Historical Analysis of 24-Hour Storms in the Upper Midwest

Andy Erickson; Noah Gallagher; John Gulliver

Extreme precipitation estimates, particularly for rare storms with large return periods like the 100-year storm, are a useful tool for watershed managers and engineers to evaluate the flood risk for their communities. While national level estimates such as Technical Paper 40 (Hershfield, 1961) and Atlas 14 (Perica et al., 2013) are publicly available, these estimates do not consider temporal dynamics (i.e., non-stationarity) of extreme precipitation. Identifying any temporal trends at a single weather station can be challenging because most modern observation records extend fewer than 150 years, which is insufficient to run statistical tests or identify trends over time for rare events. This presentation examines the evidence for both long term trends in extreme precipitation frequency, and any breakpoints in such trends using long term precipitation records from over 1,000 stations in the Midwest USA. First, we identify non-stationarity in the aggregate frequency of occurrence of top 50% and top 10% annual maximum precipitation events, as well as in the occurrence of largest recorded event at each station. Second, our results show that a piecewise linear fit with two inflection points occurring during the years 1920 and 1974 accurately models observed extreme event non-stationarity. The Dustbowl period in the American Midwest explains the 1920 breakpoint and the 1974 breakpoint coincides with surface temperature increases for the continental United States. We will discuss these findings.

Uncertainty Monsters and Nonstationary Flood-Frequency Analysis

Karen Ryberg

Flood-frequency analysis is undergoing a paradigm shift. Traditionally grounded in stationary assumptions, the field now faces growing pressure to adopt nonstationary methods that account for climate and land-use change. This shift is not just technical—it is philosophical, methodological, and deeply uncertain. Drawing on consulting experience, I examine how practitioners navigate these uncertainties using the “uncertainty monster” metaphor, which captures the tension between competing analytical frameworks. Originally developed in the Netherlands, the “uncertainty monster” metaphor describes situations where mutually exclusive categories, knowledge and ignorance, facts and values, prediction and speculation, stationary and nonstationary, coexist and create tension. In flood-frequency analysis, the monster emerges in the choice between stationary and nonstationary assumptions, each with distinct methodological and interpretive challenges (more monsters). Six strategies for dealing with uncertainty monsters have been identified: denial, exorcism, adaptation, assimilation, embracement, and anesthesia. These are not just theoretical—they are visible in practice. Denial persists in exclusive reliance on Bulletin 17C; exorcism in efforts to eliminate uncertainty through complex models; adaptation in scenario-based approaches; assimilation in rethinking foundational categories; embracement in celebrating complexity; and anesthesia in consensus-driven decisions. The role of consultants is to reduce the monster—not by denying or exorcising it, but by translating complexity into actionable insight. This presentation draws on case studies to illustrate how adaptation and assimilation strategies can be utilized in flood-frequency analysis.

Bayesian Reconstruction of Rainfall Extremes: Impacts on Infrastructure Sizing Across 45 U.S. Cities

Omid Emamjomehzadeh; Dawar Qureshi; Lauren Cook; Omar Wani

Urban drainage infrastructure is traditionally designed using Intensity-Duration-Frequency (IDF) curves derived through classical statistical methods, such as Maximum Likelihood Estimation (MLE) and the method of linear moments (L-moments). However, MLE and L-moments provide only point estimates and fail to adequately account for parametric uncertainty—especially under changing climatic conditions and when dealing with a limited number of samples, such as those related to precipitation extremes. This study presents a Bayesian framework for analyzing extreme rainfall statistics, employing Hamiltonian Monte Carlo techniques to derive full posterior distributions of Generalized Extreme Value parameters for both historical and projected rainfall data across 45 U.S. cities. The Bayesian approach produces systematically higher rainfall level estimates compared to MLE and L-moments, particularly for return periods beyond 25 years. These differences have a significant impact on the design of stormwater infrastructure. Using a rational method-based pipe sizing analysis, we demonstrate that Bayesian-informed IDF values increase pipe diameter requirements in several cities, in both historical and projected climate scenarios. We also demonstrated that, in repetitive control experiments assuming stationary, noisy stationary, and noisy nonstationary rainfall processes, the Bayesian method yields less negative bias and less undersized drainage networks. Additionally, in most of the trials, it produces less absolute bias than the MLE and L-moments methods. The results highlight the inadequacy of relying solely on point estimates in the face of intensifying extreme events.

Discharge frequency-magnitude analysis of extreme weather events under future conditions

Lia Vergara-Pena; Felix Santiago Collazo; John Ramirez-Avila

Coastal communities in low-gradient basins face escalating flood hazard as tropical cyclones (TCs) become more frequent and intense under climate change. A key challenge in flood risk management is the resulting hazard shift, or return-period compression, in which events previously considered “exceptionally unlikely” occur more frequently. Furthermore, existing conventional hydrologic and statistical frameworks fall short at simulating unprecedented extreme events. Therefore, this study develops a semi-distributed hydrologic model (HEC-HMS) for the upper part of the St. Marys River watershed, located on the eastern border between Florida and Georgia, driven by spatiotemporal variable rainfall generated by a Parametric Precipitation Model (P-CLIPPER) to produce rainfall fields from 500 synthetic TC, for each scenario, present and future conditions. This approach enables the representation of the upper tail of the discharge distribution at very low annual exceedance probabilities (AEPs). Projected changes in Land Use/Land Cover and TC rainfall are incorporated to evaluate future climate and anthropogenic change scenarios. To quantify hazard shift, simulated peak discharges were fitted to six cumulative distribution functions, allowing the estimation of AEPs beyond the 0.1% threshold without relying solely on limited historical gauge records. The study also evaluates whether calibrated and validated hydrologic parameters remain applicable under extreme storms predicted for future scenarios. These findings support the development of flood hazard maps that account for environmental and anthropogenic changes, enhancing infrastructure reliability and long-term resilience under increasing flood frequency and intensity.