{"id":"e0967e31-daff-4a60-bc03-7e6d59c06731","arxiv_id":"2412.00026","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"This paper adds a bilateral-filter-based spatial variable to a causal Bayesian network for post-earthquake hazard estimation, reporting modest AUC gains on two events but using a self-referential model construction.","lead":"After an earthquake, this paper adds a spatial smoothing variable to a Bayesian network that estimates landslides, liquefaction, and building damage from satellite radar change maps. The authors report better accuracy on two earthquakes, but the spatial variable is defined from the same hidden data it is supposed to predict, creating a circular inference loop.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The spatial variable is not an independent parent: Eq. 3 builds s_h from the latent x_h field (including the central cell), and Sec. 2.2 estimates p(s_h) from posterior samples of x_h, so the model is self-referential and not a well-defined causal Bayesian network.","rationale":"I read the paper in good faith. It is a plausible empirical post-processing scheme: smoothing the posterior hazard field with a bilateral filter and feeding it back could indeed improve AUC on the same event, because spatial context is informative. But the paper frames this as a causal Bayesian network, and that framing fails in two linked places. First, the spatial variable is not causally exogenous: it is built from the hazard field it is meant to explain. Second, the distribution over that variable is estimated from Monte Carlo samples of the same latent field, so the variational objective in Eq. 11 contains a term that depends on the variational distribution in a way standard variational inference does not allow. The evaluation is also on the same events used to fit the weights, with no error bars, so the high AUC numbers do not independently validate the causal story. I agree with the reader's weakest assumption. The claims could in principle be salvaged by redefining s_h from external covariates (e.g., peak ground acceleration, slope, lithology) and fixing p(s_h) a priori, but as written the model is not a valid causal Bayesian network. The REJECT verdict is therefore unchanged.","tokens_in":12119,"tokens_out":5740,"duration_ms":58448,"concrete_test":"Run an ablation on the 2020 Puerto Rico and 2021 Haiti data: compute s_h_{i,j} using only the eight neighboring x_h values, omitting the central cell and its intensity from the range kernel, and fix p(s_h) once from the USGS prior hazard maps before variational inference rather than re-estimating it from q(x_h) samples. If the four AUC values fall back to near the VCBI baselines (0.9121, 0.9309, 0.9032, 0.9021), the reported improvements are an artifact of circular feedback. As a complementary analytical check, evaluate ∂s_h_{i,j}/∂x_h_{i,j}; any nonzero derivative confirms that Eq. 3 makes s_h a function of the child x_h, so the network is not acyclic.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that adding a spatial variable improves causal Bayesian estimation of ground failures and damage—requires s_h_{i,j} to be a well-defined exogenous spatial context that can act as a parent of x_h_{i,j}. Equation 2 places s_h_{i,j} in the logistic activation for x_h_{i,j}. Equation 3 defines s_h_{i,j} as a normalized bilateral-filtered sum of neighboring x_h values, with the range kernel f_r comparing I(x_h_{i',j'}) to I(x_h_{i,j}). Since I is not separately defined and x_h is stated to be binary, I(x_h_{i,j}) must be a function of the latent hazard at the same cell; thus s_h_{i,j} depends on x_h_{i,j}, and Eqs. 2 and 3 form a directed cycle x_h -> s_h -> x_h. The graph in Fig. 1 is therefore not a DAG. Even if I were an external shaking intensity, Sec. 2.2 states that p(s_h) is estimated by tallying s_h values computed from N Monte Carlo samples of x_h drawn from the current model/variational posterior. The 'spatial prior' is thus re-estimated from the very quantity it is supposed to constrain, which is not a valid prior and can produce arbitrary self-reinforcing spatial patterns. The reported AUC gains over USGS and VCBI are therefore not attributable to a sound causal spatial model; they may be an artifact of this feedback loop. No held-out validation, error bars, or alternative smoothing baseline is provided to rule this out.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spatial-variant causal Bayesian network for rapid post-earthquake estimation of ground failures (landslides, liquefaction) and building damage from satellite-derived damage proxy maps. The main novelty is the introduction of a spatial variable s_h for each hazard h at each location (i,j), computed with a bilateral filter over neighboring latent hazard values, and its inclusion as a parent of the latent hazard variable in the logistic activation model. The authors evaluate the method on the 2020 Puerto Rico and 2021 Haiti earthquakes, reporting AUC values of 0.9451 (Puerto Rico liquefaction), 0.9720 (Puerto Rico building damage), 0.9373 (Haiti landslides), and 0.9425 (Haiti building damage), all outperforming USGS priors and the prior VCBI model.","tokens_in":12498,"tokens_out":2826,"duration_ms":26976,"significance":"The problem addressed is practically important: rapid and accurate post-earthquake hazard and damage estimation can inform disaster response. The paper applies the method to real earthquake events with field-verified ground truth and compares against established baselines, which is commendable. If the spatial extension were well-posed, the reported improvements could be of practical value. However, the central methodological innovation is not a valid causal Bayesian network as defined: the spatial variable is a function of the very latent variable it is supposed to help predict, and its distribution is estimated from posterior samples in a circular fashion. These issues undermine the attribution of the reported gains to a sound spatial model, so the significance of the contribution cannot be assessed as presented.","major_comments":[{"comment":"The spatial variable s_h_{i,j} is defined in Eq. (3) as a bilateral-filtered function of the neighboring latent hazard values, including the central cell's intensity I(x_h_{i,j}), and then used in Eq. (2) as a parent of x_h_{i,j} in the logistic activation. Since x_h is binary and I is not separately defined, I(x_h_{i,j}) must be a function of the same latent variable, creating a directed cycle x_h -> s_h -> x_h. The graph in Fig. 1 is therefore not a DAG, and the joint distribution is not well-defined. This is a load-bearing error because the central claim of a spatial causal extension depends on s_h being an exogenous parent.","section":"Section 2.1, Eqs. (2) and (3)"},{"comment":"The distribution p(s_h) is estimated by drawing N samples of the full spatial field x_h from the current model, computing s_h from those samples via the bilateral filter, and tallying frequencies. This means the 'spatial prior' is re-estimated from the posterior of the very quantity it is supposed to constrain, making the inference circular. The spatial variable provides no independent information, and the reported AUC improvements (e.g., 0.9451 vs. 0.9121 for Puerto Rico liquefaction) may be artifacts of this feedback loop rather than evidence of capturing spatial heterogeneity.","section":"Section 2.2, Monte Carlo estimate paragraph"},{"comment":"The evaluation reports point AUC values without held-out validation, confidence intervals, or error bars, and it does not compare against a non-circular spatial baseline such as a fixed spatial prior, a spatial autoregressive model, or a simple smoothing of the USGS prior. Given the circular construction, these controls are necessary to establish that the gains are not self-reinforcing artifacts. The authors should also show that the model's spatial variable is identifiable and report performance on held-out events or cross-validated splits.","section":"Section 3, evaluation methodology"}],"minor_comments":[{"comment":"The log-normal distribution in Eq. (1) is written with unclear notation: the mean and standard deviation arguments are not cleanly separated, and the conditioning on epsilon_y is ambiguous.","section":"Section 2.1, Eq. (1)"},{"comment":"The intensity function I(x) is never defined, and since x is binary, the range kernel f_r is undefined unless I is specified (e.g., as ground shaking intensity, but then it is not a function of x_h).","section":"Section 2.1, Eq. (3)"},{"comment":"The variational lower bound in Eq. (5) has mismatched parentheses and unclear integration variables, making it difficult to verify the subsequent derivations.","section":"Section 2.2, Eq. (5)"},{"comment":"The subsection is titled 'Stochastic Optimization' but the previous subsection is already named 'Variational Inference for Posteriors Approximation'; the second heading appears to be a typo for 'Stochastic Variational Inference' or similar.","section":"Section 2.2, subsection heading"},{"comment":"There are duplicate entries (Xu et al. 2022 appears twice; Yun et al. 2015 appears twice), and some references are incomplete or informal (e.g., 'Web, R.' and the Allstadt et al. 2017 entry).","section":"References"},{"comment":"The ROC curves are shown but not discussed in detail; reporting AUC to four decimal places without confidence intervals or significance tests overstates the precision of the comparisons.","section":"Section 3, Figures 4a-4d"}],"recommendation":"reject","confidential_remarks":"The fundamental issue is a directed cycle in the proposed graphical model and a circular estimation of the spatial prior, which cannot be fixed by minor edits. A revision would require redefining the spatial variable as an exogenous quantity (e.g., based on ground-motion intensity or a fixed spatial field) and redoing the inference and evaluation. The paper may have a useful idea for incorporating spatial smoothing into causal Bayesian estimation, but as written the central claim is not supported. I would also note that the paper reads like a conference submission (WCEE2024) and may be more appropriate for a venue with workshop-style scrutiny of methodology."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe short version: the paper's spatial variable is defined from the latent hazard field it is supposed to predict, so the causal Bayesian network isn't actually a DAG. That's a load-bearing problem, not a cosmetic one. The reported AUC gains over the prior VCBI model of 2-4 points on the two events could well be an artifact of this self-referential smoothing.\n\nThat said, the authors are working on a real and important problem—rapid post-earthquake estimation of landslides, liquefaction, and building damage from radar. The bilateral-filter idea is a reasonable attempt to bring spatial context into the model, and the evaluation on the 2020 Puerto Rico and 2021 Haiti earthquakes, with USGS priors and ground-truth comparisons, is the right kind of evidence to bring. The writing is clear about the motivation, and building on Xu et al. (2022) is sensible.\n\nThe soft spots are substantial. Equation 3 defines s_h as a bilateral-filtered function of neighboring x_h values and the central cell's intensity I(x_h_{i,j}); Equation 2 then uses s_h as a parent of x_h. That's a directed cycle. Section 2.2 makes it worse: p(s_h) is estimated from Monte Carlo samples of x_h drawn from the current posterior, so the 'spatial prior' is re-estimated from the very quantity it is supposed to constrain. This is not a causal prior; it's a feedback loop. There are no error bars, no hyperparameter values, no held-out events, and no code or data, so nothing in the paper rules out the feedback artifact. The missing details would be minor if the core definition were sound, but here they compound the central flaw.\n\nThe paper is probably most useful to readers working on spatial Bayesian networks or rapid damage mapping—as an example of how easy it is to make a spatial feature self-referential. For a journal, I'd say it deserves a serious referee only if the authors can fix the definition: make s_h depend on external data or on neighbors only (excluding the center), and treat p(s) as a fixed prior. That would be a major rewrite, so if I were the editor I'd send it back for major revision rather than accept anything close to the current form. If the journal isn't willing to wait for that, desk-rejecting is defensible.","headline":"The spatial variable is built from the latent field it is meant to explain, making the model circular and the AUC gains uninterpretable, though the empirical setup is reasonable.","tokens_in":13071,"tokens_out":6338,"would_cite":false,"duration_ms":60213,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["91.30.Px"],"model":"deepseek-v4-flash","headline":"This paper claims that adding a bilateral-filtered spatial variable to a causal Bayesian network improves post-earthquake estimates of ground failures and building damage, with AUC values of 0.9451, 0.9720, 0.9373, and 0.9425 across the…","keywords":["causal Bayesian network","spatial heterogeneity","bilateral filter","seismic ground failure","building damage estimation","landslide","liquefaction","variational inference"],"falsifier":"On the same two events, replace the spatial variable $s^h_{i,j}$ with a fixed Gaussian-smoothed version of the prior hazard map and re-run the same variational inference; if the AUC values do not drop, the model's gains are not caused by the bilateral-filter spatial parent as described.","tokens_in":11835,"feed_emoji":"🌍","tokens_out":13483,"duration_ms":117121,"temperature":0.7,"pith_summary":"This paper tries to show that post-earthquake maps of landslides, liquefaction, and building damage can be made more accurate by giving the causal Bayesian network a spatial variable: for each location, a bilateral-filtered summary of hazard in the surrounding cells is added as an additional cause of the local hazard. The motivation is that seismic damage is spatially clustered, and ignoring this clustering leaves estimates that miss local concentrations of failure. The authors test the idea on the 2020 Puerto Rico and 2021 Haiti earthquakes and report that the spatial-variant model outperforms both the standard empirical priors and the earlier causal Bayesian model on all four tasks, with AUC values of 0.9451, 0.9720, 0.9373, and 0.9425. If the approach holds, rapid response teams could get sharper damage maps in the first hours after a quake, when field data are still sparse.","feed_headline":"Spatial context lifts quake-impact map scores to 0.97 AUC","feed_subtitle":"Bilateral-filter spatial priors sharpen earthquake hazard and building-damage estimates beyond standard causal models.","key_machinery":"The central object is the spatial variable $s^h_{i,j}$, a bilateral-filtered summary of the three-by-three neighborhood around each cell. The bilateral filter is an edge-preserving image-processing smoother: it weights a neighbor's contribution by a Gaussian of spatial distance and by a Gaussian of intensity similarity to the central cell, so nearby hazards of similar severity matter most while sharp boundaries between affected and unaffected zones are kept. The paper inserts $s^h_{i,j}$ as an extra parent in the hazard node's Bernoulli activation probability and estimates its distribution inside variational inference by Monte Carlo sampling from the current posterior over the neighboring hazard field, rather than fixing it from external data. The range and spatial kernel widths determine how much the final map is smoothed versus how much local detail is preserved.","core_discovery":"The paper's central claim is that spatial context carries causal information about seismic hazards, and that a bilateral-filtered spatial variable can inject that context into a Bayesian network. Each latent hazard node $x^h_{i,j}$ for hazard $h$ at cell $(i,j)$ is given a new spatial parent $s^h_{i,j}$, a normalized weighted sum of the hazard values in the eight surrounding cells, where the weight of each neighbor is a Gaussian of the difference between its intensity and the central cell's intensity times a Gaussian of spatial distance. This edge-preserving summary enters the logistic activation probability of the local hazard, alongside the usual geospatial and shaking parents. On the 2020 Puerto Rico and 2021 Haiti earthquakes, the paper reports AUCs of 0.9451 for liquefaction, 0.9720 and 0.9425 for building damage, and 0.9373 for landslides, each above the corresponding standard prior and above the earlier causal Bayesian model.","pith_inferences":["If the reported gains hold, a direct ablation that replaces the bilateral filter with a fixed Gaussian smoother of matched width would isolate whether edge preservation or mere spatial smoothing causes the improvement; this goes beyond the paper's comparisons.","The same spatial-parent construction could be transferred to other cascading hazards, such as tsunami inundation or fire following earthquake, whenever the causal graph already links a hazard to its neighbors.","Because the spatial prior is estimated from Monte Carlo samples of the hazard field itself, the reported AUCs may be sensitive to sample count and mini-batch schedule; a sensitivity analysis would reveal whether the improvement is stable or tuning-dependent.","If the mechanism generalizes, it suggests that spatial context should be modeled as a causal parent in other remote-sensing Bayesian updating tasks where the target quantity clusters, not just in seismology."],"forward_implications":["Spatial-variant causal updating yields higher AUC than the standard empirical priors and the earlier causal model on all four reported tasks: liquefaction and building damage in Puerto Rico, and landslide and building damage in Haiti.","Posterior maps concentrate probability near observed clusters of ground failure and damage, meaning rescue resources could be targeted more precisely from the same satellite imagery.","The gain comes from a balance between site-specific features and neighborhood tendency, so the approach should transfer to other events where damage is spatially clustered.","Because the method uses the same inputs as the earlier model, the spatial variable acts as a drop-in enhancement rather than a new sensing requirement."],"supporting_citations":[{"why":"Establishes the earlier causal Bayesian network and variational inference procedure that this paper extends with a spatial variable.","marker":"Xu et al., 2022"},{"why":"Introduces the bilateral filter whose range and spatial kernels define the new spatial variable.","marker":"Tomasi and Manduchi, 1998"},{"why":"Provides the global empirical landslide model used as the prior landslide probability and a baseline for comparison.","marker":"Nowicki et al., 2018"},{"why":"Provides the geospatial liquefaction model used as the prior liquefaction estimate and a baseline for comparison.","marker":"Zhu et al., 2015"},{"why":"Supplies the building fragility curve prior model for building damage and the baseline damage estimates.","marker":"FEMA, 2020"},{"why":"Provides the damage proxy maps that serve as the sensing observations for the Puerto Rico case.","marker":"ARIA, 2020"},{"why":"Supplies ground-truth observations of ground failure for the 2020 Puerto Rico earthquake evaluation.","marker":"Allstadt et al., 2022"},{"why":"Supplies building damage field observations used in the 2021 Haiti evaluation.","marker":"Kijewski-Correa et al., 2021"},{"why":"Supplies reconnaissance ground-truth data for landslides and damage in the Haiti earthquake.","marker":"GEER, 2021"}],"fun_headline_variants":["Spatial context lifts quake-impact AUC to 0.97","Causal Bayesian net with spatial priors sharpens quake damage maps","Bilateral-filter spatial variable boosts seismic hazard estimation","Spatial-aware causal inference improves rapid post-quake impact mapping","Edge-preserving spatial context refines quake hazard and damage estimates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a location's bilateral-filtered summary of its neighbors can be treated as an independent cause of that location's hazard, even though the summary is computed from those neighbors' hazards and from the same shaking intensity that already drives the local hazard estimate.","fun_headline_variants_meta":{"raw":{"variants":["Spatial context lifts quake-impact AUC to 0.97","Causal Bayesian net with spatial priors sharpens quake damage maps","Bilateral-filter spatial variable boosts seismic hazard estimation","Spatial-aware causal inference improves rapid post-quake impact mapping","Edge-preserving spatial context refines quake hazard and damage estimates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00073,"raw_usage":{"total_tokens":3300,"prompt_tokens":1006,"completion_tokens":2294,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":2205}},"tokens_in":622,"tokens_out":2294,"duration_ms":17103,"temperature":1.0,"reasoning_tokens":2205,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:18:25.524440+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the same two events, replace the spatial variable $s^h_{i,j}$ with a fixed Gaussian-smoothed version of the prior hazard map and re-run the same variational inference; if the AUC values do not drop, the model's gains are not caused by the bilateral-filter spatial parent as described.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the damage proxy maps that serve as the sensing observations for the Puerto Rico case."}],"review_version":1}