Establishing flood thresholds for sea level rise impact communication

High tide flooding (HTF) represents minor, recurrent coastal floods driven by elevated sea levels that disrupt daily activities and degrade infrastructure without causing major structural damage. Although single events are modest, their cumulative impacts pose growing socioeconomic and infrastructure challenges as SLR accelerates HTF frequency. Effective adaptation requires local-scale understanding because relative SLR deviates from global averages due to regional processes. However, coastal monitoring relies heavily on sparse, unevenly distributed tide gauges, leaving many U.S. coastal segments ungauged—about 75% of 10 km coastal intervals lack a gauge within 10 km. Current HTF thresholds are often derived from local tidal records and flood impact monitoring and are commonly referenced to mean lower low water (MLLW). Prior work (Sweet et al.) proposed a linear relation between HTF thresholds and greater tidal range (GTR), but thresholds above MHHW (the more logical datum for flood impact relative to local high tide) show that a univariate linear model fails to capture regional variability, especially where tidal ranges are large. This study addresses these gaps by developing an ML-based system to estimate HTF thresholds above MHHW and relative SLR at 10 km resolution along U.S. coasts, enabling spatially distributed thresholds and SLR rates suitable for ungauged locations and improved risk communication.

Past studies document increasing HTF occurrence along U.S. coasts driven primarily by SLR, with projections developed using hybrid process-based and statistical approaches that quantify uncertainty across climate scenarios. Still, widely applicable, interpretable metrics to monitor SLR impacts over space and time remain limited. Conventional thresholding combines tidal records, satellite altimetry, and local flood impact monitoring but is constrained by gauge sparsity and inconsistent impact severity representation across sites. Sweet et al. introduced a normalization and a linear relation between HTF above MLLW and GTR to standardize thresholds; however, measuring thresholds above MHHW reduces embedded tidal range effects and reveals that univariate linear regression inadequately captures spatial variability. Studies also highlight regional differences in SLR drivers (e.g., vertical land motion, ocean circulation, isostatic adjustments) and the importance of local processes, motivating regionally tailored, multi-feature approaches. Recent ML efforts have shown promise in predicting regional sea level changes and enabling high-resolution SLR estimates, suggesting potential to generalize HTF thresholding beyond point gauges.

The approach comprises clustering, regional ML model development, feature preparation, training/validation, and spatial application. K-means clustering (elbow method) partitioned coastal segments into three regions with similar dynamics: West Coast, Gulf and Southeast Coasts, and Northeast Coast. For each cluster, Random Forest (RF) regressors were developed to estimate (1) relative SLR rates and (2) HTF thresholds above MHHW (HTF_MHHW). Hyperparameters (e.g., number of estimators, max depth, bootstrap) were tuned via randomized search. Input features for SLR included reanalysis-based ocean variables (ocean heat, circulation, salinity, sea level pressure, surface pressure, sea surface temperature), and vertical land motion; latitude/longitude captured geographic gradients. HTF models used relative SLR, GTR, continental shelf slope, coastal elevation, latitude, and longitude; feature importance analyses led to removal of coastal elevation and shelf slope from HTF models to improve performance and efficiency. Coarse-resolution inputs were interpolated to target points using Empirical Bayesian Kriging, and multi-feature values were extracted at 10 km coastal points. Targets: SLR rates from Kopp et al. (50th percentile, year 2020, RCP 4.5 baseline) and HTF thresholds from NOAA/NWS and additional recent estimates at 100 gauges. Data were split randomly into 85% training and 15% validation, assuming stationarity of feature patterns over the study timeframe; validation used 15% held-out samples, repeated 100 times, reporting average metrics. Separate RF models per cluster were trained and validated for SLR and HTF_MHHW. The HTF ML predictions were compared against a univariate linear regression (LR) approach (HTF_MHHW = 0.04 × GTR + 0.5) to assess added value. Final trained models generated 10 km maps of spatially distributed SLR rates and HTF thresholds along the CONUS coastline.

• Clustering: Elbow method indicated three optimal clusters—West Coast, Gulf and Southeast Coasts, and Northeast Coast—consistent with regional still-water level dynamics and HTF behavior. - SLR estimation performance: Under current climate, RF regressors showed strong agreement with target SLR rates (Kopp et al.). Regional validation metrics: West (NSE 0.74, KGE 0.82, MAE 0.54 mm/yr, R^2 0.77), Gulf/SE (NSE 0.95, KGE 0.94, MAE 0.32 mm/yr, R^2 0.95), Northeast (NSE 0.80, KGE 0.85, MAE 0.29 mm/yr, R^2 0.85). The Northeast cluster exhibited the highest median SLR (~5.25 mm/yr) and smallest spread; Gulf/SE showed a wide, negatively skewed distribution with sites exceeding the median; West had a lower median (~2 mm/yr) with positive skew, including locations below the global average. - SLR feature importance: West dominated by latitude; Gulf/SE primarily by ocean circulation; Northeast by latitude and vertical land motion. Eliminating low-importance features degraded performance, indicating locally important contributions even from regionally minor variables. - HTF threshold estimation: ML predictions of HTF_MHHW matched observations substantially better than the LR method. In direct observed-vs-ML comparisons, Fig. 4a reported ML NSE 0.98, KGE 0.97, R^2 0.98; LR performed poorly (NSE -34.55, KGE -10.52, R^2 -0.08). Regional cross-validated averages for HTF ML: West (NSE 0.42, KGE 0.60, MAE 0.04 m, R^2 0.52), Gulf/SE (NSE 0.40, KGE 0.54, MAE 0.08 m, R^2 0.44), Northeast (NSE 0.30, KGE 0.48, MAE 0.13 m, R^2 0.36). - HTF feature importance: West—latitude most important; Gulf/SE—longitude then latitude; Northeast—latitude followed by GTR (which varies from ~5.8 m in northern Maine to ~1.7 m near Charleston, SC). SLR, longitude, and GTR had comparable importance on West and Northeast coasts. Removing coastal elevation and continental shelf slope improved HTF model performance and reduced computation time. - Spatial patterns at 10 km resolution: Highest relative SLR rates along Louisiana and east Texas (often 2–4× global average), consistent with literature; Pacific Northwest showed lowest SLR rates, partially due to isostatic adjustments. HTF thresholds reflected regional variability and local flood defenses (e.g., higher thresholds near Texas hurricane seawalls and New York). - Overall: The system provides ungauged-coast estimates of SLR and HTF thresholds at 10 km spacing, complementing point-based gauges and enabling improved communication of chronic HTF risks. Training and validation against NOAA data yielded promising skill, with an average KGE reported as 0.77 in the abstract.

The study demonstrates that a multi-feature, regionally trained ML framework can overcome limitations of univariate, point-based approaches for coastal HTF thresholding. By estimating HTF thresholds above MHHW and explicitly accounting for spatially varying drivers (SLR, GTR, geography, circulation, VLM), the method captures regional and local variability that linear regression on GTR alone misses—particularly in high tidal range regions such as the Northeast. The results provide operationally relevant, spatially distributed thresholds and SLR rates, improving communication of chronic HTF risks and enabling more informed adaptation planning, infrastructure management, and insurance risk assessment. Spatial heterogeneity in HTF thresholds near metropolitan areas underscores the influence of human interventions (e.g., seawalls, mitigation in New York) and highlights inconsistencies in NOAA’s impact-based thresholds when extrapolated alongshore. Feature importance patterns corroborate known regional controls (e.g., circulation in Gulf/SE, isostatic/VLM influences in Northeast/West), while maps align with literature-reported hotspots (Louisiana/Texas) and low-rate regions (Pacific Northwest). Collectively, these findings support shifting from impact-only or univariate methods to regionally tailored, multi-variable, nonlinear approaches for consistent, scalable HTF impact monitoring and communication.

This work advances coastal flood risk communication by delivering 10 km, spatially distributed estimates of relative SLR rates and HTF thresholds above MHHW along U.S. coasts using clustered, region-specific Random Forest models. The approach addresses gauge sparsity, improves upon linear regression methods, and captures regional drivers of both SLR and HTF threshold variability. The resulting products can support policymakers, emergency managers, and insurers in risk-informed decision-making and adaptation planning by documenting the chronic HTF signal. Future work should incorporate dynamically downscaled inputs, improved observations in poorly gauged regions (e.g., Louisiana), explicit representation of flood defenses, and comprehensive dynamical modeling to refine feature sets and reduce interpolation uncertainties. Extending the methodology beyond the CONUS and integrating evolving climate scenarios can further generalize the framework for global application.

• Sparse and uneven tide gauge coverage limits training data representativeness, particularly in regions like Louisiana where few long records constrain learning of VLM contributions. - Input datasets are often coarse and require spatial interpolation (Empirical Bayesian Kriging), introducing uncertainty into local estimates. - Flood defense data were not available for inclusion, yet such measures can substantially alter local HTF thresholds and complicate alongshore generalization. - NOAA impact-based thresholds can be inconsistent across sites in terms of impact-depth relationships, challenging spatial extrapolation and validation. - Assumption of stationary feature patterns for splitting/validation may not fully capture temporal evolution. - Feature importance is regionally variable; excluding weak regional features can degrade local performance, indicating potential over-sensitivity to local conditions and the need for richer, localized data. - Linear regression baselines are limited for comparison, but ML metrics vary across clusters, with lower cross-validated skill for HTF thresholds in some regions relative to SLR.