Search

Accurate flood risk assessment requires a comprehensive understanding of flood sensitivity to regional drivers and climate factors. This paper presents the scaling of floods (duration, peak, volume) with geomorphologic characteristics of the basin (i.e., drainage area, slope, elevation) and precipitation patterns (rainfall accumulation, variability). Long‐term daily streamflow observations over the 20th and early 21st centuries from Hydro‐Climatic Data Network streamgages across the conterminous United States are used to create a flood event database based on their flood stage information. Antecedent daily rainfall accumulation and variability corresponding to these floods are computed using Global Historical Climatology Network daily data set. Two Bayesian scaling models are developed, and the spatial organization of scaling exponents is investigated. The baseline model quantifies the scaling of floods to geomorphologic characteristics. The dynamic model quantifies the scaling of floods to antecedent precipitation distribution which is further conditioned on geomorphologic characteristics. Results show that small and low‐elevation basins have a stronger response to antecedent rainfall distribution in amplifying flood peaks, while high‐elevation steeper basins have a lower response for flood duration and volume. The dynamic models demonstrate that there are significant variations in the flood scaling rates, with the largest rates up to 40% and 4.5% for flood duration, 64% and 44% for peak, and 98% and 40% for volume found across the Northeast, Coastal Southeast, and Northwest with intensifying rainfall accumulation and variability, respectively. This study advances flood predictions by better informing the flood attributes in the context of dynamical land‐atmosphere perturbations. Plain Language Summary: The relationships between floods, physical properties of river basins, and the spatio‐temporal distribution of precipitation define the regional flood scaling laws. Improved regional flood scaling models are required for updating flood risk scoring tools to help insurers and decision‐makers in the light of anthropogenic climate changes and natural variability of Earth systems. Two scaling models are developed here to examine how the physical properties of basins and precipitation patterns can inform flood attributes across the conterminous United States (US). Using long records of streamflow and precipitation observations, we identified different scaling relationships between the catchment factors, preceding precipitation, and the duration, peaks, and volume of floods. Larger basins located at higher elevations are found to have lower scaling rates with flood peaks when interacting with the preceding precipitation patterns. Overall, low‐elevation flat basins across the Eastern, Southern, and Northwestern US have a larger positive response to precipitation in amplifying flood attributes. These new scaling laws can help advance our understanding of controlling mechanisms for flash floods and mega‐floods, which is required for a dynamic flood risk assessment in a changing environment. Key Points: Two Bayesian scaling models are developed to understand the sensitivity of floods to geomorphologic characteristics and catchment rainfallLow‐elevation and flat basins with different sizes have greater scaling to rainfall in amplifying flood duration and volumesLow‐elevation and small basins with different steepness show larger sensitivity to rainfall distribution that enhances flood peaks [ABSTRACT FROM AUTHOR]

Trends in flood magnitudes vary across the conterminous USA (CONUS). There have been attempts to identify what controls these regionally varying trends, but these attempts were limited to certain—for example, climatic—variables or to smaller regions, using different methods and datasets each time. Here we attribute the trends in annual maximum streamflow for 4,390 gauging stations across the CONUS in the period 1960–2010, while using a novel combination of methods and an unprecedented variety of potential controlling variables to allow large‐scale comparisons and minimize biases. Using process‐based flood classification and complex networks, we find 10 distinct clusters of catchments with similar flood behavior. We compile a set of 31 hydro‐climatological and land use variables as predictors for 10 separate Random Forest models, allowing us to find the main controls the flood magnitude trends for each cluster. By using Accumulated Local Effect plots, we can understand how these controls influence the trends in the flood magnitude. We show that hydro‐climatologic changes and land use are of similar importance for flood magnitude trends across the CONUS. Static land use variables are more important than their trends, suggesting that land use is able to attenuate (forested areas) or amplify (urbanized areas) the effects of climatic changes on flood magnitudes. For some variables, we find opposing effects in different regions, showing that flood trend controls are highly dependent on regional characteristics and that our novel approach is necessary to attribute flood magnitude trends reliably at the continental scale while maintaining sensitivity to regional controls. Key Points: A wide variety of controls are necessary to explain flood magnitude trends across the United States between 1960 and 2010Climatic changes and land cover conditions are of similar importance for flood magnitude trends at the regional scaleControls on flood trends can have highly nonlinear effects and can have opposing effects in different hydro‐climatological subregions [ABSTRACT FROM AUTHOR]

Many studies investigate river floods by analyzing annual maximum series that record the largest flow of each year, including many within‐bank events inconsequential for human communities. Fewer focus on larger floods, especially at the continental scale. Using 473 streamgages across the conterminous United States with near‐natural flow from 1966 to 2015, we characterized the seasonality, occurrence, and climatic associations of the 10 largest and 2 largest floods at each site. These are often overbank events that have ecosystem functions and pose risks to humans. We grouped sites into 14 clusters corresponding to climatic and physiographic regions and characterized their flood seasonality at a monthly resolution using a probabilistic method. We then evaluated annual occurrence regionally and nationally, and seasonal occurrence regionally, by complementing a traditional approach to trend analyses with a novel method based on expected occurrence. Relationships between flood occurrence and climate indices were also investigated. Large floods have strong seasonality in some regions, but in areas with numerous flood‐generating mechanisms, seasonality is more complex. There is little evidence nationally that large riverine floods are more or less frequent than expected in recent years and only two regions show significant trends in annual counts; few show seasonal trends. We found some regional relationships between flood counts and climate indices, annually and seasonally; nationally the Pacific North American pattern is related to annual counts of the 2 largest floods. Large‐flood occurrence was generally stable across the United States in the last five decades; this may or may not continue with projected warming. Plain Language Summary: Flood studies usually focus on records of the largest streamflow of each year because they are widely available and provide relatively large samples. However, the largest flow of each year is not necessarily big enough to inundate the floodplain and pose risks to human communities. Our study investigated only the largest observed floods in the record, ones that occur roughly every five to fifty years. We used a set of 473 gages that measure near‐natural river flow across the United States. We studied many aspects of these consequential events at regional and national scales for the period 1966–2015, including seasonality, changes over time, and associations with known climate patterns. We found that the seasonality of large floods is strong in some regions, but weak or complex in others where floods have a variety of causes. There is little evidence that large floods have become more or less frequent with time, but we found some regional and national associations between large‐flood occurrence and known climate patterns such as the Pacific North American pattern. Large‐flood occurrence has generally been stable in the last five decades, but continued re‐evaluation of the historical record is important as new data are collected. Key Points: First broad characterization of large‐flood occurrence in the conterminous United States in the modern hydroclimate, including seasonalityBoth traditional and novel methods to assess temporal distribution of large floodsFew significant temporal trends, but significant relations to climate indices for some regions and seasons [ABSTRACT FROM AUTHOR]

Long short‐term memory (LSTM) networks represent one of the most prevalent deep learning (DL) architectures in current hydrological modeling, but they remain black boxes from which process understanding can hardly be obtained. This study aims to demonstrate the potential of interpretive DL in gaining scientific insights using flood prediction across the contiguous United States (CONUS) as a case study. Two interpretation methods were adopted to decipher the machine‐captured patterns and inner workings of LSTM networks. The DL interpretation by the expected gradients method revealed three distinct input‐output relationships learned by LSTM‐based runoff models in 160 individual catchments. These relationships correspond to three flood‐inducing mechanisms—snowmelt, recent rainfall, and historical rainfall—that account for 10.1%, 60.9%, and 29.0% of the 20,908 flow peaks identified from the data set, respectively. Single flooding mechanisms dominate 70.7% of the investigated catchments (11.9% snowmelt‐dominated, 34.4% recent rainfall‐dominated, and 24.4% historical rainfall‐dominated mechanisms), and the remaining 29.3% have mixed mechanisms. The spatial variability in the dominant mechanisms reflects the catchments' geographic and climatic conditions. Moreover, the additive decomposition method unveils how the LSTM network behaves differently in retaining and discarding information when emulating different types of floods. Information from inputs within previous time steps can be partially stored in the memory of LSTM networks to predict snowmelt‐induced and historical rainfall‐induced floods, while for recent rainfall‐induced floods, only recent information is retained. Overall, this study provides a new perspective for understanding hydrological processes and extremes and demonstrates the prospect of artificial intelligence‐assisted scientific discovery in the future. Key Points: The expected gradients method was used to retrieve flooding mechanisms learned by long short‐term memory networks in 160 catchments in the United StatesSnowmelt, recent rainfall, and historical rainfall induce 10.1%, 60.9%, and 29.0% of the 20,908 identified flow peaks, respectivelyThe additive decomposition method unveiled models' distinct behaviors in retaining and discarding information for different flood types [ABSTRACT FROM AUTHOR]