipitation → NDWI → Landslide; Elevation → Precipitation → Temperature → Snow Depth → NDWI → Landslide); (2) Tectonic-controlled pathways (Distance to faults → PGA → Landslide); (3) Topographic dominated pathways (Slope → Curvature → Landslide); and (4) Hydrological driven pathways (Distance to rivers → NDWI → Landslide). Using a novel importance-weighted decoupling method, we generated pathway-specific susceptibility maps. These four chains account for 18.32%, 15.74%, 17.67%, and 16.76% of the high-susceptibility areas, respectively. These areas are predominantly clustered in mid–high mountainous, high-intensity seismic, and weakened lithological belt regions. Our proposed framework advances LSA from statistical prediction to dependency-informed explanation, providing decision-makers with a scientific basis for interpreting susceptibility variations across different spatial and environmental settings.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Natural Hazards and Earth System Sciences.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.- Preprint
(2359 KB) - Metadata XML
-
Supplement
(312 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-2637', Anonymous Referee #1, 22 May 2026
-
AC1: 'Reply on RC1', Qingkai Meng, 06 Jul 2026
Dear Reviewer,
Thank you for your valuable comments on our manuscript. Please find our responses below. We welcome any further guidance and discussion.
Comment 1: A major concern is the low quality and limited representativeness of the fundamental dataset. The landslide inventory contains only 1,198 landslides over a very large and geologically complex region, which is inadequate to support the proposed highly parameterized DAG-GNN framework.
Response: We respectfully clarify that the landslide inventory was constructed through a systematic multi-source approach: visual interpretation of multi-temporal high-resolution optical remote sensing imagery (Google Earth, Gaofen-1/2) covering 1990–2023, combined with field verification. In total, 1,832 geohazard points were mapped across the basin, comprising 1,198 landslides, 144 rockfalls/collapses, and 490 debris flows, of which approximately 36% of the interpreted landslide locations were validated through field surveys [432 landslides]. This inventory has been reviewed and archived on the official data-sharing platform of the Third Xinjiang Scientific Expedition Program (the archived subset covers 1990–2005; https://cstr.cn/33110.11.XIEG.XJSEDATA.2024.00100302), involving independent expert review prior to release.
Regarding whether the sample number is sufficient to support the proposed framework, we highlight that our manuscript already includes a systematic data-ablation experiment (Section 5.1, Figure 10b), in which the training sample proportion was progressively reduced from 80% to 20%. Even under the most data-scarce condition, all evaluation metrics remained above 0.7, and the inferred DAG structures remained largely stable in terms of edge count and retained connections (Figure 11). This demonstrates that the current inventory size is sufficient to support stable structure-learning results within this basin. Our identified high-susceptibility patterns in Tian, china are broadly consistent with regional-scale tectonic and topographic controls documented across the wider Tien Shan orogenic system, Kyrgze and Tajikistan (Rosi et al., 2023), though a direct quantitative comparison was beyond the scope of this study due to differing study extents and input datasets.We agree, nonetheless, that supplementing the inventory with additional fine-scale landslide mapping in future work would further strengthen the robustness of the data-driven components of the framework.
Comment 2: Many environmental factors are derived from coarse-resolution or secondary datasets resampled to 30 m resolution, yet the impacts of spatial uncertainty and scale effects are not rigorously evaluated.
Response: We thank the reviewer for raising this issue. We clarify the native resolution of each factor group: (1) climatic factors (precipitation, temperature, snow depth) are derived from the ERA5 reanalysis product at approximately 0.1° (~9–11 km); (2) geological and tectonic factors (engineering geological lithology, distance to faults) are compiled from the Natural Resources Archives of the Xinjiang Uygur Autonomous Region at a nominal scale of 1:200,000; (3) PGA is derived from the fifth national geohazard risk census at a nominal scale of 1:100,000; and (4) the remaining 10 topographic, hydrological, and land-cover factors are derived directly from 30 m DEM or 30 m/10 m remote sensing products. All layers were harmonized to a common 30 m grid using bilinear resampling for computational consistency across this multi-source dataset.
We acknowledge that this resolution mismatch introduces a scale-effect concern, however, climatic reanalysis products and geological archival maps rarely reach sub-100 m resolution even in recent regional-scale LSA studies, making this constraint largely unavoidable given current data availability. In the revision, we will conduct a multi-resolution sensitivity analysis by re-running the structure-learning pipeline at a coarser common resolution and report the stability of retained edges and pathway weights as a new supplementary analysis.
Comment 3: The proposed causal interpretations appear largely speculative and are not adequately supported by physical evidence or independent validation. Several "dependency pathways" merely reflect common statistical associations rather than revealing genuinely new scientific mechanisms. The DAG structure is strongly dependent on expert intervention and manual correction, which substantially weakens the objectivity and reproducibility of the framework.
Response: We address the three aspects raised in turn.
- On the speculative nature of causal interpretations.We agree that the directed dependencies inferred by structure-learning algorithms represent statistical dependency structures consistent with, but not proof of, causal mechanisms, and we will consistently use "spatial dependency" rather than "causal mechanism" throughout the revised manuscript. We note that several reported pathways are already supported by independent evidence discussed in the manuscript: the spatial congruence between the seismic-dynamic chain (DF→PGA→LS) and the documented rupture zones of the 1812 Nilka Mw 8.0 and 2003 Zhaosu Mw 6.1 earthquakes (Havenstrite et al., 2015), and the consistency between the snowmelt-driven chain and independent laboratory findings on freeze–thaw degradation of Ili loess (Li et al., 2023; Meng et al.,2026). We will frame the DAG-derived pathways explicitly as a hypothesis-generating tool requiring future InSAR deformation monitoring or field-based geophysical validation, added as a concrete future-work direction.
- On expert intervention weakening objectivity and reproducibility.The "expert-guided refinement" step is not an ad hoc subjective correction, but consists of three explicit, rule-based constraints applied uniformly across all candidate edges: (i) an end-node constraint requiring all paths to terminate at "Landslide"; (ii) a fixed weight threshold (smoothed weight < 0.30) for pruning unreliable connections; and (iii) a logical-direction constraint requiring dependency chains to proceed from regional topographic/structural background factors, through climatic and hydrological mediating factors, to surface processes, and ultimately to landslide occurrence — not the reverse. This ordering is grounded in well-established literature (Gong et al.,2018) rather than individual subjective judgment, and is applied deterministically, making the F-DAG fully reproducible given the same I-DAG input. We agree the manuscript did not adequately report sensitivity to these design choices; in the revision we will add a sensitivity analysis varying the pruning threshold (e.g., 0.25 and 0.35), and we plan to release the model code and refinement rule set publicly upon publication.
- On pathways reflecting common statistical associations rather than new mechanisms.We acknowledge that several elevation–climate–hydrology chains (Chains 1–10) are consistent with well-established geomorphological understanding. We would argue, however, that quantitatively confirming and spatially ranking these mechanisms at the basin scale, with explicit directionality and pathway-specific susceptibility mapping, still provides operational value for targeted mitigation (e.g., prioritizing hydrological monitoring versus slope-stabilization engineering in different sub-regions). More importantly, we highlight Chain 12 (DF→EG→LU→NDVI→NDWI→LS) as a genuinely non-obvious finding: it reveals that faults influence landslide susceptibility not only through direct ground shaking, but also indirectly by shaping long-term geomorphic units that guide land-use decisions, which in turn regulate vegetation cover and soil moisture. In the revision, we will restructure the Discussion to explicitly distinguish these confirmatory pathways from this counterintuitive pathway, which we present as the primary novel mechanistic contribution of this study.
Comment 4: The scientific novelty is overstated because similar graph-based or explainable AI approaches have already been widely explored in recent years. The manuscript does not convincingly demonstrate a substantial methodological breakthrough.
Response: We agree that both post-hoc explainable AI (e.g., SHAP, LIME) and graph-based approaches have been explored in the LSA literature, and clarify the distinctions between our framework and these paradigms. Post-hoc XAI methods rank the statistical contribution of each factor but do not capture directionality among factors, making them susceptible to confounding and spurious-correlation artifacts. Existing graph-based LSA approaches typically construct graphs over spatial units (grid cells or slope units) connected by spatial adjacency, aiming primarily to improve spatial feature representation and predictive accuracy [Zeng et al., 2022; Xia et al., 2024; Zhang et al., 2024].
In contrast, the graph in our framework is constructed over environmental factors themselves rather than spatial units, with directed edges representing dependency propagation among factors. This is achieved through a three-stage pipeline that, to our knowledge, has not been jointly applied in LSA: (i) comparative structure learning across four algorithms to select the most robust initial dependency graph; (ii) knowledge-guided rule-based refinement to enforce geologically plausible directionality; and (iii) graph attention reweighting to refine local dependency strengths. Beyond structure discovery, our framework further introduces an importance-weighted pathway decomposition method that allows practitioners to attribute high-susceptibility zones to specific mechanistic pathways (e.g., seismic-driven vs. snowmelt-driven), which is, to our knowledge, rarely offered by existing graph-based or post-hoc XAI approaches for regional LSA. We will revise the Introduction and Discussion to more explicitly articulate these distinctions, and would welcome the reviewer's guidance on any specific prior work offering directly comparable pathway-level, direction-aware susceptibility decomposition.
Comment 5: The expression and presentation quality are also poor... the discussion section remains descriptive and lacks deep scientific analysis regarding uncertainty, transferability, and physical implications.
Response: We will address them in revision documents to improve the article structure and figure presentation and enhance the clarity of our work.
Reference
Rosi, A., Frodella, W., Nocentini, N., Caleca, F., Havenith, H.B., Strom, A., Saidov, M., Bimurzaev, G.A., Tofani, V., 2023. Comprehensive landslide susceptibility map of Central Asia. Nat. Hazards Earth Syst. Sci. 23, 2229–2250. https://doi.org/10.5194/nhess-23-2229-2023
Meng, Q.K., 2022. Geological hazard survey data of the Ili River Basin at 1:250,000 scale (1990–2005). Institute of Mountain Hazards and Environment, Chinese Academy of Sciences and Ministry of Water Resources [creator]. Xinjiang Third Scientific Expedition Data Sharing Service Platform [distributor], 2024-10-10. https://cstr.cn/33110.11.XIEG.XJSEDATA.2024.00100302 (in Chinese).
Li, Y., Yang, G., Ye, W., et al., 2023. Deterioration law and microscopic mechanism of hydraulic characteristics of undisturbed loess in Ili under freeze-thaw cycles. Eng. Geol. 41, 1234–1245. https://doi.org/10.12401/j.nwg.20220730 (in Chinese).
Meng, Y., Meng, Q.K., Wu, H., et al., 2026. Effects of freeze–thaw action on the pore structure and shear characteristics of Ili loess. Arid Zone Res. 1–16. https://link.cnki.net/urlid/65.1095.X.20260627.1608.004 (in Chinese).
Havenstrite, R., et al., 2015. Tien Shan Geohazards Database: earthquakes and landslides. Geomorphology 249, 104–121. https://doi.org/10.1016/j.geomorph.2015.03.020
Zeng, H., Zhu, Q., Ding, Y., Hu, H., Chen, L., et al., 2022. Graph neural networks with constraints of environmental consistency for landslide susceptibility evaluation. Int. J. Geogr. Inf. Sci. 36, 1–24. https://doi.org/10.1080/13658816.2022.2103819
Xia, D., Tang, H., Glade, T., et al., 2024. KNN-GCN: a deep learning approach for slope-unit-based landslide susceptibility mapping incorporating spatial correlations. Math. Geosci. 56, 1011–1039.
Zhang, Q., He, Y., Zhang, L., Lu, J., Gao, B., Yang, W., Chen, H., Zhang, Y., 2024. A landslide susceptibility assessment method considering the similarity of geographic environments based on graph neural network. Gondwana Res. 132, 323–342. https://doi.org/10.1016/j.gr.2024.04.013
Gong, X.P., Wang, Z.G., Ma, H.B., et al., 2018. Study on causes and evaluation of geological hazards in Yili Valley, Xinjiang. Geological Publishing House, Beijing (in Chinese).
Citation: https://doi.org/10.5194/egusphere-2026-2637-AC1 -
RC2: 'Reply on AC1', Anonymous Referee #1, 08 Jul 2026
After carefully considering the authors’ response and the current manuscript, I do not think the major concerns have been adequately addressed. The response mainly provides explanations or promises for future revision, but the fundamental weaknesses remain. The landslide inventory is still limited for supporting a highly parameterized DAG-GNN framework over a large and geologically complex basin, and the completeness, temporal consistency, mapping uncertainty, and independent validation of the inventory remain insufficiently demonstrated. The scale mismatch among the input factors is also a serious concern: several key climatic, geological, and seismic variables are derived from coarse-resolution or secondary datasets and then resampled to 30 m, but the effects of spatial uncertainty and scale dependency have not been rigorously assessed. The proposed “dependency pathways” remain largely statistical and hypothesis-generating rather than physically validated mechanisms. The final DAG still depends substantially on expert-guided correction, which weakens the claimed objectivity and reproducibility. In addition, the novelty is overstated, as graph-based and explainable AI methods have already been widely explored in landslide susceptibility studies. The discussion remains descriptive, and the manuscript still contains many problems in logic, expression, figure readability, and overinterpretation. Overall, the manuscript does not provide sufficient data reliability, methodological advance, or mechanism validation to meet the publication standard. Therefore, I do not recommend publication.
Citation: https://doi.org/10.5194/egusphere-2026-2637-RC2
-
AC1: 'Reply on RC1', Qingkai Meng, 06 Jul 2026
-
CC1: 'Comment on egusphere-2026-2637', J.Q. Zhuang, 15 Jul 2026
This study proposes a directed acyclic graph (DAG)-informed interpretable framework that integrates structure-learning algorithms, expert-knowledge constraints, graph attention networks, and pathway-specific landslide susceptibility mapping. The topic is timely, and the attempt to move beyond conventional post-hoc feature attribution toward explicit directed dependency structures is potentially valuable. The overall methodological workflow is logically organized, and the results provide useful insights into how different environmental factors may jointly contribute to the spatial distribution of landslide susceptibility. i think the following aspects could be further improved before accepted.
- Landslide inventory issue. It is recommended that additional information be provided, including landslide type, occurrence time and triggering factors. In that case, you can using DAG method to explore direct or indirect evidence relationship in each types of landslides. If sufficiently reliable type- or trigger-specific information is available, it would be valuable to examine whether the inferred dependency structures differ among rainfall-induced, snowmelt-related, earthquake-induced, or other landslide categories. Such an analysis could provide stronger evidence for the proposed direct and indirect dependency pathways.
- The rationale for selecting variables such as NDVI, NDWI, annual and monthly mean temperature, precipitation, and snow depth should be further strengthened. For example, why NDVI represents root-reinforcement/interception, why NDWI proxies antecedent soil saturation, why snow depth matters for freeze-thaw, that will improve physical justification.
- We found you have partially addressed already via the ablation and data-sparsity experiments (Figures 10–11), further discussion is recommended regarding the feasibility, robustness, and transferability of the model, including the stability and reliability of identified spatial dependency under different data partitions and parameter settings.
4 Random splitting on spatially autocorrelated environmental data will inflate reported accuracy/AUC, since train and test pixels can be near-neighbors. It is recommended that spatial cross-validation be incorporated to reduce spatial dependence between training and testing samples and to evaluate model performance across multiple independent training-testing splits. This would provide a more objective assessment of the models generalization ability.
5 Derivation and validation of pathway-specific susceptibility maps.Extracting dominant dependency pathways and generating pathway-specific susceptibility maps are among the most distinctive contributions of this study. Please add explain context about the construction and interpretation of these maps.
Citation: https://doi.org/10.5194/egusphere-2026-2637-CC1 -
RC3: 'Comment on egusphere-2026-2637', Anonymous Referee #2, 15 Jul 2026
This study proposes a directed acyclic graph (DAG)-informed interpretable framework that integrates structure-learning algorithms, expert-knowledge constraints, graph attention networks, and pathway-specific landslide susceptibility mapping. The topic is timely, and the attempt to move beyond conventional post-hoc feature attribution toward explicit directed dependency structures is potentially valuable. The overall methodological workflow is logically organized, and the results provide useful insights into how different environmental factors may jointly contribute to the spatial distribution of landslide susceptibility. i think the following aspects could be further improved before accepted.
- Landslide inventory issue. It is recommended that additional information be provided, including landslide type, occurrence time and triggering factors. In that case, you can using DAG method to explore direct or indirect evidence relationship in each types of landslides. If sufficiently reliable type- or trigger-specific information is available, it would be valuable to examine whether the inferred dependency structures differ among rainfall-induced, snowmelt-related, earthquake-induced, or other landslide categories. Such an analysis could provide stronger evidence for the proposed direct and indirect dependency pathways.
- The rationale for selecting variables such as NDVI, NDWI, annual and monthly mean temperature, precipitation, and snow depth should be further strengthened. For example, why NDVI represents root-reinforcement/interception, why NDWI proxies antecedent soil saturation, why snow depth matters for freeze-thaw, that will improve physical justification.
- We found you have partially addressed already via the ablation and data-sparsity experiments (Figures 10–11), further discussion is recommended regarding the feasibility, robustness, and transferability of the model, including the stability and reliability of identified spatial dependency under different data partitions and parameter settings.
4 Random splitting on spatially autocorrelated environmental data will inflate reported accuracy/AUC, since train and test pixels can be near-neighbors. It is recommended that spatial cross-validation be incorporated to reduce spatial dependence between training and testing samples and to evaluate model performance across multiple independent training-testing splits. This would provide a more objective assessment of the models generalization ability.
5 Derivation and validation of pathway-specific susceptibility maps.Extracting dominant dependency pathways and generating pathway-specific susceptibility maps are among the most distinctive contributions of this study. Please add explain context about the construction and interpretation of these maps.
Citation: https://doi.org/10.5194/egusphere-2026-2637-RC3
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 147 | 65 | 20 | 232 | 35 | 11 | 9 |
- HTML: 147
- PDF: 65
- XML: 20
- Total: 232
- Supplement: 35
- BibTeX: 11
- EndNote: 9
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript presents a DAG-guided explainable framework for landslide susceptibility assessment; however, the overall scientific quality and practical contribution remain insufficient for publication in its current form. A major concern is the low quality and limited representativeness of the fundamental dataset. The landslide inventory contains only 1,198 landslides over a very large and geologically complex region, which is inadequate to support the proposed highly parameterized DAG-GNN framework and the claimed discovery of “robust dependency pathways.” The manuscript lacks sufficient information regarding inventory completeness, temporal consistency, mapping uncertainty, validation strategy, and the balance between landslide and non-landslide samples. Many environmental factors are derived from coarse-resolution or secondary datasets resampled to 30 m resolution, yet the impacts of spatial uncertainty and scale effects are not rigorously evaluated. In addition, the proposed causal interpretations appear largely speculative and are not adequately supported by physical evidence or independent validation. Several “dependency pathways” merely reflect common statistical associations among topography, precipitation, and hydrological indicators rather than revealing genuinely new scientific mechanisms. The manuscript repeatedly emphasizes explainability and causal inference, but the DAG structure is strongly dependent on expert intervention and manual correction, which substantially weakens the objectivity and reproducibility of the framework. Moreover, the scientific novelty is overstated because similar graph-based or explainable AI approaches for landslide susceptibility assessment have already been widely explored in recent years. The manuscript does not convincingly demonstrate a substantial methodological breakthrough or significant improvement over existing approaches. The expression and presentation quality are also poor. The manuscript contains numerous grammatical problems, unclear logical transitions, repetitive descriptions, and overinterpretation of results. Several figures are visually cluttered and difficult to interpret, while the discussion section remains descriptive and lacks deep scientific analysis regarding uncertainty, transferability, and physical implications. Overall, the manuscript suffers from insufficient data reliability, limited scientific advancement, weak mechanism validation, and poor presentation quality. Therefore, I do not recommend publication, and rejection is suggested.