Single Shift Segmentation Improves Moderate Flood Estimates under Nonstationary Conditions across the United States

Precipitation, particularly at high quantiles, has been reported to increase in various regions across the globe, raising pluvial flood risk. One of the main challenges in reliable flood frequency analysis is handling nonstationarity arising from climate variability or anthropogenic disturbances such as land use/cover change or river regulation. To separate these nonstationary footprints, we analyzed annual maximum peak flow records from 18 reference (minimally disturbed) and 66 non-reference stream gages, each with more than 100 years of flood records across the United States. Next, we used a nonparametric Pettitt test to identify statistically significant change points. When present, the flood record was split into pre- and post-change segments with a Log-Pearson III distribution fitted to each. Depending on the region and site type, using a segmented record improved the quantile estimate. At the majority of reference sites, post-change data produced the highest flood quantiles, reflecting recent climate-driven nonstationarity. Conversely, at several non-reference sites, pre-change data returned larger estimates, indicating that long-standing anthropogenic disturbances can attenuate the signal of climatic variations. Our study confirms that fitting a flood frequency model to the segment that minimizes nonstationarity, rather than the entire record, returns more reliable estimates for moderate flood magnitudes of up to a 25-year return interval. The approach highlights the need to understand the population from which flood records are extracted, to separate those populations where appropriate, and then fit a statistical distribution. This practical approach offers a simple thought process for updating moderate flood forecasts to guide infrastructure design or rehabilitation in the current dynamic environment, an era of constant change that needs flexibility in everything we design.