REVIEW 3 major objections 6 minor 26 references
Spatial-variant causal Bayesian inference for rapid seismic ground failures and impacts estimation
T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read 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…
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (3)
- [Section 2.1, Eqs. (2) and (3)] 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 2.2, Monte Carlo estimate paragraph] 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 3, evaluation methodology] 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.
minor comments (6)
- [Section 2.1, Eq. (1)] 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 2.1, Eq. (3)] 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 2.2, Eq. (5)] The variational lower bound in Eq. (5) has mismatched parentheses and unclear integration variables, making it difficult to verify the subsequent derivations.
- [Section 2.2, subsection heading] 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.
- [References] 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 3, Figures 4a-4d] 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.
Circularity Check
Spatial variable is not an independent parent: Eq. (3) constructs 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 reported AUC gains are not attributable to a sound causal spatial model.
-
self definitional
[Section 2.1, Equations (2) and (3), Figure 1]
"For each x_{i',j'}^{h} in Ω_{i,j}, we consider its intensity value I(x_{i',j'}^{h}), its spatial distance from (i,j), and the intensity difference between I(x_{i',j'}^{h}) and I(x_{i,j}^{h}) to compute the filter response. ... 𝑠_{i,j}^{h}=1/𝑀_{i,j}∑_{𝑥_{i',j'}^{h}∈Ω_{i,j}} 𝑥_{i',j'}^{h}×𝑓_{𝑟}(Δ𝑥_{i',j',i,j})×𝑔_{𝑠}(||(𝑖′,𝑗′)−(𝑖,𝑗)||) ... 𝑝(𝑥_{i,j}^{h}=1|𝒫(ℎ_{i,j}),𝜖_{i,j}^{h})=1/(1+exp(−(∑_{𝑘∈𝒫(ℎ_{i,j})} 𝑤_{𝑘}^{h} 𝑥_{𝑘,i,j}^{h}+𝑤_{+}^{h} 𝜖_{i,j}^{h}+𝑤_{0}^{h}+𝑤_{𝑠}^{h} 𝑠_{i,j}^{h})))"
Equation (2) makes s_h a parent of x_h. Equation (3) defines s_h as a bilateral-filtered aggregate of neighboring x_h values, and the range kernel compares the central intensity I(x_h_{i,j}) with neighboring intensities. Thus s_h depends on x_h at the central cell and at neighboring cells, while those neighboring x_h are themselves determined by Equations (2)-(3) through their own s_h terms. The graph in Figure 1 is therefore not a DAG: it contains directed cycles x_h -> s_h -> x_h. The spatial variable is not an exogenous causal parent; it is a deterministic function of the very latent field the model claims to infer. This invalidates the causal interpretation of the spatial extension.
-
fitted input called prediction
[Section 2.2, Monte Carlo estimate of E(log p(s_h)) and Equation (11)]
"We employ Monte Carlo estimates to evaluate 𝔼(𝑙𝑜𝑔𝑝(𝑠_{i,j}^{h})), where 𝑠_{i,j}^{h} derived from neighboring values of 𝑥_{i',j'}^{h}... our first step is to sample from 𝑝(𝑥^{h}). To do this, we draw 𝑁 samples... For each spatial sample 𝑥_{n}^{h}, we calculate the value of 𝑠_{i,j,n}^{h} using the bilateral... we tally the occurrence frequency of each distinct value of 𝑠_{i,j,n}^{h} within the dataset. The probability 𝑝(𝑠_{i,j,n}^{h}) is then the frequency of a specific value divided by the total number of data points 𝑁."
The quantity p(s_h) used in the variational lower bound is not a fixed external prior. It is the empirical frequency of s_h values recomputed from Monte Carlo draws of the latent hazard field x_h taken from the current approximate posterior. The spatial term is therefore re-estimated from the very variable it is supposed to constrain, every iteration. This makes the spatial prior a self-referential regularizer rather than an independent causal input. Consequently, the reported improvements in AUC over the USGS priors and the VCBI model cannot be attributed to a sound spatial causal prior; they may reflect this feedback loop rather than a genuine spatial causal effect.
full rationale
The central contribution of the paper is the spatial variable s_h. That contribution is circular by construction. Eq. (2) places s_h in the logistic activation for x_h, while Eq. (3) defines s_h as a bilateral-filtered function of the x_h field, including the central intensity I(x_h_{i,j}) through the range kernel. This creates the directed cycle x_h -> s_h -> x_h, so the model is not a causal Bayesian network with a DAG as claimed. Section 2.2 compounds the problem: p(s_h) is estimated by tallying s_h values computed from Monte Carlo samples of x_h drawn from the current variational posterior, so the spatial prior is re-estimated from the quantity it is supposed to constrain. The reported AUC values are measured against external ground truth, so the empirical evaluation is not wholly circular, but the spatial mechanism that is claimed to drive the improvement reduces to a self-referential fit. No ablation with a non-referential smoothing baseline or held-out validation is provided to rule out the feedback explanation. The self-citations to the authors' prior VCBI model supply the baseline and model family, but they are not the load-bearing circular step; the circularity lies in the model equations themselves.
Assumptions & free parameters
free parameters (3)
- causal weights w for observation and hazard activation models (w_k^y, w_0^y, w_eps^y, w_sd^y, w_0^h, w_eps^h, w_s^h) =
not reported in text
- bilateral filter kernel widths sigma_r and sigma_d =
not reported
- Monte Carlo sample count N for p(s) empirical distribution =
not specified
assumptions (6)
- domain assumption DPM observation y is log-normally distributed around a weighted sum of active hazards and noise (Eq.1)
- domain assumption Hazard activation x_h follows a Bernoulli logistic model with parents, noise, and spatial variable (Eq.2)
- ad hoc to paper Spatial variable s_h is a well-defined random variable with a distribution p(s) separate from x
- standard math Mean-field factorization q(X)=prod q_h and Jensen/Taylor bounds give a valid ELBO
- domain assumption Ground truth field observations are complete and accurate for ROC evaluation
- domain assumption USGS prior models and fragility curves provide valid priors
invented entities (1)
-
spatial variable s_h_{i,j} for each hazard and grid location
Cite this review
Pith. "Pith review of Spatial-variant causal Bayesian inference for rapid seismic ground failures and impacts estimation." pith.science (2026). https://pith.science/paper/OJO33W36
@misc{pith2026241200026,
author = {Pith},
title = {Pith review of: Spatial-variant causal Bayesian inference for rapid seismic ground failures and impacts estimation},
year = {2026},
howpublished = {\url{https://pith.science/paper/OJO33W36}},
note = {Machine review of arXiv:2412.00026}
}
read the original abstract
Rapid and accurate estimation of post-earthquake ground failures and building damage is critical for effective post-disaster responses. Progression in remote sensing technologies has paved the way for rapid acquisition of detailed, localized data, enabling swift hazard estimation through analysis of correlation deviations between pre- and post-quake satellite imagery. However, discerning seismic hazards and their impacts is challenged by overlapping satellite signals from ground failures, building damage, and environmental noise. Previous advancements introduced a novel causal graph-based Bayesian network that continually refines seismic ground failure and building damage estimates derived from satellite imagery, accounting for the intricate interplay among geospatial elements, seismic activity, ground failures, building structures, damages, and satellite data. However, this model's neglect of spatial heterogeneity across different locations in a seismic region limits its precision in capturing the spatial diversity of seismic effects. In this study, we pioneer an approach that accounts for spatial intricacies by introducing a spatial variable influenced by the bilateral filter to capture relationships from surrounding hazards. The bilateral filter considers both spatial proximity of neighboring hazards and their ground shaking intensity values, ensuring refined modeling of spatial relationships. This integration achieves a balance between site-specific characteristics and spatial tendencies, offering a comprehensive representation of the post-disaster landscape. Our model, tested across multiple earthquake events, demonstrates significant improvements in capturing spatial heterogeneity in seismic hazard estimation. The results highlight enhanced accuracy and efficiency in post-earthquake large-scale multi-impact estimation, effectively informing rapid disaster responses.
Reference graph
Works this paper leans on
-
[1]
Introduction Due to their destructive nature, earthquakes not only cause direct building damage from ground shaking but also can trigger cascading ground failures like landslides and liquefaction. Along with ground shaking, these ground failures can have significant consequences, including building damage and loss of life. For example, the 2021 Haiti eart...
work page 2021
-
[2]
Geospatial prior and posterior liquefaction estimation models of the 2020 Puerto Rico earthquake. (a) Prior liquefaction model. (b) Prior liquefaction model with ground truth observations. (c) Posterior liquefaction model. (d) Posterior liquefaction model with ground truth observations. Evaluation findings indicate that the posterior liquefaction model, d...
work page 2020
-
[3]
However, their application is limited when geotechnical data are absent
and the liquefaction potential index (Zhu et al., 2016). However, their application is limited when geotechnical data are absent. Despite their foundation in primary physical processes, these models can be prone to errors, often from oversimplified representations of intricate physical dynamics. Building fragility curve, a log-normal function that estimat...
work page 2016
-
[4]
primarily focuses on direct causal relations and does not account for the complex nature of these spatial variables and their relationships. As a result, it tends to overlook the nuanced interplay of spatial relationships, leaving room for improvements in both accuracy and comprehensiveness. Recognizing these spatial relationships among neighboring areas ...
work page 1998
-
[5]
,$, which is bounded by [0,1], to refer to our sensing observation. We denote 𝑥!
Causal Spatial-Variant Bayesian Network for multi-hazards and impacts estimation in seismic events. (𝑖,𝑗) in the feature refers to the location coordinate. Green rectangles refer to the known variables. Blue circles refer to unobserved nodes. 𝛾! are the unknown causal parameters that quantify the causal relations among parent and child nodes. 𝜇! are the u...
work page 2022
-
[6]
It is computed as the sum of all the weights given by the product of the range and spatial kernels for all 𝑥!".,$. in Ω(",$). Utilizing the above distributions and conditional distribution assumptions, we develop a Bayesian network rooted in the causal graph, adeptly representing the interdependencies among ground failures, building damage, remote sensing...
work page 1999
-
[7]
Evaluations In this section, we evaluate our spatial-variant causal Bayesian inference method for rapid seismic ground failures, which are landslides and liquefaction in our case, and building damage estimation using two events: the 2020 Puerto Rico and the 2021 Haiti earthquakes. We utilize the current USGS ground failure products for landslides and liqu...
work page 2020
-
[9]
(a) DPM with significant signal changes
Geospatial prior and posterior building damage estimation models of the 2020 Puerto Rico earthquake. (a) DPM with significant signal changes. (b) posterior building damage model with building footprints. (c) posterior building damage model with building footprints and ground truth observations. Figure 3 illustrates the performance of our model in estimati...
work page 2020
Show all 26 references
-
[10]
(b) The 2020 Puerto Rico building damage estimation
ROC curve of baseline models and our posterior model for (a) The 2020 Puerto Rico liquefaction estimation. (b) The 2020 Puerto Rico building damage estimation. (c) The 2021 Haiti landslide estimation. (d) The 2021 Haiti building damage estimation. 3.2. The evaluation results o...
2020
-
[11]
(a) Prior landslide model
Geospatial prior and posterior landslide estimation models of the 2021 Haiti earthquake. (a) Prior landslide model. (b) Prior landslide model with ground truth observations. (c) Posterior landslide model. (d) Posterior landslide model with ground truth observations. In the con...
2021
-
[12]
(a) DPM with significant signal changes
Geospatial prior and posterior building damage estimation models of the 2021 Haiti earthquake. (a) DPM with significant signal changes. (b) posterior building damage model with building footprints. (c) posterior building damage model with building footprints and ground truth o...
2021
-
[15]
https://www.fema
Hazus-MH 2.1 AEBM User’s Manual. https://www.fema. gov/sites/default/files/2020- 09/fema_hazus_advanced- engineering- building- module_user- manual.pdf Geotechnical Extreme Events Reconnaissance (GEER). 2021, 2021 HAITI EARTHQUAKE, https://geerassociation.org/component/geer_re...
2020
-
[16]
StEER-Aegean Sea Earthquake (30 October 2020),
Preliminary virtual reconnaissance report (pvrr). StEER-Aegean Sea Earthquake (30 October 2020),
2020
-
[20]
Designsafe-CI
StEER-Puerto Rico Earthquake Sequence December 2019 to January 2020: early access reconnaissance report (EARR). Designsafe-CI. Newmark, N.M.,
2019
-
[22]
4 (7 September 2021)
Haiti: Earthquake Situation Report No. 4 (7 September 2021). Xiang, S., & Zhang, H.,
2021
-
[23]
Nature Communications, 13(1), p.7793
Seismic multi-hazard and impact estimation via causal inference from satellite imagery. Nature Communications, 13(1), p.7793. Xu, S., Li, X., Noh, H. and Wald, D.J., 2022, December. Deep Causal Bayesian Network for Modeling Spatial Seismic Building Damage from Remote Sensing O...
2022
-
[25]
Geomorphology, 415, p.108419
Evaluation of factors controlling the spatial and size distributions of landslides, 2021 Nippes earthquake, Haiti. Geomorphology, 415, p.108419. Zhao, C. and Lu, Z.,
2021
-
[26]
Bulletin of the Seismological Society of America, 110(4), pp.1549-1566
Liquefaction and related ground failure from July 2019 Ridgecrest earthquake sequence. Bulletin of the Seismological Society of America, 110(4), pp.1549-1566
2019
-
[1971]
Journal of the Soil Mechanics and Foundations division, 97(9), pp.1249-1273
Simplified procedure for evaluating soil liquefaction potential. Journal of the Soil Mechanics and Foundations division, 97(9), pp.1249-1273. Tomasi, C., & Manduchi, R., 1998, January. Bilateral filtering for gray and color images. In Sixth international conference on computer...
1998
-
[2002]
Havenith, H.B., Guerrier, K., Schlögel, R., Mreyen, A.S., Ulysse, S., Braun, A., Victor, K.H., Saint-Fleur, N., Cauchie, L., Boisson, D
Anomalous concentrations of seismically triggered rock falls in Pacoima Canyon: Are they caused by highly susceptible slopes or local amplification of seismic shaking? Bulletin of the Seismological Society of America, 92(8), pp.3180-3189. Havenith, H.B., Guerrier, K., Schlögel...
2022
-
[2015]
Seismological Research Letters, 86(6), pp.1549-1556
Rapid damage mapping for the 2015 M w 7.8 Gorkha earthquake using synthetic aperture radar data from COSMO–SkyMed and ALOS-2 Satellites. Seismological Research Letters, 86(6), pp.1549-1556. Yun, S.H., Fielding, E.J., Webb, F.H. and Simons, M., California Institute of Technolog...
2015
-
[2017]
Natural Hazards and Earth System Sciences, 17(5), pp.781-800
Evaluating simplified methods for liquefaction assessment for loss estimation. Natural Hazards and Earth System Sciences, 17(5), pp.781-800. Koller, D. and Friedman, N., 2009, December. Probabilistic graphical models: principles and techniques. MIT press. Jensen, F.V. and Niel...
2009
-
[2020]
Allstadt, K.E., Thompson, E.M., Hearne, M., Jessee, M.N., Zhu, J., Wald, D.J
Available at https://aria- share.jpl.nasa.gov/20200106- Puerto_Rico_EQ/DPM/. Allstadt, K.E., Thompson, E.M., Hearne, M., Jessee, M.N., Zhu, J., Wald, D.J. and Tanyas, H., 2017, January. Integrating landslide and liquefaction hazard and loss estimates with existing USGS real-ti...
2017
-
[2021]
Remote Sensing of Environment, 262, p.112508
Exploring the relationship between InSAR coseismic deformation and earthquake-damaged buildings. Remote Sensing of Environment, 262, p.112508. Li, X., Bürgi, P.M., Ma, W., Noh, H.Y., Wald, D.J. and Xu, S., 2023, August. DisasterNet: Causal Bayesian Networks with Normalizing Fl...
2023
-
[2022]
Seismological Research Letters, 93(2A), pp.594-608
Ground failure triggered by the 7 january 2020 m 6.4 puerto rico earthquake. Seismological Research Letters, 93(2A), pp.594-608. Buades, A., Coll, B., & Morel, J. M.,
2020
-
[2023]
Fundamental physical processes form the basis of physical models, exemplified by the Newmark displacement-centric landslide model (Jibson, 2000; Newmark,
and statistical counterparts (Jessee et al., 2016; Nowicki et al., 2014; Seed and Idriss, 1971; Zhu et al., 2015). Fundamental physical processes form the basis of physical models, exemplified by the Newmark displacement-centric landslide model (Jibson, 2000; Newmark,
2016
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.