Floods across the eastern United States are projected to last longer

The eastern United States is one of the most flood-prone regions in the country due to diverse generating mechanisms, including heavy precipitation, snowmelt, and tropical and extratropical cyclones. Recent billion-dollar flood disasters have disproportionately impacted this region, and with climate change and population growth expected to increase exposure, understanding how flood characteristics—magnitude, frequency, and especially duration—may change is critical for preparedness and risk mitigation. Prior work has largely focused on annual maximum peak discharge, which can mask true flooding because annual maxima may not exceed flood thresholds. Peaks-over-threshold (POT) approaches better capture meaningful flood behavior but have been less leveraged for projecting changes in duration. This study fills the gap by applying an attribution-and-projection framework to seasonal flood duration at 378 USGS stations across the eastern U.S., focusing on NWS minor and moderate flood levels due to data limitations for major floods. The authors develop season- and site-specific statistical models linking seasonal flood duration to concurrent and antecedent precipitation and temperature, and then use CMIP6 projections under multiple SSPs to assess future changes.

Most literature has examined changes in flood magnitude using annual maxima, which can obscure true flooding when values do not exceed flood thresholds. POT analyses have shown stronger evidence of increasing flood frequency and can reveal changes concealed by annual maxima. Although trends in POT-driven characteristics (frequency, duration) have been studied, extrapolating observed trends is challenging because they may not persist. A promising approach is to establish statistical relationships between flood characteristics and physical drivers (e.g., precipitation and temperature) and then use projected drivers for future assessments. Previous studies have applied such attribution-and-projection frameworks to flood magnitude and frequency in various regions, underscoring the value of climate-informed flood projections. Building on this, the present study targets flood duration, a relatively understudied characteristic at regional scale.

Data and study area: The analysis covers USGS streamgages across the eastern U.S. Daily mean discharge series (1949–2019) were used. Stations were screened by: (1) availability of basin boundary data (USGS Streamgage NHDPlus Version 1 Basins 2011); (2) years/seasons with <10% missing daily observations; (3) minimum of 30 years of observations in 1980–2019. Stations were classified into reference and non-reference groups using GAGES-II to assess human-disturbance effects. Flood thresholds and durations: NWS provides stage/discharge thresholds for minor, moderate, and major flooding; thresholds and rating curves were obtained per station. For each station and severity level, discharge thresholds were used (or stages converted via rating curve). Seasonal flood duration was defined as the count of days per meteorological season—fall (Sep–Nov), winter (Dec–Feb), spring (Mar–May), summer (Jun–Aug)—with mean daily discharge exceeding the severity-specific threshold. Flood events spanning seasons were split into corresponding seasons. Climate drivers: Basin-averaged precipitation and temperature were derived from PRISM monthly products (2.5 arcmin resolution). Monthly values were aggregated/averaged to seasonal scales to form four predictors per season: concurrent precipitation (Pcon), concurrent temperature (Tcon), lagged precipitation from the prior season (Plag), and lagged temperature (Tlag). Statistical attribution models: Seasonal flood duration (number of flood days) was modeled using GAMLSS with a binomial distribution for counts within a season (n = number of days in the season). The occurrence probability of flooding (μ) was linked to predictors via a logit link: log(μ/(1−μ)) = μ0 + μ1 Pcon + μ2 Tcon + μ3 Plag + μ4 Tlag. For each site, severity level, and season, models were developed only if at least five years had nonzero flood duration; otherwise, no model was fit and future flooding was assumed absent. All linear combinations of the four predictors were considered; the best model was selected using Schwarz’s Bayesian criterion (BIC). Performance was evaluated by the Spearman correlation between observed and modeled seasonal flood durations (median of fitted distribution). Leave-one-out cross-validation (LOOCV) assessed extrapolative skill; for each year, the model was refit excluding that year and used to predict the excluded year. GCM data and bias correction: Monthly precipitation (kg m−2 s−1) and near-surface air temperature (K) were obtained from 36 CMIP6 GCMs with nominal spatial resolution ≤250 km for historical and SSP1-2.6, SSP2-4.5, SSP3-7.0, SSP5-8.5 future scenarios. Basin averages were computed, converted to monthly totals and °C as needed, and empirically bias-corrected by month via quantile mapping using PRISM as reference. Bias-corrected series were aggregated to seasonal scales and input to the fitted attribution models to generate seasonal flood durations for historical and future periods. GCM selection for projections: To evaluate GCM suitability, for each site and severity level the time series of annual flood durations (from PRISM-driven models and each GCM-driven model) for 1951–2014 were used to fit nonstationary binomial models with a time covariate: log(μ/(1−μ)) = μ0 + μ1 t. A GCM-site pair was classified as Match if trends μ1 from GCM and PRISM were both significant with the same sign or both not significant at 5%; otherwise Mismatch. GCMs with >60% Match sites for both minor and moderate flooding were retained, yielding 15 selected GCMs. Ensemble medians of seasonal and annual flood durations were computed from selected GCMs. Significance testing of projected shifts: One-sided Student’s t-tests (α=0.05) assessed significant increases or decreases in flood durations between 1985–2014 (historical) and 2071–2100 (future) for each season and annually. Results were also stratified by reference vs non-reference sites.

- Statistical attribution performance: Across seasons and for both minor (378 sites) and moderate (161 sites) NWS flood levels, models captured interannual variability well, with median Spearman correlations ~0.6 between observed and modeled seasonal durations; annual aggregations performed particularly well in the Coastal Plain. Cross-validation median correlations were ~0.5 across seasons and severities. Reference and non-reference basins showed similar performance (differences in median correlation < ~0.06), indicating dominant control by climate drivers. - Dominant predictors: Concurrent precipitation was selected at nearly all sites in all seasons with a positive contribution to flood occurrence probability. Lagged precipitation was regionally variable, more often selected in the Coastal Plain. Temperature was less frequently selected; in the Coastal Plain during spring and summer it often had a negative effect, consistent with higher evapotranspiration reducing soil moisture and flood likelihood. Results were similar for moderate flooding. - GCM evaluation and selection: Of 36 CMIP6 GCMs considered, 15 were retained for projections based on reproducing the sign/significance of historical trends in annual flooding occurrence probability (Match rate >60% for both severities). - Projected seasonal shifts (2071–2100 vs 1985–2014): For minor flooding, winter shows the highest share of sites with significant increases in flood duration (18.7–53.3% depending on SSP), followed by fall (12.6–21.2%); other seasons show increases at <16% of sites. Decreases are more common in spring and summer (12.2–26.8% of sites across scenarios) and limited in fall and winter (≤7.1%). Moderate flooding shows similar spatial-seasonal patterns with somewhat lower percentages. - Emissions dependence: Higher-emission scenarios lead to more sites with significant increases, especially longer winter durations. Annually, the percentage of sites with significantly longer flood durations increases with emissions from 26.2% to 47.6% (minor) and from 19.3% to 32.3% (moderate). Reference vs non-reference sites display similar seasonal and scenario-dependent behaviors.

By linking seasonal flood durations to basin-averaged concurrent and antecedent precipitation and temperature, the study directly addresses how climate drivers control interannual variability in flood duration. The models’ robust performance across diverse basins and their limited sensitivity to human modifications suggest that seasonal climate conditions are the primary determinants of flood duration. The dominance of concurrent precipitation across seasons and locations evidences its central role, while the regional and seasonal influence of temperature (often negative in warm-season Coastal Plain) aligns with process understanding of evapotranspiration-soil moisture interactions. Using CMIP6 projections with a stringent GCM selection protocol, the study indicates widespread lengthening of flood durations—most prominently in winter, the primary flood season in much of the eastern U.S.—and under higher emission scenarios. These longer durations have implications for risk management because duration contributes to both direct damages and indirect health and societal impacts. The findings thus reinforce the importance of both climate mitigation (to limit the magnitude of changes) and adaptation (e.g., structural measures, planning) to manage increasing flood duration risks.

This work develops and validates season- and site-specific statistical attribution models for flood duration across the eastern U.S., demonstrating that seasonal precipitation (especially concurrent) and, to a lesser extent, temperature explain much of the interannual variability in days above NWS minor and moderate flood thresholds. Projections using selected CMIP6 models under multiple SSPs indicate that floods are expected to last longer, particularly in winter and under higher emissions, with meaningful fractions of sites showing significant increases in duration at seasonal and annual scales. The approach is flexible and transferable to other regions or individual gages wherever flood severity thresholds are defined, enabling climate-informed assessments of future flood duration. Potential future work includes applying dominance analysis to quantify the relative contribution of each predictor, extending the framework to additional hydrometeorological drivers where available (e.g., soil moisture, snow metrics), and broadening the spatial scope beyond the eastern U.S.

- Modeling threshold: Models were only fit where at least five years had nonzero seasonal flood duration; otherwise, no-flooding was assumed. This could underestimate the fraction of sites projected to experience increasing durations. - Data availability for severities: Major floods were excluded due to limited station-years above major thresholds, constraining conclusions to minor and moderate levels. - Predictor set: Only precipitation and temperature were used as predictors due to their consistent availability across GCMs; other potentially relevant variables (e.g., soil moisture, evapotranspiration, snowmelt) were not included. - GCM dependence: Although a screening ensured reasonable historical trend representation, projections still depend on GCM biases and scenario uncertainties. - Seasonal aggregation: Using seasonal means/totals may miss sub-seasonal processes that influence flood duration at event scales.