Fusing Sparse Observations and Dense Simulations for Spatial Extreme Value Analysis: Application to U.S. Coastal Sea Levels

Brian N. White, Brian Blanton, Rick Luettich, Richard L. Smith, [6pt] Department of Statistics and Operations Research, University of North Carolina at Chapel Hill, [6pt] Correspondence: [email protected] · 2026

Read Fusing Sparse Observations and Dense Simulations for Spatial Extreme Value Analysis: Application to U.S. Coastal Sea Levels by Brian N. White, Brian Blanton, Rick Luettich, Richard L. Smith, [6pt] Department of Statistics and Operations Research, University of North Carolina at Chapel Hill, [6pt] Correspondence: [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03247.txt.

Introduction

Brian White, Brian Blanton, Rick Luettich, and Richard Smith fuse sparse tide gauges with dense simulations to map coastal flood risk. Extreme sea level estimates from a thin gauge network carry wide uncertainty, yet physics-based simulators produce detailed output everywhere. Their 2026 two-stage framework combines both sources for the United States coast.

First, extreme value distributions are fitted at each site, allowing trends over time. Then a spatial model links the fitted parameters across sources, with correlations between interspersed gauge and simulation sites carrying information without any shared locations. Applied to four decades of records from twenty-nine gauges and one hundred simulation sites, fusion cut the error of hundred-year flood estimates by about a third versus gauges alone. Block validation confirms the gain survives extrapolation.

Catalog record

CREATORBrian N. White, Brian Blanton, Rick Luettich, Richard L. Smith, [6pt] Department of Statistics and Operations Research, University of North Carolina at Chapel Hill, [6pt] Correspondence: [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51937
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.03247

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Fusing Sparse Observations and Dense Simulations for Spatial Extreme Value Analysis: Application to U.S. Coastal Sea Levels Brian N. White, Brian Blanton, Rick Luettich, Richard L. Smith, [6pt] Department of Statistics and Operations Research, University of North Carolina at Chapel Hill, [6pt] Correspondence: [email protected] Abstract Estimating spatial extremes from sparse observational networks produces uncertain return level maps, but dense output from physics-based simulation models is often available as a complementary data source. We develop a two-stage frequentist framework for fusing observations and simulations. In Stage 1, generalized extreme value (GEV) distributions are fitted independently at each site, with a nonstationary location parameter where appropriate to accommodate observed trends. In Stage 2, the parameter estimates from all sources are modeled jointly as a high-dimensional spatial process through a linear model of coregionalization (LMC). Cross-source correlations, estimated from spatially interspersed networks without co-located sites, provide the mechanism for information transfer; an analytic gradient for the resulting likelihood keeps estimation computationally practical. We apply the framework to U.S. coastal sea levels over 1979–2021, fusing 29 NOAA tide gauge records with 100 ADCIRC hydrodynamic simulation sites. Leave-one-out cross-validation shows a 35% reduction in 100-year return level RMSE relative to a gauge-only model. Geographic block cross-validation confirms that fusion benefits persist under spatial extrapolation. The approach is implemented in the R package evfuse . Keywords: data fusion, coregionalization, return levels, storm surge, geostatistics, cross-validation 1 Introduction Extreme value distributions vary in space, but observational networks are seldom dense enough to estimate this variation precisely. Physics-based simulation models can produce the quantity of interest at far greater spatial density, and fusing their output with sparse observations can improve both the coverage and precision of return level estimates, the T -year thresholds expected to be exceeded on average once every T years. We develop a general framework for this fusion and demonstrate it by combining tidal gauge observations with hydrodynamic simulation output to map extreme coastal sea levels along the U.S. East and Gulf Coasts. The two data sources exemplify a structure common in environmental monitoring. The NOAA tidal gauge network provides reliable long-term sea level records, but gauges are sparse and their records are short relative to the return periods of interest, producing substantial uncertainty in tail parameter estimates at individual sites. Physics-based hydrodynamic models such as ADCIRC ( Luettich et al. 1992 ) can simulate sea levels at hundreds of coastal locations, providing the spatial density that gauges lack, but their output reflects model physics rather than direct observation. Statistical approaches to spatial extremes include max-stable processes and hierarchical GEV models with spatially varying parameters ( Davison et al. 2012 ; Cooley et al. 2007 ; Sang & Gelfand 2010 ) . Max-stable models are theoretically appealing and capture the spatial extent of extreme events, but their likelihood is intractable for even moderate numbers of stations. When the primary interest is marginal return levels rather than joint tail dependence, latent process models are often preferred: the GEV parameters at each location are treated as realizations of a spatial process, and inference proceeds either through Bayesian hierarchical modeling ( Cooley et al. 2007 ; Sang & Gelfand 2010 ; Gelfand et al. 2004 ; Finley et al. 2008 ) or through a two-stage frequentist approach where pointwise MLEs are smoothed spatially ( Russell et al. 2020 ) . Data fusion combining in-situ observations with gridded model output or remote sensing has a rich history in spatial statistics ( Cressie & Wikle 2011 ; Nguyen et al. 2012 ; Berliner 2003 ) . Despite this, fusion methods have rarely been applied to extreme value parameters, where the goal is to combine heterogeneous sources to improve tail inference rather than to interpolate mean fields. We adopt a two-stage frequentist approach for computational tractability and interpretability, building on Russell et al. 2020 , who combined a linear model of coregionalization (LMC) with two-stage frequentist inference for spatially varying GEV parameters, applied to a single network of Gulf Coast precipitation stations. Our contribution extends this framework in two directions. First, we generalize the model to represent GEV parameters from multiple data sources as a single high-dimensional spatial process. The LMC is identifiable under heterotopic sampling ( Wackernagel 2003 ) , so cross-source correlations can be estimated from spatially interspersed networks without co-located sites. A selection matrix extracts the observed components from the full latent covariance, yielding a standard Gaussian likelihood. An analytic gradient for the resulting 33-parameter log-likelihood, derived from the LMC structure (Section S1), reduces optimization from approximately 90 minutes to under 3 minutes on a desktop workstation, making multi-start optimization and bootstrap inference practical. Second, we provide cross-validation evidence quantifying the gain from fusion. Under a nonstationary Stage 1 model that accommodates observed trends, the joint model reduces 100-year return level LOO-CV RMSE by 35% relative to a NOAA-only baseline. Geographic block cross-validation shows that the benefit persists under spatial extrapolation, though at reduced magnitude. 2 Data The spatial domain extends from southern Texas ( 26.1^{\circ} N, 97.2^{\circ} W) to northern Maine ( 44.9^{\circ} N, 67.0^{\circ} W). We analyze annual maximum sea levels over 1979–2021 ( T_{0}=43 years) from two sources. 2.1 NOAA Tidal Gauge Observations We obtained hourly sea-level measurements from the NOAA Tides & Currents portal in May 2022. All measurements are in meters relative to the mean sea level (MSL) tidal datum (1983–2001 epoch). We applied a two-step quality filter: (i) any year with fewer than 90% of expected hourly observations was excluded; (ii) any station with fewer than 20 annual maxima after step (i) was omitted. This yielded L_{N}=29 stations, each with at least 34 annual maxima (24 of 29 exceeding 40). The 90% threshold guards against missing the true annual maximum. These records are short relative to the return periods of interest (100 years): estimating the shape parameter \xi , which controls tail behavior, from 34–43 annual maxima produces substantial uncertainty at individual sites, motivating spatial borrowing of strength. If tidal gauges fail during extreme events, the resulting censorship of the largest observations would bias \hat{\xi} downward. This provides additional motivation for fusion: simulations are not subject to instrument failure, and the cross-source shape correlation in Stage 2 can mitigate such bias. 2.2 ADCIRC Simulated Data The Advanced Circulation (ADCIRC) model is a finite-element hydrodynamic model that solves the shallow water equations on unstructured grids ( Luettich et al. 1992 ; Dietrich et al. 2011 ) . Simulated hourly sea-level values over 1979–2021 at L_{A}=100 coastal sites were generated by the Renaissance Computing Institute (RENCI) at the University of North Carolina at Chapel Hill, driven by CFSR reanalysis atmospheric forcing. These are historical reconstructions generated from model physics alone. Although ADCIRC operates on unstructured grids with thousands of nodes, we use 100 representative coastal sites, a density sufficient to resolve the large-scale spatial patterns of interest along 6,000 km of coastline. Finer-resolution questions would require a denser draw from the ADCIRC grid and corresponding extensions to the spatial model (Section 5.4 ). Because the record is complete by design, all 43 annual maxima are available at each site. 2.3 Study Design The total number of sites is L=L_{N}+L_{A}=129 , with no site having both NOAA and ADCIRC observations. The ADCIRC sites fill spatial

Read the complete text