{"id":"fd187bae-44dc-4d71-9da5-ea6a2bc76fbc","arxiv_id":"2507.08742","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A pixel-based Bayesian model of Gorkha earthquake landslides finds that channel steepness predicts landslide location but not size.","lead":"This paper builds statistical maps of where earthquake-triggered landslides are likely to occur, using the steepness of river channels as a new terrain predictor. It reports that steeper channels mark more landslide-prone slopes, but do not predict larger landslides.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central ksn effect may be a spatial-confounding artifact: the covariate-only model omits spatial random effects and standard DEM covariates, so the claimed association is not yet identified.","rationale":"The reader's weakest assumption highlights the reliance on a complete covariate set and the untested transferability, which is closely related to my concern. However, I focus specifically on the identification of the ksn effect on landslide susceptibility rather than on the separate transferability promise. The self-acknowledged risk in Section 4.3.4 that covariates may absorb spatially structured variability strengthens the concern. A straightforward sensitivity analysis with a spatial random effect or additional DEM covariates would settle whether the ksn association is robust. Because the paper is otherwise careful (proper scoring rules, concavity sensitivity analysis, transparent limitations in Section 5.1), the appropriate verdict remains conditional on this additional check, matching the reader's CONDITIONAL verdict.","tokens_in":25840,"tokens_out":5753,"duration_ms":69332,"concrete_test":"Refit model fit6a (and fit1a as a check) with an additional Matérn SPDE spatial random effect on the same mesh, or with extra fixed covariates such as slope, elevation, aspect, TWI, and distance to faults. Compare the posterior mean and 95% credible interval of the log-ksn effect against the original covariate-only fit. If the effect attenuates by more than 50% or its credible interval includes zero, the central claim is not supported; if the effect remains stable, the confounding concern is mitigated. Report the resulting ksn effect and the spatial random effect's marginal variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that elevated ksn is strongly associated with landslide susceptibility depends on the covariate-only model in Section 3.3, which deliberately excludes spatial random effects. That design is valid only if all spatially relevant landslide controls are captured by PGA, ksn, distance-to-channel metrics, land cover, and geology. This assumption is not tested: the model omits standard DEM covariates (slope, aspect, curvature, TWI, distance to faults) that are correlated with ksn because both are derived from the same DEM (Section 2.2.1). If any omitted variable is spatially correlated with both ksn and landslide intensity, the positive RW2 effect of log ksn in Figure 13a is biased. The internal grid-based cross-validation (Section 3.4) cannot detect this bias, since a spatially structured omitted variable would still yield good within-region predictions. The paper itself acknowledges in Section 4.3.4 that 'in the absence of an explicit spatial random field, spatially structured variability may be absorbed by the covariates themselves, potentially leading to spurious or overfitted patterns.' The manual ksn filling in hanging valleys (Section 2.2.2) is a second, more localized source of bias, but the confounding concern affects the entire study area and directly threatens the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Suen et al. present a pixel-based Bayesian spatial model for earthquake-induced landslides triggered by the 2015 Gorkha earthquake in Nepal. Landslide centroids are modelled as a Poisson point process and log-transformed landslide areas as a Gaussian mark, with fixed covariates including peak ground acceleration (PGA), the normalised channel steepness index (k_sn), relief and flow distance to channels, land cover, and geology. The models are fitted with the inlabru framework to handle spatial misalignment between raster covariates and point observations. The authors deliberately omit spatial random effects to favour transferability, compare six model pairs via proper scoring rules under thinning and grid-based cross-validation, and conclude that elevated k_sn is strongly associated with increased landslide susceptibility but not with landslide magnitude. They also include a concavity sensitivity analysis and uncertainty maps for the best-fitting models.","tokens_in":26099,"tokens_out":9420,"duration_ms":102821,"significance":"The paper's strengths are methodological: it applies strictly proper scoring rules (LS and CRPS) with both thinning and grid-based cross-validation, uses inlabru for coherent mesh-based disaggregation of misaligned data, and is transparent about many limitations. If the k_sn association is real, the framework would be a useful, transferable alternative to slope-unit-based INLA landslide models, and the distinction between controls on landslide location and size would be a valuable geomorphic finding. However, the central claim is not yet fully identified: the covariate-only model omits standard DEM covariates that are correlated with k_sn because both are derived from the same DEM, the size-model conclusion is based on removing k_sn from the final model rather than testing it directly, and the manual k_sn filling in hanging valleys is not subjected to sensitivity analysis. These issues are fixable and do not undermine the value of the modelling framework itself.","major_comments":[{"comment":"The linear predictor in Eq. (3) contains only fixed covariates, and Section 4.3.4 explicitly acknowledges that without a spatial random field, spatially structured variability may be absorbed by covariates and produce spurious or overfitted patterns. This is a direct threat to the central k_sn claim because k_sn is derived from the same DEM as slope, aspect, curvature, TWI, and distance to streams (Section 2.2.1, Table 1), and the positive rw2(log k_sn) effect in Figure 13a may be an artefact of an omitted topographic control. The grid-based cross-validation in Section 3.4 cannot detect this bias, since a spatially structured omitted variable would still improve within-region interpolation. Please add a robustness model that includes standard DEM covariates or a spatial random effect, or validate on an external holdout region, and report how the k_sn effect changes.","section":"Section 3.3, Eq. (3) and Section 4.3.4"},{"comment":"The claim that k_sn is not associated with landslide size is not directly tested. In Section 3.3, k_sn is removed from fit5b and fit6b because 'initial findings indicated that its inclusion did not enhance predictive performance for landslide size', but no model is reported that includes k_sn together with Fd2Ch or Rf2Ch, and the score differences among size models in Table 6 are very small (for example, LS 2.4616 vs 2.4620 for fit6b vs fit5b in Thinning Set A). The second half of the abstract therefore needs a direct comparison, ideally with posterior intervals for the k_sn coefficient in a size model that also includes the distance-to-channel covariate.","section":"Section 3.3, Table 3 and Table 6"},{"comment":"The manual replacement of k_sn values in hanging valleys with values from nearby geomorphologically similar regions is a subjective step without sensitivity analysis. Section 2.2.2 itself states that the standard algorithm systematically misestimates local k_sn in these areas, so the estimated rw2(log k_sn) effect in Figure 13a could depend on the filling rule. Please report a sensitivity analysis (for example, excluding glacial pixels, varying the substitution neighbourhood, or treating k_sn as missing in those areas) to show that the association is not an artefact of the manual filling.","section":"Section 2.2.2 and Appendix B.1"},{"comment":"The exclusion of the Siwalik Group, Gangetic Plain, and Sub-Himalayan regions (Section 2.3) is justified by numerical stability, but it restricts the study domain to the high-relief Himalaya. The paper's transferability claim (Section 3.3 and Abstract) should therefore be qualified as applying to similar mountain terrain, and a sensitivity test of the exclusion boundary would help establish how much the results depend on this subjective choice.","section":"Section 2.3"}],"minor_comments":[{"comment":"The row entries for fit5b and fit6b are ambiguous; please clarify which covariates enter each model, in particular whether PGA is included in the size models.","section":"Table 3"},{"comment":"The squared error (SE) is proper but not strictly proper; the text should avoid presenting SE and DS as equally principled as the strictly proper LS and CRPS rules.","section":"Section 3.4.1"},{"comment":"The reported LS differences between fit6b and fit5b are extremely small; add Monte Carlo standard errors or a formal comparison to support the claim that fit6b is clearly the best size model.","section":"Table 6"},{"comment":"The statement that aftershock PGA 'did not improve the results' is unsupported; please provide a summary of the comparison or a reference to an appendix containing the numbers.","section":"Section 2.2.3"},{"comment":"The concavity sensitivity analysis is described only qualitatively; report a quantitative summary (for example, correlations between k_sn maps for different concavity values) to substantiate the claim that the results are unchanged.","section":"Section 4.1.2 and Appendix D.2"},{"comment":"The text refers to 'Appendix 20' where 'Appendix D.2' is intended, and the reference to 'Gorhka' should read 'Gorkha'; please proofread the bibliography and cross-references.","section":"Minor language issues"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the spatial-confounding concern raised in the stress-test note is, in my reading, a genuine identification problem rather than a preference for a different modelling school. The authors are explicit about their transferability motivation, but the covariate-only design plus in-sample variable selection leaves the central k_sn claim insufficiently protected. I do not see grounds for rejection, because the methodological framework is sound and the requested robustness checks are implementable within the manuscript's scope. The paper fits the journal's scope; the novelty is incremental but appropriate for stat.AP."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a well-executed applied case study, but the central claim about ksn is not yet identified. The authors deliberately drop spatial random effects for transferability, then interpret the ksn effect as causal. Their own Section 4.3.4 concedes that without an explicit spatial field, spatially structured variability may be absorbed by covariates and produce spurious patterns. That is exactly the soft spot.\n\nWhat the paper does well: using ksn as a pixel-based covariate in an inlabru point process model is a legitimate new application. The mesh-based disaggregation for misaligned data is sensible, and the evaluation with strictly proper scoring rules plus two cross-validation schemes is a real step up from WAIC/DIC summaries. The concavity sensitivity analysis is thorough, and the empirical suggestion that ksn predicts location but not size is a useful process insight if it holds.\n\nWhere it gets soft, in proportion. First and most important: ksn is derived from the DEM, as are slope, aspect, curvature, and TWI. The models include PGA, ksn, distance-to-channel, land cover, and geology, but not slope or the other standard DEM derivatives. If any omitted DEM variable is spatially correlated with both ksn and landslide intensity, the positive RW2 effect of log ksn in Figure 13a is biased. The grid-based CV cannot detect this: a spatially structured omitted variable still predicts well within the same region. So the abstract's \"strongly associated\" is not yet supported. This is not a minor caveat; it is the main result.\n\nSecond, the manual replacement of ksn in hanging valleys is subjective and could bias the local effect, though this is more localized. Third, ksn was dropped from the size models based on in-sample \"initial findings,\" which flatters the \"not size\" conclusion; the CV provides some check, but not a clean one. Fourth, the transferability promise is only tested with a within-region chequerboard split, not an external holdout region.\n\nThe authors are honest about limitations, and the analysis is careful enough that this deserves a serious referee. But the paper should not be accepted as is. The central claim needs to be tested against a model that includes slope and other DEM covariates as controls, or that fits a spatial field and compares the ksn effect, or both.\n\nWho gets value from this: landslide susceptibility modellers and geomorphologists working with earthquake inventories. It deserves peer review, but revision should be major. I would not cite it in its current form.","headline":"The ksn-location association is plausible but not identified: the covariate-only design omits spatial random effects and standard DEM covariates, and the paper's own Section 4.3.4 admits the consequence.","tokens_in":26597,"tokens_out":1699,"would_cite":false,"duration_ms":22324,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P12","62M30","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Elevated river-channel steepness is strongly tied to where the 2015 Gorkha earthquake triggered landslides, but not to how large those landslides were.","keywords":["earthquake-induced landslides","Gorkha earthquake 2015","channel steepness index ksn","landslide susceptibility modelling","spatial point process","inlabru / INLA","spatially misaligned data","proper scoring rules"],"falsifier":"Re-run the best-fitting model with $k_{sn}$ recomputed at concavities 0.4 and 0.6 and without the manual hanging-valley substitutions; if the $k_{sn}$ credible interval shifts to include zero or flips sign under any of these recomputations, the claimed link between channel steepness and landslide susceptibility is not robust. As a second check, apply the fitted covariate-only model to a different earthquake's landslide inventory and compare its out-of-region predictive scores with those of a version that includes an explicit spatial random field: if the spatial-field version wins out-of-region, the transferability premise fails.","tokens_in":2243,"feed_emoji":"🏔","tokens_out":3739,"duration_ms":136496,"temperature":0.7,"pith_summary":"This paper seeks to establish that the normalized channel steepness index ($k_{sn}$), a river-network metric that proxies how actively streams are incising the landscape, is a physically meaningful covariate for earthquake-induced landslide susceptibility. Using the 2015 Gorkha earthquake in Nepal as the test bed, the authors build pixel-based Bayesian spatial point process models of landslide centroids plus separate Gaussian models of log-transformed landslide area, linking them through inlabru's mesh-based disaggregation to overcome misalignment between pixel covariates and point observations. The central result is that elevated $k_{sn}$ is strongly associated with a higher probability of landslide initiation, while landslide magnitude is instead associated with flow distance to the channel. A sympathetic reader would care because the framework deliberately excludes spatial random effects, aiming for a covariate-only model that could transfer to other earthquakes in complex terrain where only DEMs and shaking maps are available.","feed_headline":"River steepness predicts quake landslide sites but not sizes","feed_subtitle":"Nepal's 2015 Gorkha quake: channel incision sets landslide likelihood; flow distance sets landslide size.","key_machinery":"The load-bearing object is the normalized channel steepness index $k_{sn}$, derived from the stream-power incision law $S = k_{sn} A^{-\\theta}$, which expresses local channel slope as a function of drainage area and thereby isolates the steepness that reflects erosion and incision processes. Around it the paper assembles: (i) a Poisson point process for landslide centroids with intensity modelled in log space; (ii) a Gaussian model for log-transformed landslide area; (iii) DEM-derived distance metrics (relief and flow distance to channel) computed with LSDTopoTools; and (iv) inlabru's mesh-based disaggregation, which resolves spatial misalignment between raster covariates and point observations by projecting both onto a common triangular mesh. The deliberate exclusion of spatially structured random effects (Matérn SPDE fields) is itself part of the machinery: the authors argue that including such fields would inflate posterior uncertainty, bias fixed effects through spatial confounding, and undermine the stated goal of transferring fitted covariate relationships to unseen regions. Model comparison runs on four cross-validation splits (two random thinnings and two 3 km by 3 km checkerboard grids) scored with strictly proper scoring rules, primarily the logarithmic score and the continuous ranked probability score.","core_discovery":"The paper's central claim is that the normalized channel steepness index $k_{sn}$ — the steepness of a river channel normalized by drainage area under a stream-power incision law with concavity fixed at 0.5 — carries a strong and distinct signal for where earthquake-induced landslides initiate, and no signal for how large they become. In a six-model comparison for the Gorkha earthquake, the centroid model (fit6a) combining a smoothed $k_{sn}$ effect, peak ground acceleration, an exponentially decaying flow-distance-to-channel term, land cover, and geology scores best under strictly proper scoring rules (CRPS and RMSE), and the posterior shows elevated $k_{sn}$ clearly increasing susceptibility above roughly $\\log k_{sn} = 2.5$, with a flattening plateau at the steepest end. For landslide magnitude, models with $k_{sn}$ are outperformed by fit6b, which uses flow distance to the channel: the longer the flow path to the channel, the larger the landslide. The paper interprets this as $k_{sn}$ modulating triggering probability while flow distance governs failure volume, and it frames the whole approach — pixel-based covariates instead of slope-unit aggregation, and no spatial random field — as a deliberately transferable alternative to existing INLA-based landslide models built on slope-unit polygons.","pith_inferences":["Editorial inference: the decoupling of initiation control ($k_{sn}$) from size control (flow distance) points to two distinct physical mechanisms — fluvial incision setting a triggering threshold on adjacent hillslopes, while runout and contributing flow path set volume — which implies that risk planning needs two separate map products, occurrence probability and expected size.","Editorial inference: the paper validates transferability only inside the Gorkha region; the natural next test is to refit on one earthquake and score on another (for example, a different large mountain earthquake with a mapped inventory), which would directly quantify how much of the covariate relationship is universal rather than regional.","Editorial inference: the unexplained bump in the smoothed PGA effect near 0.42 g is the paper's own hint that residual spatial structure may have leaked into the covariate curves; re-running with an orthogonalized spatial effect would test whether the $k_{sn}$ coefficient is inflated by unmeasured spatial processes.","Editorial inference: because the response uses landslide centroids rather than crowns or full polygons, the $k_{sn}$–susceptibility association could partly reflect where centroids fall inside mapped polygons; re-estimating with crown-based points would check whether the association is an artifact of centroid placement."],"forward_implications":["Earthquake response susceptibility maps can be produced from DEM and shaking data alone: channel steepness does the geomorphic work that slope-unit geometry previously did, so no per-region parameterization of slope units is needed.","Hazard assessments should separate location from magnitude, because where landslides start is governed by fluvial incision intensity ($k_{sn}$) while how large they grow tracks flow distance to the channel.","The covariate-only design means the fitted relationships can in principle be retrained for any mountain region with a DEM, a channel network, a land-cover map, and a PGA raster, without waiting for a new landslide inventory.","Traditional DEM covariates that are nonlinear and mutually correlated (slope, elevation, curvature) can be replaced by a smaller set of process-based metrics such as $k_{sn}$ and flow distance, simplifying models and improving interpretability.","Per-grid-cell strictly proper scoring rules expose where a model fails locally, such as monsoon-preconditioned slopes, a diagnostic that aggregated mean scores like WAIC or DIC would conceal."],"supporting_citations":[{"why":"Supplies the response dataset: the Gorkha earthquake landslide inventory of polygons whose centroids and areas define the point pattern and marks.","marker":"Valagussa et al., 2022"},{"why":"Primary source inventory of 24,915 landslides integrated into the catalogue the modelling is built on.","marker":"Roback et al., 2018"},{"why":"LSDTopoTools software used to compute $k_{sn}$, relief to channel, and flow distance to channel from the DEM.","marker":"Mudd et al., 2022"},{"why":"The inlabru framework that enables the mesh-based disaggregation of spatially misaligned covariates and point observations.","marker":"Lindgren et al., 2024"},{"why":"The coherent disaggregation method that justifies the mesh resolution and the alignment of pixel covariates with point-response data.","marker":"Suen et al., 2025"},{"why":"Hilltop flow routing algorithm that produces the relief-to-channel and flow-distance metrics used as covariates.","marker":"Grieve et al., 2016"},{"why":"Stream-power incision model linking channel steepness to erosion, the physical basis for using $k_{sn}$ as a landslide covariate.","marker":"Whipple and Tucker, 1999"},{"why":"Establishes strictly proper scoring rules (logarithmic score and CRPS) used to compare predictive performance across models.","marker":"Gneiting and Raftery, 2007"},{"why":"ShakeMap Atlas source of the peak ground acceleration raster used as the seismic trigger covariate.","marker":"Marano et al., 2024"}],"fun_headline_variants":["River steepness flags quake landslide sites, not sizes","Quake landslide location set by river steepness; size by flow distance","Channel steepness controls where quake landslides start, not how big","Gorkha quake: steep channels raise landslide risk, not magnitude","For landslides, steep rivers set the trigger; flow path sets the size"],"cache_read_input_tokens":28800,"weakest_assumption_plain":"The model's transferability premise is that its five covariates — PGA, $k_{sn}$, flow distance to channel, land cover, and geology — leave no spatially structured control unmodelled, so dropping the spatial random field cannot bias the covariate effects; that premise is tested only by a checkerboard split inside the Gorkha region rather than against an external region, and it also depends on the manually substituted $k_{sn}$ values in glacial hanging valleys being accurate.","fun_headline_variants_meta":{"raw":{"variants":["River steepness flags quake landslide sites, not sizes","Quake landslide location set by river steepness; size by flow distance","Channel steepness controls where quake landslides start, not how big","Gorkha quake: steep channels raise landslide risk, not magnitude","For landslides, steep rivers set the trigger; flow path sets the size"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000435,"raw_usage":{"total_tokens":2247,"prompt_tokens":1012,"completion_tokens":1235,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":1143}},"tokens_in":628,"tokens_out":1235,"duration_ms":11948,"temperature":1.0,"reasoning_tokens":1143,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:10:00.170571+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the best-fitting model with $k_{sn}$ recomputed at concavities 0.4 and 0.6 and without the manual hanging-valley substitutions; if the $k_{sn}$ credible interval shifts to include zero or flips sign under any of these recomputations, the claimed link between channel steepness and landslide susceptibility is not robust. As a second check, apply the fitted covariate-only model to a different earthquake's landslide inventory and compare its out-of-region predictive scores with those of a version that includes an explicit spatial random field: if the spatial-field version wins out-of-region, the transferability premise fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the response dataset: the Gorkha earthquake landslide inventory of polygons whose centroids and areas define the point pattern and marks."},{"cited_title":"K., West, A","cited_arxiv_id":null,"evidence_quote":"Primary source inventory of 24,915 landslides integrated into the catalogue the modelling is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LSDTopoTools software used to compute $k_{sn}$, relief to channel, and flow distance to channel from the DEM."},{"cited_title":"H., Naylor, M., and Lindgren, F","cited_arxiv_id":null,"evidence_quote":"The coherent disaggregation method that justifies the mesh resolution and the alignment of pixel covariates with point-response data."},{"cited_title":"W., Mudd, S","cited_arxiv_id":null,"evidence_quote":"Hilltop flow routing algorithm that produces the relief-to-channel and flow-distance metrics used as covariates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Stream-power incision model linking channel steepness to erosion, the physical basis for using $k_{sn}$ as a landslide covariate."},{"cited_title":"D., Hearne, M., Jaiswal, K","cited_arxiv_id":null,"evidence_quote":"ShakeMap Atlas source of the peak ground acceleration raster used as the seismic trigger covariate."}],"review_version":1}