arxiv: 2604.18840 · v1 · submitted 2026-04-20 · 📊 stat.AP · stat.CO· stat.ME

Recognition: unknown

Spatial Extremes at Scale: A Case Study of Surface Skin Temperature and Heat Risk in the United States

Ben Seiyon Lee, Emma S. Simpson, Jordan Richards, Likun Zhang, Reetam Majumder

Pith reviewed 2026-05-10 02:46 UTC · model grok-4.3

classification 📊 stat.AP stat.COstat.ME

keywords spatial extremesBayesian inferencescale mixture processamortized learningsurface skin temperatureheat riskFour Cornersstatistical modeling

0 comments

The pith

A random scale mixture process enables scalable Bayesian inference for spatial extremes in US surface temperatures.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a statistical model to handle how extreme temperature events depend on one another across many locations in space. Standard approaches for this become too slow when the number of sites grows large, as in modern climate datasets. The authors introduce a random scale mixture process and pair it with advances in spatial statistics and amortized learning to make full Bayesian analysis practical. They test the approach with large simulation studies and then apply it to high-resolution surface skin temperature records from the Four Corners region to map heat risks. Accurate spatial models of this kind matter because they support better public health planning for heat in areas with varied terrain.

Core claim

The authors claim that the random scale mixture process, together with scalable inference strategies that leverage spatial modeling and amortized learning, makes Bayesian inference feasible for spatial extremes, as shown by extensive simulation studies and an application to high-resolution surface skin temperature data in the Four Corners region of the United States.

What carries the argument

The random scale mixture process, a model that represents complex joint tail dependencies among extreme values observed at many spatial locations.

If this is right

Bayesian modeling of spatial extremes becomes feasible for datasets containing thousands of locations.
Spatially varying and seasonally changing heat extremes can be characterized in regions with complex terrain.
Surface skin temperature can be used to derive location-specific heat indices that inform public health risk.
Practitioners in climate science and environmental risk assessment gain practical guidelines for large-scale extreme value analysis.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same mixture structure could be applied to other spatial extremes such as heavy rainfall or strong winds.
Amortized components might allow the model to update quickly when new temperature observations arrive.
Extending the method across the full United States could identify national patterns in heat risk that local studies miss.
Direct comparison of the inferred extremes against output from physical climate models would test consistency between statistical and process-based views.

Load-bearing premise

The random scale mixture process and the amortized learning approximations together capture the actual joint tail dependencies in the temperature data without introducing bias or artifacts that change the heat risk conclusions.

What would settle it

On a smaller data subset where exact but slow Bayesian inference is still possible, the scalable method produces noticeably different estimates of extreme quantiles or risk measures than the exact version.

Figures

Figures reproduced from arXiv: 2604.18840 by Ben Seiyon Lee, Emma S. Simpson, Jordan Richards, Likun Zhang, Reetam Majumder.

**Figure 1.** Figure 1: Left: Location of the study region within the broader south-central and southwestern United States. The colored raster cells mark the NLDAS grid points over the Four Corners region and their elevations. Middle: Spatially averaged seasonal maximum surface skin temperature (Tskin) over 1979–2024. For each year and season (DJF, MAM, JJA, SON), seasonal block maxima are computed at each grid location and the… view at source ↗

**Figure 2.** Figure 2: Seasonal maps of the maximum a posteriori (MAP) estimate of the marginal GEV shape field {ξ(s)} over the study region for each season. The white points denote the held-out NLDAS grid cells excluded from model fitting. The four triangle markers identify the GHCN stations—Winslow, Gallup, Albuquerque, and Santa Fe—used for subsequent station-based validation. State boundaries are overlaid for geographic refe… view at source ↗

**Figure 3.** Figure 3: Spatial maps of seasonal maxima for the Four Corners region in 2024. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

**Figure 4.** Figure 4: Schematic of the modeling framework linking the latent L´evy random scale [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗

**Figure 5.** Figure 5: Low-dimensional study. Comparison of parameter estimation accuracy and computational cost across inference methods under four representative simulation settings. Here, Vec(m) denotes the Vecchia approximation with conditioning set size m, Tap(p%) denotes covariance tapering with sparsity level p%, and LR denotes the low-rank method. Boxplots show posterior median estimates of the dependence parameter α obt… view at source ↗

**Figure 6.** Figure 6: High-dimensional study. Comparison of parameter estimation accuracy and computational cost across inference methods under four representative simulation settings. Boxplots show posterior median estimates of the dependence parameter α obtained from 100 simulated datasets, with the true parameter value (dashed red line). Blue crosses denote the walltime (in minutes) required for each method. Panels represent… view at source ↗

**Figure 7.** Figure 7: Estimated χu(h) extremal dependence curves for the NLDAS Tskin data, at spatial lag h = 0.177 (∼ 75 km). Results are shown for the four seasons (DJF, MAM, JJA, SON). The left and right panels correspond to Mat´ern covariance functions with smoothness parameters ν = 0.5 and ν = 1.5, respectively. Modeling approaches include the full Gaussian process (Full GP), the Vecchia approximation with conditioning set… view at source ↗

**Figure 8.** Figure 8: Top row: Observed (from GHCN) and predicted summer maximum 2-m air temperature at four validation sites. The LRSM-based predictions are obtained by first fitting the LRSM to Tskin, then propagating the resulting posterior predictions of Tskin through the trained XGBoost calibration model to obtain T2m. Bottom row: Signed prediction errors (predicted minus observed), where negative values indicate underest… view at source ↗

**Figure 9.** Figure 9: Posterior median fields of Tskin, T2m, and heat index (HI), and exceedance probabilities Pr(HI > 32◦C) for 2000. Rows show inference methods: Full GP (top), Vecchia (m = 5; middle), and covariance tapering (20% sparsity; bottom). The threshold Pr(HI > 32◦C) corresponds to the National Weather Service “extreme caution” category. Supplementary Figure S.15.1 displays pointwise 95% posterior credible interva… view at source ↗

**Figure 10.** Figure 10: Heat risk and socioeconomic vulnerability in the Four Corners region. Median [PITH_FULL_IMAGE:figures/full_fig_p029_10.png] view at source ↗

read the original abstract

Understanding and mapping extreme heat is critical for risk management and public health planning, particularly in regions with complex terrain and heterogeneous climate. We present a case study of extreme heat in the Four Corners region of the United States, using high-resolution surface skin temperature data from the North American Land Data Assimilation System to characterize spatially heterogeneous and seasonally varying extremes across complex terrain, and to assess their implications for heat-related public health risks. Spatial extremes exhibit complex dependencies across geographic regions, which require sophisticated statistical models to capture. While recent advances in spatial extreme value modeling provide flexible representations of joint tail dependencies, statistical inference remains computationally demanding, especially for datasets with a large number of locations. To address this, we propose a random scale mixture process that facilitates Bayesian inference of spatial extremes, and develop scalable inference strategies that leverage advances in spatial modeling and amortized learning. We evaluate the proposed inference methods through large-scale simulation studies, representing the first such extensive study in spatial extremes, and a high-resolution surface skin temperature application in the Four Corners region. Surface skin temperature is particularly useful as a predictor for air temperature, for studying heatwaves and related environmental phenomena, and to calculate heat indices reflecting downstream health risks at any location. Our findings provide insights into efficient, data-driven approaches for modeling spatial extremes, and serve as guidelines for practitioners in the fields of climate science, environmental risk assessment, and beyond.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper introduces a random scale mixture process with amortized inference to make Bayesian spatial extremes feasible at large scales, backed by simulations and a Four Corners heat application.

read the letter

The main thing to know is that this work proposes a random scale mixture process paired with amortized learning to handle Bayesian inference for spatial extremes on big datasets. It targets the computational bottleneck that has limited these models when the number of locations gets large, like in high-resolution climate grids. They back it with what they call the first extensive simulation study in the area and then apply it to surface skin temperature data from the Four Corners region to look at heat risk patterns across complex terrain. That practical tie-in to public health planning is a plus and shows the method in action on real heterogeneous data. The argument structure holds up internally: the motivation from existing limits is clear, and the application matches the stated goal of assessing spatially varying extremes. Soft spots are mostly around the missing details in the abstract—no equations, no specific performance numbers on tail dependence capture or runtime gains, and no clear checks for bias from the amortized step. Without those, it's difficult to confirm how well the mixture process avoids artifacts in the joint tails or whether the scalability claims hold under the actual data conditions. The stress-test note is right that no obvious assumption is violated on the surface, but the results section will need to deliver the evidence. This is aimed at spatial statisticians and climate risk modelers who need tools for large-scale extremes work. A reader focused on scalable Bayesian methods or environmental applications would get concrete value from the simulation design and the case study. I would send it for peer review—the computational gap it addresses is genuine and the setup looks worth a full look even if revisions are needed on the validation.

Referee Report

0 major / 2 minor

Summary. The manuscript proposes a random scale mixture process to enable Bayesian inference for spatial extreme value models, develops scalable inference strategies that combine advances in spatial modeling with amortized learning, evaluates the methods via large-scale simulation studies (described as the first extensive study of its kind), and applies the framework to high-resolution North American Land Data Assimilation System surface skin temperature data over the Four Corners region to characterize spatially heterogeneous extremes and assess implications for heat-related public health risks.

Significance. If the proposed process and inference strategies prove accurate and scalable, the work would address a recognized computational barrier in spatial extremes modeling, enabling Bayesian analyses at scales relevant to climate and environmental applications. The emphasis on extensive simulations and a concrete public-health case study strengthens practical utility and could provide useful guidelines for practitioners in climate science and risk assessment.

minor comments (2)

The abstract would be strengthened by including one or two quantitative highlights from the simulation studies (e.g., runtime reductions or coverage probabilities) to substantiate the scalability claims.
Clarify in the methods or results whether the random scale mixture process introduces any additional assumptions on tail dependence that are not directly validated against the surface skin temperature data.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of our manuscript, accurate summary of the proposed random scale mixture process and amortized inference strategy, and recommendation for minor revision. We are pleased that the significance for addressing computational barriers in spatial extremes modeling, the extensive simulation studies, and the public-health application to heat risk are recognized.

Circularity Check

0 steps flagged

No significant circularity; proposal is a new modeling framework evaluated externally

full rationale

The paper introduces a random scale mixture process and amortized inference strategies motivated by computational limitations of existing spatial extremes methods. It supports these via large-scale simulation studies (described as the first of their kind) and a real-data application to Four Corners surface skin temperature, without any equations or claims reducing predictions to fitted parameters by construction. No self-citation chains, uniqueness theorems, or ansatzes are invoked as load-bearing justifications for the core results. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit equations or implementation details, preventing identification of specific free parameters, axioms, or invented entities; the model is described at a high level only.

pith-pipeline@v0.9.0 · 5567 in / 1068 out tokens · 35988 ms · 2026-05-10T02:46:39.480780+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

176 extracted references · 8 canonical work pages

[1]

and Williams, A

Abatzoglou, John T. and Williams, A. Park , title =. Proc. Natl. Acad. Sci. U.S.A. , year =
[2]

Methods to calculate the heat index as an exposure metric in environmental health research , author=. Environ. Health Persp. , volume=
[3]

Environmetrics , volume=

Hu, Chenglei and Baltazar Bispo, Regina and Rue, H. Environmetrics , volume=. 2026 , publisher=

2026
[4]

Integral representations for

Askey, Richard and Fitch, James , journal=. Integral representations for
[5]

Baddeley, Adrian and Rubak, Ege and Turner, Rolf , year=
[6]

Banerjee, Sudipto and Carlin, Bradley P and Gelfand, Alan E , year=
[7]

Flux parameterization over land surfaces for atmospheric models , author=. J. Appl. Meteorol. Clim. , volume=
[8]

Berg, Christian and Mateu, Jorge and Porcu, Emilio , journal=. The. 2008 , publisher=

2008
[9]

Julia: A fresh approach to numerical computing , author=. SIAM. 2017 , publisher=

2017
[10]

The Annals of Applied Statistics , pages=

Spatial modeling of extreme snow depth , author=. The Annals of Applied Statistics , pages=. 2011 , volume=

2011
[11]

Spatially adaptive covariance tapering , author=. Spat. Stat. , volume=. 2016 , publisher=

2016
[12]

High--order composite likelihood inference for max--stable distributions and processes , author =. J. Comput. Graph. Stat. , volume =
[13]

Bayesian spatial modeling of extreme precipitation return levels , author=. J. Am. Stat. Assoc. , volume=. 2007 , publisher=

2007
[14]

and Johannesson, G

Cressie, N. and Johannesson, G. , date-added =. J. R. Stat. Soc. B. , number =
[15]

Basis-function models in spatial statistics , author=. Annu. Rev. Stat. Appl. , volume=. 2022 , publisher=

2022
[16]

Cressie, Noel and Wikle, Christopher K , year=
[17]

Geostatistics of dependent and asymptotically independent extremes , author=. Math. Geosci. , volume=. 2013 , publisher=

2013
[18]

A spectral representation for max-stable processes , author=. Ann. Probab. , pages=. 1984 , volume=

1984
[19]

Flexible space--time models for extreme data , author=. Spat. Stat. , volume=. 2025 , publisher=

2025
[20]

Climate change and changes in global precipitation patterns:

Dore, Mohammed H I , journal=. Climate change and changes in global precipitation patterns:. 2005 , publisher=

2005
[21]

Weather Clim

Detection and attribution of climate extremes in the observed record , author=. Weather Clim. Extrem. , volume=. 2016 , publisher=

2016
[22]

and Capon, Anthony and Berry, Peter and Broderick, Carolyn and de Dear, Richard and Havenith, George and Honda, Yasushi and Kovats, R

Ebi, Kristie L. and Capon, Anthony and Berry, Peter and Broderick, Carolyn and de Dear, Richard and Havenith, George and Honda, Yasushi and Kovats, R. Sari and Ma, Wei and Malik, Arunima and Morris, Nathan B. and Nybo, Lars and Seneviratne, Sonia I. and Vanos, Jennifer and Jay, Ollie , title=. The Lancet , year=
[23]

Evaluating station, satellite, & combined data for

Sch. Evaluating station, satellite, & combined data for. Atmos. Res. , pages=. 2025 , publisher=

2025
[24]

Extremes , volume=

Gradient boosting with extreme-value theory for wildfire prediction , author=. Extremes , volume=. 2023 , publisher=

2023
[25]

Extremes , volume=

Gradient boosting for extreme quantile regression , author=. Extremes , volume=. 2023 , publisher=

2023
[26]

Extremes , volume=

Extremal dependence of random scale constructions , author=. Extremes , volume=. 2019 , publisher=

2019
[27]

Covariance tapering for interpolation of large spatial datasets , author=. J. Comput. Graph. Stat. , volume=. 2006 , publisher=

2006
[28]

Journal of Multivariate Analysis , volume=

Compactly supported correlation functions , author=. Journal of Multivariate Analysis , volume=. 2002 , publisher=

2002
[29]

Stochastic models that separate fractal dimension and the

Gneiting, Tilmann , journal=. Stochastic models that separate fractal dimension and the. 2004 , publisher=

2004
[30]

Strictly Proper Scoring Rules, Prediction, and Estimation , author=. J. Am. Stat. Assoc. , volume=
[31]

Technometrics , volume =

Joseph Guinness , title =. Technometrics , volume =
[32]

and Schramm, Paul J

Hayden, Mary H. and Schramm, Paul J. and Beard, Charles B. and Bell, Jesse E. and Bernstein, Aaron S. and Bieniek-Tobasco, Ashley and Cooley, Nikki and Diuk-Wasser, Maria and Michael K. Dorsey and Ebi, Kristie L. and Ernst, Kacey C. and Gorris, Morgan E. and Howe, Peter D. and Khan, Ali S. and Lefthand-Begay, Clarita and Maldonado, Julie and Saha, Shubhay...
[33]

A case study competition among methods for analyzing large spatial data , author=. J. Agr. Biol. Envir. St. , volume=. 2019 , publisher=

2019
[34]

Space--time modelling of extreme events , author =. J. R. Stat. Soc. B. , volume =
[35]

Stat , volume =

Full likelihood inference for max--stable data , author =. Stat , volume =
[36]

Bridging asymptotic independence and dependence in spatial extremes using

Huser, Rapha. Bridging asymptotic independence and dependence in spatial extremes using. Spat. Stat. , volume=. 2017 , publisher=

2017
[37]

Modeling of spatial extremes in environmental data science: Time to move away from max--stable processes , author=. Environ. Data Sci. , volume=. 2025 , publisher=

2025
[38]

Stein and Peng Zhong , title =

Raphaël Huser and Michael L. Stein and Peng Zhong , title =. J. Comput. Graph. Stat. , volume =
[39]

Modeling spatial processes with unknown extremal dependence class , author=. J. Am. Stat. Assoc. , volume=. 2019 , publisher=

2019
[40]

WIREs Comput

Advances in statistical modeling of spatial extremes , author=. WIREs Comput. Stat. , volume=. 2022 , publisher=

2022
[41]

Land surface skin temperature climatology: Benefitting from the strengths of satellite observations , author=. Environ. Res. Lett. , volume=. 2010 , publisher=

2010
[42]

Fixed--width output analysis for

Jones, Galin L and Haran, Murali and Caffo, Brian S and Neath, Ronald , journal=. Fixed--width output analysis for. 2006 , publisher=

2006
[43]

Stationary max--stable fields associated to negative definite functions , author =. Ann. Probab. , volume =
[44]

Stat , volume=

Spatial extremal modelling: A case study on the interplay between margins and dependence , author=. Stat , volume=. 2024 , publisher=

2024
[45]

Vecchia approximations of

Katzfuss, Matthias and Guinness, Joseph and Gong, Wenlong and Zilber, Daniel , journal=. Vecchia approximations of. 2020 , publisher=

2020
[46]

Keetch, John James and Byram, George Marsden , year=
[47]

Space--time extremes of severe

Koh, Jonathan and Koch, Erwan and Davison, Anthony C , journal=. Space--time extremes of severe. 2024 , volume=

2024
[48]

Factor copula models for replicated spatial data , author=. J. Am. Stat. Assoc. , volume=. 2018 , publisher=

2018
[49]

Hydrolog

The implications of projected climate change for freshwater resources and their management , author=. Hydrolog. Sci. J. , volume=. 2008 , publisher=

2008
[50]

Biometrika , volume =

Statistics for near independence in multivariate extreme values , author =. Biometrika , volume =
[51]

Leeper and Tyler Harrington and Michael A

Ronald D. Leeper and Tyler Harrington and Michael A. Palecki and Kelley DePolt and Emma Scott and Jennifer Runkle and Howard J. Diamond. The Influence of Drought on Heat Wave Intensity, Duration, and Exposure. J. Appl. Meteorol. Clim. 2025. doi:10.1175/JAMC-D-24-0072.1

work page doi:10.1175/jamc-d-24-0072.1 2025
[52]

Lehmann, Erich L and Casella, George , year=
[53]

Neural networks for parameter estimation in intractable models , author=. Comp. Stat. Data An. , volume=. 2023 , publisher=

2023
[54]

and Rowhani, P

Lesk, C. and Rowhani, P. and Ramankutty, N. , title =. Nature , year =
[55]

Chapter 13: Transportation

Liban, Cris B and Kafalenos, Robert and Alessa, Lilian and Anenberg, Susan and Chester, Mikhail and DeFlorio, Joshua and D \'o \ n ez, Francisco J and Flannery, Aimee and Sanio, Michael R and Scott, Beverly A and Stoner, Anne Marie K. Chapter 13: Transportation
[56]

An explicit link between

Lindgren, Finn and Rue, H. An explicit link between. J. R. Stat. Soc. B. , year =
[57]

2023 , comment =

A deep learning synthetic likelihood approximation of a non--stationary spatial model for extreme streamflow forecasting , journal =. 2023 , comment =. doi:https://doi.org/10.1016/j.spasta.2023.100755 , author =

work page doi:10.1016/j.spasta.2023.100755 2023
[58]

Modeling extremal streamflow using deep learning approximations and a flexible spatial process , volume =

Majumder, Reetam and Reich, Brian J and Shaby, Benjamin A , journal =. Modeling extremal streamflow using deep learning approximations and a flexible spatial process , volume =
[59]

Chapter 2: Climate Trends

Marvel, Kate and Su, Wenying and Delgado, Roberto and Aarons, Sarah and Chatterjee, Abhishek and Garcia, Margaret E and Hausfather, Zeke and Hayhoe, Katharine and Hence, Deanna A and Jewett, Elizabeth B and Robel, Alexander and Singh, Deepti and Tripati, Aradhna and Vose, Russell S. Chapter 2: Climate Trends
[60]

Management Science , volume=

Scoring rules for continuous probability distributions , author=. Management Science , volume=. 1976 , publisher=

1976
[61]

The Annals of Statistics , volume=

What is a statistical model?\ (with discussion) , author=. The Annals of Statistics , volume=. 2002 , publisher=

2002
[62]

Bernoulli , volume=

Polynomial covariance functions on intervals , author=. Bernoulli , volume=. 2003 , publisher=

2003
[63]

Weather-Related Fatality and Injury Statistics , year =
[64]

A low-to-no snow future and its impacts on water resources in the western

Siirila-Woodburn, Erica R and Rhoades, Alan M and Hatchett, Benjamin J and Huning, Laurie S and Szinai, Julia and Tague, Christina and Nico, Peter S and Feldman, Daniel R and Jones, Andrew D and Collins, William D and others , journal=. A low-to-no snow future and its impacts on water resources in the western. 2021 , publisher=

2021
[65]

Nolan, John P , publisher=
[66]

Terminology in thermal infrared remote sensing of natural surfaces , author=. Agr. Forest Meteorol. , volume=. 1995 , publisher=

1995
[67]

Statistical post--processing of forecasts for extremes using bivariate

Oesting, Marco and Schlather, Martin and Friederichs, Petra , journal=. Statistical post--processing of forecasts for extremes using bivariate. 2017 , publisher=

2017
[68]

Remotely sensed land skin temperature as a spatial predictor of air temperature across the conterminous

Oyler, Jared W and Dobrowski, Solomon Z and Holden, Zachary A and Running, Steven W , journal=. Remotely sensed land skin temperature as a spatial predictor of air temperature across the conterminous
[69]

Likelihood--based inference for max-stable processes , author=. J. Am. Stat. Assoc. , volume=. 2010 , publisher=

2010
[70]

and Lewis, Sophie C

Perkins-Kirkpatrick, Sarah E. and Lewis, Sophie C. , title =. Nature Communications , year =
[71]

Asymptotic properties of the maximum in a stationary

Pickands, James , journal=. Asymptotic properties of the maximum in a stationary
[72]

Environmetrics , volume=

Fast parameter estimation of generalized extreme value distribution using neural networks , author=. Environmetrics , volume=. 2024 , publisher=

2024
[73]

The Annals of Applied Statistics , volume=

A hierarchical max--stable spatial model for extreme precipitation , author=. The Annals of Applied Statistics , volume=. 2012 , publisher=

2012
[74]

, year =

Resnick, Sidney I. , year =
[75]

Journal de la Soci

Spatial extremes: Max-stable processes at work , author=. Journal de la Soci
[76]

Statistica Sinica , volume=

Bayesian inference from composite likelihoods, with an application to spatial extremes , author=. Statistica Sinica , volume=. 2012 , publisher=

2012
[77]

evgam: An

Youngman, Benjamin D , journal=. evgam: An
[78]

and Huser, R

Richards, J. and Huser, R. , editor =. Extreme Quantile Regression with Deep Learning , booktitle =
[79]

Flexible modeling of nonstationary extremal dependence using spatially fused

Shao, Xuanjie and Hazra, Arnab and Richards, Jordan and Huser, Rapha. Flexible modeling of nonstationary extremal dependence using spatially fused. Technometrics , volume=. 2025 , publisher=

2025
[80]

Biometrika , pages=

Dependence modelling for spatial extremes , author=. Biometrika , pages=. 2012 , publisher=

2012

Showing first 80 references.