1. Introduction
Flooding is one of the deadliest natural hazards in the United States, according to the National Oceanic and Atmospheric Administration (NOAA) National Weather Service’s (NWS) United States Natural Hazard Statistics (NOAA 2020c). In 2017 alone, there were 116 flood-related deaths in the United States, with over $60 billion dollars in damages (NOAA 2020c). The most flood-related fatalities have occurred in the state of Texas, with 205 reported from 2010 to 2018 (NOAA 2020c) and 760 from 1950 to 2008 (Ashley and Ashley 2008). The NWS reports that during flooding, most fatalities involve vehicles that are likely trying to cross flooded roads (NOAA 2019b).
In this work we focus on low-water crossings (LWXs). These structures are at grade, or slightly raised, with respect to a streambed (Gautam and Bhattarai 2018). As such, LWXs are designed to be overtopped by high flows and closed to traffic (Carstens and Woo 1984). If a crossing is not closed, then flood hazards at an LWX can cause harm by pushing, overturning, or floating a vehicle and could result in drowning of the individuals inside the vehicle. In Texas, the Department of Transportation has developed signage strategies to bring awareness to the possible hazards these structures pose (Balke et al. 2011). These strategies include signage and automated lights that flash when a hazard is detected by an associated streamflow gauge.
An LWX is an appropriate structure for roads with average daily traffic of less than 25 vehicles (Gautam and Bhattarai 2018; Rossmiller et al. 1983). This requirement results in LWXs being implemented in remote locations. The remoteness of these structures makes timely emergency response difficult. As a result, flood hazards at LWXs can be deadly. In June of 2016, nine U.S. Army soldiers drowned at Fort Hood, Texas. The military vehicle was swept downstream while attempting to cross a flooded LWX. The Army’s Investigating Officer stated that procedural improvements could be made to better mitigate future risks. In April of 2017, another soldier was killed at Fort Hood when his personal vehicle was swept away during a flash flood, and a civilian diver drowned a few days later while looking for the original victim.
NOAA/NWS works to prevent the loss of life and property by providing operational forecasts and warnings. There are 13 River Forecast Centers (RFCs; NOAA 2019a) across the United States responsible for regionally upholding the NWS mission with regards to flood hazards. In 2016, NOAA launched the National Water Model (NWM; NOAA 2020b) to predict streamflow at 2.67 million streams in the United States. Flood monitoring forecasts, such as those from the RFCs and NWM, predict the severity and area affected of an upcoming event. Early warning systems ingest flood forecasts, then communicate and distribute that information to protect the public. Preventative action based on forecasts from operational agencies can mitigate the risk associated with a flood hazard before harm is inflicted.
When the possible consequences of flood hazards are deadly, it is imperative that warning systems provide skillful and timely forecasts (Emerton et al. 2016; Pagano et al. 2014). The nonlinearity of atmospheric and hydrologic processes challenges the timeliness and accuracy of predictions (Lorenz 1969; Kumar 2011). Predictability challenges include initial value, boundary value, and parameter estimation problems that each introduce uncertainty in forecasts. That uncertainty cascades through chained models and into the flood forecasts used for decision-making (Zappa et al. 2010).
Ensemble prediction systems (EPSs) allow for the operational prediction of uncertainty (Emerton et al. 2016; Palmer 2017). EPSs are rooted in monthly weather forecasting (Murphy and Palmer 1986) to address the chaotic nature of the atmosphere. Ensemble stochastic schemes introduce slight variations into the components of models (Buizza et al. 2007; Palmer 2001) to generate a spread of equally probable scenarios. The ensemble spread can provide information on the outcome that is most likely, which can be more useful than a single “best guess” from a deterministic approach. Additionally, a probabilistic approach can predict the worst-case scenario, assuming the system accounts for all sources of uncertainty. Due to their benefits, EPSs have evolved the focus of numerical weather prediction from determinism to probability (Palmer 2017).
The ability of EPSs to extend lead times and better quantify predictability makes them an appealing technique for flood forecasting (Cloke and Pappenberger 2009; Emerton et al. 2016). Started in 2004, the Hydrological Ensemble Prediction Experiment (HEPEX) is an international project aimed at advancing hydrologic prediction technologies (Schaake et al. 2007). HEPEX brings together the international meteorological and hydrological communities to produce and utilize ensemble streamflow predictions (ESPs). ESP is the name originally given to long-range ensemble forecasts in the United States and only recently has the United States deployed short-to-medium range (~1–15-day forecast) ESPs. Verification studies of short- to medium-range ESPs are widespread (Duan et al. 2019; Bartholmes et al. 2009; Jaun and Ahrens 2009; Demargne et al. 2010; Hopson and Webster 2010; Addor et al. 2011; Zappa et al. 2013). Several different operational forecasting centers across the globe have adopted ensemble techniques to produce ESPs (Cloke and Pappenberger 2009). The NWS currently operates an end-to-end Hydrologic Ensemble Forecast Service (HEFS; Demargne et al. 2014) as part of the Advanced Hydrologic Prediction Service (AHPS; McEnery et al. 2005). AHPS is available at over 2500 U.S. Geological Survey (USGS) streamflow gauges across the United States with plans to further expand (NOAA 2020a).
Communicating the risk of flood hazard and the uncertainty of that risk to the community is as important as the ESPs forecast skill. The communication, perception, and use of alerts from ESPs has received far less attention than the development of the ESPs themselves (Demeritt et al. 2013). The HEPEX project has highlighted the need for ESP post processors that consider user needs (Schaake et al. 2007). Visualizing probabilistic flood forecasts as maps is a valuable communication tool but displaying uncertainty does not guarantee that uncertainty is understood or considered in the decision-making process (Pappenberger et al. 2013). For communication to be effective, the user should be able to understand the hazard and the uncertainty conveyed by a probabilistic communication strategy (Ramos et al. 2010). The success of a probabilistic communication strategy is therefore dependent on the way information is communicated. This is of added importance when dealing with the public due to false alarms negatively impacting public responsiveness to flood warnings (Demeritt et al. 2007).
The first objective of this work is to present proof of concept for an integrated modeling system that ingests probabilistic weather forecasts from the National Center of Atmospheric Research (NCAR) experimental EPS and propagates the uncertainty of the precipitation through hydrologic and hydraulic models. The premise behind an EPS such as this is considered a well-accepted approach around the world (Duan et al. 2019). In contrast to previous efforts, our work goes beyond streamflow by incorporating hydraulic simulations. In doing so, we can provide information about flow velocity and depth at remote locations. To make probabilistic hydraulic flow velocity and depth predictions, a hydraulic model is added to the integrated modeling system. This provides information that is directly relevant to decision-makers. We further communicate the information in a way that helps managers plan for road closures to specific vehicle classes. The proposed model chain is tested with a case study at a remote LWX on Fort Hood, in central Texas. The two hydrologic events that resulted in loss of life, as well as three additional events are hindcasted for the case study. Our goal is to use these five cases as a proof of concept illustration of a case study application related to hydraulic hazards.
The second objective of this work is to communicate flood hazards at a remote LWX, as well as the uncertainty of that prediction through ensemble member agreement. Flood hazard is currently communicated with gauge specific NWS high-water categories. As an alternative to high-water categories, the multiplication of flow velocity and flow depth is defined as the hydraulic stability threshold (HST). HST values are set on the thresholds that cause vehicles to float, slide, or overturn (Shand et al. 2011). As such, HST values are vehicle specific and not site specific making it preferable in remote areas. The goal of the second objective is to examine if alternative strategies for hazard communication can have added benefit in conjunction with current methods.
2. Study site and model setup
a. Cowhouse Creek watershed
We perform the integrated modeling study in Cowhouse Creek watershed. Eleven flood-related deaths occurred in this watershed between 2016 and 2017 as the victims were attempting to traverse flooded LWXs. The watershed is USGS hydrologic unit code (HUC) 12070202 and is herein referred to as the Cowhouse Creek catchment. The 1625-km2 catchment is in central Texas (Fig. 1a) with bounding coordinates 31°44′N, 97°46′W, 31°04′N, 98°35′W spanning Bell, Comanche, Coryell, Hamilton, Lampasas, and Mills Counties (Fig. 1b). The catchment is mostly open rangeland used for grazing cattle and military maneuver training. The 2016 National Land Cover Database (NLCD; Yang et al. 2018) has 74% of the area classified as grassland, shrub, or pasture and only 5% classified as developed. Half of the basin’s soils are characterized by the Natural Resource Conservation Service (NRCS) into Hydrologic Soil Group (HSG) D (NRCS 2019). HSG D soils are clayey with low infiltration rates and a high runoff potential. The flooding potential is compounded by the lack of built infrastructure to attenuate large flows. The infrastructure along Cowhouse Creek primarily consists of improved ford LWXs. Most of the LWXs have signage in place with flashing lights to warn of hazards, as part of the “Turn Around Don’t Drown” campaign in Texas (NOAA 2019b). Predicting floods in this region is difficult due to a lack of observations. The basin is minimally gauged with a single USGS stream gauge located at the confluence of Cowhouse Creek and Bee House Creek (black dot, Fig. 1a). The gauge is at an elevation of 230 m, while the basin’s highest elevation is 540 m and the outlet is 200 m [USGS National Elevation Dataset (NED); Gesch et al. 2002].
Watershed location map: (a) the Cowhouse Creek catchment with the blue lines denoting the stream network’s major channels, the black circle denoting the location of the USGS stream gauge, and red diamond denoting the LWX. (b) The location of the Cowhouse Creek catchment, denoted in red, within the state of Texas.
Citation: Journal of Hydrometeorology 21, 5; 10.1175/JHM-D-19-0238.1
b. Ensemble precipitation forcing
The NCAR experimental EPS was selected as the precipitation forecast for this study (Schwartz et al. 2015). Historical NCAR EPS hourly precipitation accumulation is publicly available for 2015–17. The NCAR EPS has 10 members, 48-h forecasts initialized daily at 0000 UTC, and a 3-km horizontal grid spacing over the contiguous United States (CONUS). Although the NCAR EPS is no longer operational, the experimental High-Resolution Rapid Refresh Ensemble (HRRRE; Dowell et al. 2016) system run at NOAA/ESRL/GSD shares similar characteristics with 9 members, a 36-h forecast, and 3-km CONUS grid. The NCAR EPS was selected in anticipation of using the operational HRRRE as a precipitation forecast replacement if this study’s methodology were to be implemented in an operational environment. During preprocessing, we regridded the precipitation using bilinear interpolation, and projected the precipitation to a 1-km2, Albers equal-area grid that could be ingested into the hydrologic model. The interpolation was necessary because the hydrologic model currently does not accept forcing at the NCAR EPS native 3-km resolution. Daily 24-h forecasts were combined to generate a multiday precipitation time series without gaps or overlaps. A discontinuity is introduced between the last forecast hour of one day and the first hour of the next day’s forecast because of this. The discontinuity errors were assumed to be negligible. The stitching of daily forecasts together results in a 24-h streamflow forecast initialized at 0000 UTC. The chaotic nature of the atmosphere results in continued divergence from reality with time, therefore events forecasted earlier in the 24-h horizon are subject to less uncertainty. To limit this, only the first 24 h of the NCAR EPS 48-h forecasts are utilized. Computing latencies decreases the forecast horizon of 24 h at 0000 UTC to approximately 22 h.
c. Hydrologic model
Each precipitation ensemble member is used to force a basin-scale hydrologic model simulation built with the Hydrologic Engineering Center Hydrologic Modeling System software (HEC-HMS; Feldman 2000; Scharffenberg et al. 2018). HEC-HMS is an open source hydrologic modeling software developed by the U.S. Army Corps of Engineers (USACE). HEC-HMS simulates hydrologic processes, such as soil infiltration, overland flow, and hydrologic routing, at the watershed and river reach scale. Each of the precipitation members is independently forcing the same HEC-HMS model. This results in a streamflow ensemble with the same number of members as the precipitation. The model time step is hourly, and the routing calculation is subhourly and variable to help maintain model stability. The hydrologic model was configured with parameterizations tailored for basins with minimal observations (Table 1). The catchment and stream network (Fig. 2a) for the model were delineated from the USGS NED (Gesch et al. 2002) digital elevation model (DEM) (Fig. 1a). The minimum contributing area to initiate a stream was set as 10 km2. The stream definition value was validated by visually comparing the stream network delineated to the Cowhouse Creek stream network in the USGS National Hydrography Dataset (NHD; USGS 2019).
The subroutines, as well as the associated parameters, used within HEC-HMS and HEC-RAS for the Cowhouse Creek catchment hydrologic and hydraulic models.
The HEC-HMS hydrologic model and HEC-RAS hydraulic model domains: (a) the vector-based watershed delineation of the Cowhouse Creek catchment and (b) the 1-km2 grid associated with the Green and Ampt and ModClark subroutines within HEC-HMS. The soil porosity is shown as an example parameter. The subbasins are the areas that the baseflow and Muskingum–Cunge routing subroutines of HEC-HMS are parameterized by. (c) The HEC-RAS model domain that spans a single reach along Cowhouse Creek with the USGS streamflow gauge (black point) and the LWX (red diamond) along that reach.
Citation: Journal of Hydrometeorology 21, 5; 10.1175/JHM-D-19-0238.1
Within HEC-HMS, the necessary parameters of the gridded Green and Ampt loss method can be estimated from soil data alone. Such a method is preferred in minimally gauged catchments, such as the Cowhouse Creek catchment with its single USGS streamflow gauge. Saturated water content, hydraulic conductivity, and wetting front suction grids were generated using the NRCS Soil Survey Geographic database (SSURGO; NRCS 2019) and the U.S. Department of Agriculture (USDA) Agricultural Research Service (ARS) ROSETTA hydraulic parameter model (Schaap et al. 2001). Sand/silt/clay percentage, bulk density at 15 bars, and water content percentage at 15 and 33 bars were extracted from the SSURGO database and imported into ROSETTA with saturated water content and hydraulic conductivity as ROSETTA output. The wetting front suction was approximated empirically and compared to literature for validation (van Genuchten 1980; Rawls et al. 1983). The initial soil water content is an initial condition and was estimated as the top 10-cm soil moisture simulated from the mosaic North American Land Data Assimilation System (NLDAS-2: Xia et al. 2012) for a 1-month spinup simulation forced with NLDAS precipitation and evapotranspiration. The soil moisture at the end of the 1-month spinup became the initial water content for the event simulations. HEC-HMS requires the ModClark transform method be used with a gridded loss method (Fig. 2b). The ModClark method has two parameters, time of concentration and storage coefficient, that were estimated using the SCS watershed lag method (Mockus 1961; NRCS 2010). Baseflow was calculated as a constant monthly value through baseflow separation of the historical record of daily streamflow for the period 1950–2018. The digital filtering approach was used for baseflow separation (Lyne and Hollick 1979). The Muskingum–Cunge routing method was selected for its applicability in basins that are difficult to calibrate due to the lack of hydrograph observations. Routing parameterization was done by subbasin (Fig. 2a), while the other parameterizations were on a grid (Fig. 2b).
The hydrologic model was calibrated using four events. Then, we used the calibrated model to simulate five events that composed our case study. All events (calibration and case study) were selected because they exceed the NWS action level threshold. The four calibration events fell outside the NCAR EPS data availability (2015–18), but within the subdaily streamflow record (from 2009 to present). Calibration was conducted manually on the saturated water content, time of concentration, and storage coefficient parameters and evaluated as a maximization of the Nash–Sutcliffe efficiency (NSE) and Kling–Gupta efficiency (KGE) coefficients.
Hydrologic model performance was assessed in three regards following the outline by Ritter and Muñoz-Carpena (2013). The assessments include a graphical representation of the relationship between model estimates and observations, an absolute value error indicator, and a dimensionless index for quantifying the goodness of fit. We used the root-mean-square error (RMSE), NSE coefficient, and KGE coefficient.
The hydrologic model performance assessment compared streamflow observations from the USGS streamflow gauge to HEC-HMS streamflow simulations for the four calibration events. To assess the hydrologic model, the HEC-HMS hydrological model needed to be forced with precipitation that was as close to reality as possible. Otherwise uncertainty in the precipitation forcing would influence the comparison of streamflow observations and simulations. Therefore, the HEC-HMS model was forced with the NCEP Stage IV (Lin 2011) quantitative precipitation estimation (QPE). The Stage IV precipitation analysis is available on a 4-km grid across CONUS and at 1-h temporal resolution. Lin et al. (2018) showed that Stage IV can be a suitable substitute for radar in basins greater than 1000 km2. HEC-HMS streamflow simulations, with Stage IV precipitation forcing, were then compared to USGS streamflow observations to assess hydrologic model performance. The precipitation forcing was switched from Stage IV to NCAR EPS precipitation for the case study events to introduce the probabilistic element needed.
d. Hydraulic model
The hydraulic model generated a rating curve for the LWX, an improved ford, selected for this study (Fig. 2c). The hydraulic model takes HEC-HMS streamflow and routes the water along the channel to provide an estimate of flow velocity and depth at a fine spatial scale. A hydraulic model for a single reach on Cowhouse Creek was built and run with HEC’s River Analysis System software (HEC-RAS: Brunner 2016). The model spatial features were digitized in Esri’s ArcGIS with the HEC-GeoRAS toolbox (Ackerman 2009). The LWX was modeled as an inline broad-crested weir. Structure geometries and elevations were determined from point survey data collected in situ. User specified parameters included Manning’s n values for the channel and floodplain. These parameters, as well as structure flow contraction and expansion parameters were selected from values in HEC-RAS’s Reference Manual (Brunner 2016) and based off field observations. A rating curve was generated by running the HEC-RAS model successive times with a steady analysis and a subcritical flow regime. HEC-RAS solves the kinematic wave approximation to the Saint-Venant equations with initial upstream flow conditions ranging from 1 to 2500 m s−1 and a downstream boundary condition assumed to be the normal flow depth.
e. Hazard communication
Hydrologic/hydraulic simulations were compared to high-water categories specified by the NWS. The NWS uses high-water level categories at USGS stream gauges to communicate flood hazard. The high-water level terms include bank-full stage, action stage, and flood stage. The action stage is defined by the NWS as, “the stage which, when reached by a rising stream, represents the level where the NWS or a partner/user needs to take some type of mitigation action in preparation for possible significant hydrologic activity” (NOAA 2016). The flood stage is more severe and defined by the NWS as the water level that is hazardous to life and property. High-water levels are presented as stage but were converted in this study to flow using the rating curve for the USGS gauge.
The high-water levels at the USGS gauge are not applicable at the remote LWXs on Cowhouse Creek because the crossings are at grade with the channel bottom. The NWS action stage on Cowhouse Creek is 5.5 m. For the elevated bridge where the USGS gauge is located, such a threshold is suitable. However, for the at-grade LWXs on Cowhouse Creek, 5.5 m of water would submerge all vehicles. As an alternative to high-water categories, the multiplication of flow velocity and flow depth is defined as HST. HST values are set on the stability thresholds that cause vehicles to float, slide, or overturn (Shand et al. 2011). The appearance of HST in literature was reviewed by Martínez-Gomariz et al. (2018) with values from experimental and analytical studies. The thresholds are highly vehicle dependent. Although this requires comprehensive testing, a communication strategy that incorporates HST allows tailoring to specific vehicles. The best reference to date for HST criteria, as stated by the review of Martínez-Gomariz et al. (2018), is from the Australian Rainfall and Runoff (AR&R) guideline (Shand et al. 2011) and used herein. The least conservative AR&R criteria is for stationary, large four-wheel-drive vehicles. These criteria specify HST never exceeds 0.6 m2 s−1 with additional constraints that depth never exceeds 0.5 m and velocity never exceeds 3.0 m s−1 (Shand et al. 2011). In this study, two graphics were generated to report flood hazard by comparison to the HST. The first graphic overlays forecasted hydraulics with the AR&R HST criteria to communicate flood hazard. The second graphic overlays time series of flow velocity multiplied by flow depth for each model member with the HST criterion of 0.6 m2 s−1 to communicate flood hazard, as well as the forecast uncertainty through model member agreement.
3. Results
a. Event descriptions
The proposed methodology was tested by performing hindcasts of five hydrologic events that exceeded the NWS action level. These events occurred in May/June 2015, October 2015, May/June 2016, April 2016, and April 2017 (herein M15, O15, M16, A16, and A17). M16 and A17 were the two deadly events. M16 claimed 9 lives and was reported at 1630 UTC 2 June 2016. The exact time of A17 is unknown but believed to be sometime on 3 April 2017 and killed one individual. The five events are grouped into either isolated or successive hydrologic responses. M15 and M16 are successive events with multiple hydrologic responses within the time frame of analysis. As such, they were simulated over a longer period of 10 days. The isolated events, O15/A16/A17, were simulated over a 7-day period. Because of the loss of life, M16 and A17 results are given greater focus herein.
The precipitation that fell during M16 on 2 June was more intense than that of A17 on 2 April (Figs. 3a,b, black dots). The two storms are similar in duration, but they differ in that the M16 storm had 1–2 h of intense rainfall with lighter precipitation before and after, while the A17 storm was characterized by more consistent, less intense rainfall throughout (not shown). There was 42.9 mm of accumulated precipitation (ACPC) over 1200–1800 UTC 2 June (Fig. 3a), with a maximum hourly rate of 17.3 mm h−1. Over 1200–1800 UTC 2 April (Fig. 3b), there was 31.2 mm of ACPC with maximum hourly rate of 6.6 mm h−1. When the precipitation rate intensifies and surpasses the maximum infiltration rate of the soil, Hortonian overland runoff is generated as infiltration excess even if the soil is not saturated. Similar intense storms of shorter duration were also seen in M15 (Fig. 3c), while A16 and O15 were longer duration, weaker intensity patterns of intermittent precipitation over several days (Figs. 3d,e).
Temporal comparison of Stage IV precipitation to the NCAR EPS precipitation ensemble. Each NCAR EPS daily forecast has a 24-h horizon and is initialized at 0000 UTC: (a) time series of Stage IV 6-h accumulated precipitation averaged across the area that is upstream of the USGS streamflow gauge shown as black points and boxplot of the 10 NCAR EPS members with precipitation presented as a 6-h accumulated value averaged across the area that is upstream of the USGS streamflow gauge. (b)–(e) As in (a), but for the four other events.
Citation: Journal of Hydrometeorology 21, 5; 10.1175/JHM-D-19-0238.1
b. Precipitation data
The temporal variability of 6-h ACPC is examined for the contributing area to the USGS streamflow gauge (Fig. 3). Simulated ACPC across the 10 NCAR EPS ensemble members is compared to observed Stage IV ACPC. The M16 observed Stage IV ACPC is concentrated on an intense event of 52.1 mm that occurred over 1200–0000 UTC 2 June 2016 (Fig. 3a) with 1.5 mm ACPC the day before and 5.0 mm ACPC the day after. NCAR EPS median is 31.0 mm and the interquartile range (IQR) is 39.5 mm. The NCAR EPS precipitation is on average underestimating on 2 June 2016 when compared to the Stage IV (Fig. 3a). However, the NCAR EPS is generally overestimating on the days before and after 2 June 2016. The timing of the NCAR EPS precipitation is ahead of reality in some forecast members and behind in others. The erroneous timing causes an underestimation of M16 on 2 June, but an overestimation with 28.3 mm median ACPC the day before and 16.6 mm median ACPC the day after (Fig. 3a). The mistiming also contributes to the large IQR dispersion across the 10 ensemble members. This mistiming is also seen in the A16 event (Fig. 3d) with all the NCAR ensemble members lagging the Stage IV. M15 has an example where NCAR EPS completely missed a hydrologic response on 27 May that is captured by Stage IV. The A17 observed Stage IV ACPC is 51.8 mm (Fig. 3b) over 0600–1800 UTC 2 April 2017. NCAR EPS median is 49.1 mm and the IQR is 12.0 mm over those 12 h. The NCAR EPS ensemble is showing less intense, longer-duration storms on the subdaily scale compared to Stage IV (Fig. 3b). Additionally, the erroneous timing seen in the 2016 event is not present with the 2017 event.
The spatial variability of Stage IV ACPC is compared with the simulated NCAR EPS for the M16 and A17 days of greatest ACPC (Fig. 4). As expected, none of the simulated NCAR EPS members (Figs. 4b,d) perfectly match the observed Stage IV (Figs. 4a,c). A majority of the NCAR EPS members correctly forecast greater precipitation in the headwaters of the catchment in M16 and a band of precipitation in A17. The ensemble variability is expected and welcome, as the probabilistic framework of the NCAR EPS produces 10 outcomes that are equally likely to occur. Across the five case study events, there is not a general bias of the NCAR EPS forecasts when compared to Stage IV. All the successive storms in M16 were underestimated by NCAR EPS (Fig. 3c), but the isolated storms in A16 and A17 were overestimated when compared to Stage IV (Figs. 3b,d).
Spatial comparison of accumulated precipitation for each of the (b) 10 NCAR EPS ensemble members and (a) Stage IV precipitation observations on 2 Jun 2016. (c),(d) As in (a) and (b), but for 2 Apr 2017.
Citation: Journal of Hydrometeorology 21, 5; 10.1175/JHM-D-19-0238.1
M16 and A17 had similar rainfall 24-h accumulations. This is seen in the Stage IV spatial distribution of ACPC (Figs. 4a,c). However, the M16 recorded a higher peak flow in the USGS observational streamflow record on 2 June. This difference again highlights the importance of both the precipitation intensity, as well as the antecedent soil state during both events. M16 had a higher precipitation intensity than A17 producing more infiltration-excess runoff. Soils that have been recently primed with antecedent precipitation will have a greater potential to generate saturation-excess runoff during future storms relative to the same soil when dry. For Cowhouse Creek catchment and its soils with high clay content and low infiltration and hydraulic conductivity, both infiltration-excess and saturation-excess runoff are seen to be important for explaining the magnitude of hydrologic responses observed.
The observed Stage IV ACPC for both M16 and A17 have the greatest precipitation totals localized within a fraction of the catchment area (Figs. 4a,c). The maximum M16 observed ACPC is 99.9 mm and the minimum 25.6 mm (Fig. 4a). Such a large spatial variability across the catchment can be more hazardous for those that are trafficking the LWXs. Those individuals are downstream from the greatest precipitation but are not experiencing the same amount of precipitation that could create a false sense of security.
c. Hydrologic model
1) Model calibration and initial conditions
The HMS model is calibrated on four events which occurred in April 2009, October 2009, January 2010, and September 2010 (herein A09, O09, J10, and S10). The calibration is limited to the four events that exceed the NWS action level threshold and occurred prior to 2015 (when NCAR EPS data is available) but after 2009 when the historical record of subdaily streamflow is available.
Visually, the four events match observations fairly well (Fig. 5). The beginning of each simulated hydrograph rising limb match observations within one hour. The greatest discrepancy is J10 underestimating the peak flow by 358.6 m3 s−1 (Fig. 5c). RMSE values in chronological order are 45.9, 29.4, 193.0, and 28.9 m3 s−1. All four calibration events have NSE coefficients greater than 0.50 and the KGE coefficients for A09, O09, and S10 all exceed 0.75. By definition of NSE, the model is simulating streamflow better than a random prediction made off the time frame observational mean.
A graphical representation of the HEC-HMS performance for the calibration events evaluated at the USGS gauge. (a) The top plot is the event hyetograph of hourly Stage IV precipitation averaged across the area of the Cowhouse catchment that is upstream of the USGS streamflow gauge. The bottom plot is observed streamflow (black dashed line) from the USGS streamflow gauge and HEC-HMS modeled streamflow (turquoise line) from a simulation forced by the Stage IV precipitation time series in the top plot. NSE and KGE skill scores are shown for each simulation. (b)–(d) As in (a), but for three other calibration events.
Citation: Journal of Hydrometeorology 21, 5; 10.1175/JHM-D-19-0238.1
All HEC-HMS streamflow simulations (calibration and case study) require initial soil moisture conditions. For each event the initial soil moisture conditions were generated from a 1-month spinup simulation. The model water content at the last time step of the continuous simulation becomes the initial water content for event simulations. The case study initial conditions for M15, M16, and A17 are all near saturation due to substantial precipitation during the 1-month spinup (Fig. 6). O15 soils are relatively dry at the start of the simulation in comparison.
The initial soil moisture across the Cowhouse Creek catchment for the five case study simulations in HEC-HMS. The grids are the final model state from a 1-month spinup prior to each case study event. (a) May 2016, M16; (b) April 2017, A17; (c) May 2015, M15; (d) April 2016, A16; and (e) October 2015, O15.
Citation: Journal of Hydrometeorology 21, 5; 10.1175/JHM-D-19-0238.1
2) Simulation of case studies
Streamflow simulations generated using the calibrated HEC-HMS model with Stage IV precipitation are compared to USGS streamflow gauge observations (Fig. 7, yellow and purple). Visually, the modeled streamflow resembles observations. At 0400 UTC 3 June 2016 the USGS gauge stopped recording, preventing a complete comparison of M16 to observations (Fig. 7a). M15 does have complete observations and four successive hydrologic responses in 7 days. The model simulates the peaks of these four events within 0–3 h of the observed peaks and with an average difference of 13.8 m3 s−1 (Fig. 7c).
(a) The event spaghetti hydrographs from HEC-HMS output at the USGS gauge. Each green shading represents a streamflow time series corresponding to one of the ensemble members. The hyetograph has the time series of each ensemble member mean precipitation rate for the contributing area to the USGS gauge. Shadings are opaque to show when multiple ensemble members report similar values. The time series consists of NCAR EPS daily forecasts with a 24-h horizon initialized at 0000 UTC being stitched together seamlessly. Stage IV precipitation is in yellow. USGS streamflow observations are hashed in black and simulations with Stage IV are in purple. (b)–(e) As in (a), but for the four other events.
Citation: Journal of Hydrometeorology 21, 5; 10.1175/JHM-D-19-0238.1
The simulated A17 peak flow with Stage IV (Fig. 7b) underestimates by 8.7 m3 s−1 and has a 0-h lag. Similarly, A16 underestimates by 8.1 m3 s−1 and has a 0-h lag. The peak flow underestimation is probably more than what is reported because the HEC-HMS model is overestimating baseflow in April by 6.2 m3 s−1 averaged over A16 and A17. These two events highlight the importance of the initial soil moisture. A17 is near saturation at the beginning of the simulation (Fig. 6b), while A16 is unsaturated (Fig. 6d). O15 shows this as well with 19 mm ACPC from 0300 to 0700 UTC on 23 October (Fig. 7e) and dry soils (Fig. 6e) combining to not produce a hydrologic response in the Stage IV simulation.
We perform an independent hydrologic model simulation for each member of the NCAR EPS forecast. The HEC-HMS parameterization does not change between the streamflow forecasts with NCAR EPS forcing and with Stage IV forcing. This results in a 10-member streamflow ensemble for each event (Fig. 7, blue and green). Spread in the NCAR EPS propagates through the streamflow members to create streamflow spread. The precipitation in M16 is highly uncertain (Fig. 7a) as seen by the lighter blue hyetograph shading and a general overestimation of the ensemble mean streamflow for 1–2 June. The lack of complete observations prevents comparison of the deadly hydrologic response, but such an event is an example of how probabilistic prediction methods have utility when observational methods fail.
The time of accident in A17 is unknown, but the mean ESP peaks at 526.8 m3 s−1 at 2200 UTC 2 April with a standard deviation of 262.8 m3 s−1. The high standard deviation conveys how variable the hydrologic responses are at this time step with the strongest response at 1162.5 m3 s−1 and the weakest at 185.3 m3 s−1. The time of peak in the A17 NCAR EPS-forced hydrographs (Fig. 7b) ranges from 2 h ahead to 4 h after the peak of the USGS observations. There is no difference between the time of peak of the hydrologic simulation forced with Stage IV precipitation and the USGS observations (Fig. 7b). The differences are therefore a product of the spatial and temporal variability in precipitation members cascading through the streamflow ensemble.
The lag time is defined as the difference between maximum rainfall and the hydrograph peak. For A17, the average value across the 10 ensemble members is 6.7 h, while the lag from the HEC-HMS simulation with Stage IV precipitation is 10 h (Fig. 7b). The Stage IV lag time is higher partially because A17 had 7 h of near constant precipitation (mean of 6.7 mm h−1 with standard deviation of 0.8 mm h−1), but the maximum was during the first hour. For the M15 peak flow, the average ensemble lag time is 8.8 h (three members did not have a peak flow and were excluded), while the Stage IV is 5 h (Fig. 7c). The variability in lag time is caused by variability in the NCAR EPS precipitation cascading through the hydrologic model.
Interestingly, M15 has a flood that is observed and simulated well with Stage IV but missed by every NCAR EPS member (Fig. 7c). This is an example of a false negative where the ensemble is in strong agreement on the wrong answer. This is a disadvantage of probabilistic methods and supports that the most effective strategies for minimizing flood fatalities should be multidimensional incorporating predictive and observational methods.
d. Hazard communication
1) Flood hazard with high-water levels
At the USGS gauge, the median ESP and IQR is compared to NWS high-water levels converted to flows (Fig. 8). The action stage at the USGS gauge is 5.5 m which equates to an action streamflow of 260.5 m3 s−1. The flood stage is 6.1 m (streamflow of 334.1 m3 s−1). Figure 8 clearly shows how the NCAR EPS is forcing streamflow underestimation in some cases (M15) and overestimation in others (A17).
Comparison of simulated streamflow with NCAR EPS precipitation to the NWS high-water level categories specified at the USGS streamflow gauge. (a) The top plot is the 10-member median streamflow (dark green line) with the interquartile range (lighter green area) and the USGS streamflow observations (black dashed line) compared to the NWS action stage and flood stage values that are converted to flows. The bottom plot is the probability of the forecasted streamflow being greater than the NWS action flow qa taken as the ratio of the ensemble members that are greater than qa to the total ensemble members. (b)–(e) As in (a), but for the other four events.
Citation: Journal of Hydrometeorology 21, 5; 10.1175/JHM-D-19-0238.1
For M16 evaluated at the USGS gauge, there are never more than five forecast members simulating streamflow greater than the action flow, but those five members also exceed the flood flow. This equates to a 50% probability of the M16 streamflow forecast being greater than both the action and flood flow (Fig. 8a). This differs for 27 May in M16, where there is a 30% probability of exceeding the action level and a 0% probability of exceeding the flood level.
The ESP median streamflow is greater than the action flow for 11 h in A17 (Fig. 8b). USGS observed streamflow is greater than the action flow for a single time step, at 2000 UTC 2 April 2017. In A17 (Fig. 8b), the probability of the forecast streamflow being greater than the action and flood flows peak at 80%. There is majority agreement of the forecast being greater than both the action and flood level, but not unanimous. A16 and O15 have similar results, where the majority of the streamflow ensemble is exceeding the flood threshold for multiple hours (Figs. 8d,e).
The flood flow, representative of a hazardous level to life and property, was not surpassed by the USGS observations in A16 or A17. Since the high-water levels are based on historical observations in the basin, these comparisons are relative to the historical observations. A17 highlights how an observed hydrologic event can pose a deadly flood hazard, while never exceeding the flood flow. As such, hazard thresholds should not be taken deterministically, but instead as guidelines with margins arising from the uncertainty associated with their estimation.
2) Flood hazard with HST
The hydrologic results presented have all been at the USGS gauge (Fig. 1a). The following hydraulic forecasts are for LWX 15, which is 2.5 km downstream of the USGS gauge on Cowhouse Creek (Fig. 2c). Of the 480 points (48 h and 10 ensemble members) in the M16 hydraulic ensemble (Fig. 9a), 92% are exceeding one of the AR&R HST criteria of 0.6 m2 s−1, 3.0 m s−1, and 0.5 m. In a scenario without any flood hazard, such as baseflow conditions, all 480 points would be in the first bins of the velocity and depth histograms (Fig. 9f). M16 distributions of flow velocity and flow depth have additional bins populated. Such a histogram shape is representative of hazardous conditions in the forecast.
Hydraulic forecasts at LWX 15 on Cowhouse Creek for event days: There are 480 points (red/blue points) plotted with the HEC-RAS rating curve for LWX 15 (black dashed line). The 480 points come from two event days of hourly data for 10 ensemble members. Histograms of both axes (y axis: flow velocity, x axis: flow depth) are plotted to capture the distribution of the 480 points since there is considerable overlap. There is not a hysteresis loop present because of the steady state assumption in HEC-RAS. The AR&R HST criteria for stationary, large four-wheel-drive vehicles is plotted (yellow line) to show when the rating curve, and therefore forecasted hydraulic conditions, is above a hazardous level. (b)–(e) As in (a), but for the four other events. (f) The event forecast for a constant baseflow (no hazard) condition.
Citation: Journal of Hydrometeorology 21, 5; 10.1175/JHM-D-19-0238.1
For A17 report days 2–3 April, half of the first day was simulated as baseflow (Fig. 7b). The baseflow appears in the first HST graphic (Fig. 9b) with a large count in the first bin of both histograms. A17 simulated hydraulics exceeded one of the AR&R HST criteria for 69% of the points in the ensemble (Fig. 9b, red points and bins). The time steps when either forecasted depth or velocity are potentially hazardous are visible at the tails of the histograms, indicated by red bins. The A17 histograms, with a large percentage of points in the first bins, are representative of an event that is more isolated, but still poses a flood hazard. Similarly, this is seen with A16 (Fig. 9d) and O15 (Fig. 9e). The general underestimation of M15 shows an overpopulated first bin as well. Based on the AR&R HST criteria, all five case study flood hazards are due to deep water and not fast flowing water. All of the histograms of water velocity (Fig. 9) except A16, show zero points exceeding the AR&R 3.0 m s−1 velocity HST. This is depicted by all the histogram bins being colored blue except one in A16. The water depth histograms (Fig. 9) have bins that exceed the AR&R 0.5-m depth HST colored red.
To communicate if ensemble members are reporting similar hazards, the 10 time series of flow depth multiplied by flow velocity are compared to the 0.6 m2 s−1 HST criterion (Fig. 10). Values are also shown for the USGS observations and the Stage IV simulations. For the entirety of 2–3 June in M16 (Fig. 10a), 8 of the 10 members are above the 0.6 m2 s−1 HST criterion. There is high agreement of a flood hazard exceeding the 0.6 m2 s−1 HST criterion within the ensemble itself, and between the ensemble and the values computed using the USGS streamflow observations and Stage IV simulations.
(a) The May 2016 hydraulics forecast at LWX 15 compared to the 0.6 m2 s−1 HST (yellow horizontal line). USGS streamflow observations are converted to flow velocity and depth at LWX 15 and shown in black. HMS simulations with Stage IV precipitation are shown in gray. The ensemble forecast is represented as the background color for any given hour with values greater than 0.6 m2 s−1 in red and less than 0.6 m2 s−1 in blue. Shading is only shown for when five or more members agree. The greater the intensity of the red or blue background, the higher the model agreement. (b)–(e) As in (a), but for four other events.
Citation: Journal of Hydrometeorology 21, 5; 10.1175/JHM-D-19-0238.1
For A17 (Fig. 10b), all 10 members forecast hazards matching the USGS observations. The HMS baseflow error appears again as an overestimation when comparing the Stage IV values to the USGS observations in April (Figs. 10b,d), and an underestimation in October (Fig. 10e). The underestimation of NCAR streamflow in M15 is shown in Fig. 10c. There is also less confidence in the threshold exceedance as compared to the other case study events.
Type II errors, or false negative, are the time steps where the USGS observations are above the 0.6 m2 s−1 HST criterion, but more than five of the ensemble members fall below. Type I errors, or false alarms, occur when observations are below the 0.6 m2 s−1 HST and the ensemble majority above. There are periods of false negatives in M16, M15, and A16. During those times, the ensemble is disagreeing with the observations, but the Stage IV simulation is agreeing with the observations. The forecast precipitation is causing a missed detection there, and not the HMS model. There are instances, specifically in M15 (Fig. 10c) and A16 (Fig. 10d), where the receding tails of the hydrographs recede too quickly compared to the observations causing a period of missed detection. During those times, the ensemble is detecting a hazard while the Stage IV simulations are not. The largest false alarm occurs in M16 (Fig. 10a) in the days leading up to the event that destroyed the USGS gauge.
4. Discussion and conclusions
a. Factors that influence flooding
Two events, M16 and A17, hindcasted in this study claimed lives. The environmental conditions that led to the two flood hazards, and the 3 others modeled, have similarities and differences. The average observed Stage IV ACPC upstream of the USGS gauge is similar, with 52.1 mm in M16 (Fig. 3a) and 51.8 mm in A17 (Fig. 3b). The two storms are similar in duration, both lasting about 8 h. They differ in intensity with the M16 storm having 1–2 h of intense rainfall with lighter precipitation before and after, while the A17 storm was a more consistent, less intense rainfall. M16 and A17 hindcasts differ in the magnitude of the hydrologic response. The observed M16 flow when the gauge stopped reporting was 529.5 m3 s−1 (Fig. 7a), while the A17 peak flow was only 264.5 m3 s−1 (Fig. 7b). The more intense M16 precipitation generated more infiltration-excess runoff.
Another difference between these five case studies is the antecedent environmental conditions. The antecedent soil moisture increases the flooding hazard by decreasing the soil’s ability to hold more water. Once additional infiltration increases the soil water content to the saturation level, any precipitation beyond the saturated hydraulic conductivity will runoff as saturation excess. M16, A17, and M15 all have soils near saturation at the beginning of the simulation. A16 and O15 had less saturated soils, therefore saturation-excess runoff contributes to the flow later for these events. This study highlights the importance of understanding the current environmental state of a basin when predicting the hydrologic response.
b. Utility of probabilistic framework
There can be large variability in streamflow across the different ESP members, even though they are all generated from the same hydrologic model. The ESP variability is attributed to the spatial and temporal precipitation variability across NCAR EPS ensemble members, and this variability cascades through the hydrologic and hydraulic models. The ESP variability (Fig. 7) highlights the benefit of using probabilistic techniques for forcing predictive hydrologic models. This result is consistent with numerous examples that have supported the utility of ESPs (Duan et al. 2019) but is unique in its application to remote water crossings.
The spatial domain of the Cowhouse Creek watershed (1625 km2) is very small compared to that of the weather model ensemble, which is the CONUS (≈8 000 000 km2). The NCAR ensemble provides data at a 3-km resolution, therefore the Cowhouse Creek hydrologic model is forced by 181 grid cells from the ensemble. Consequently, what would be considered small deviations from reality in a weather model at the continental scale could be the difference between the model simulating precipitation within the basin or not, and the presence/absence of a hydrologic response. If one were to use a best-guess deterministic forecast, then this could lead to high confidence in inaccurate predictions. For A17, all 10 of the members show simulated peak flows exceeding 150 m3 s−1. With model output two orders of magnitude greater than the observed baseflow of 1.5 m3 s−1, decision-makers can have high confidence that a large hydrologic event is likely to occur in this basin.
c. Utility of alternative threshold metrics
The NWS uses high-water categories as deterministic thresholds to communicate flooding hazard on Cowhouse Creek. Such metrics are simple and useful for describing the current state of Cowhouse Creek relative to historical high-water levels at the location of the USGS gauge. The downside to that simplicity is that the high-water categories at the USGS gauge are not universally applicable to all the infrastructure within the basin. The HST is an alternative that focuses on what is considered safe for a specific vehicle or person. The threshold does not require a long historical record, such as metrics linked to return periods. It can therefore be easily implemented in remote locations. Alternative metrics, such as HST, to the NWS high-water categories can have added value so that the two metrics can be useful in conjunction to forecast across the different types of infrastructure within a catchment.
The NWS high-water categories are easy to visualize because water depth is a physical quantity. The conversion of hydrologic simulated flow to stage at the USGS gauge can only occur with the rating curve created with the long-term history. For the remote locations where the LWXs are located, the addition of a hydraulic model allows for forecasts to be presented in terms of flow stage and flow velocity. The use of a hydraulic model for hindcasting at a specific LWX is more valuable than a streamflow forecast alone. Without context, being told that there is a high probability that streamflow will exceed 500 m3 s−1 on LWX 15 does not have meaning for the people traversing the crossing. However, being told that the there is a high probability that Cowhouse Creek will be greater than 7.5 m deep and 2.0 m s−1 at LWX 15 is much easier for a decision-maker to comprehend and use to motivate preventative action.
Our proof of concept provides a way for probabilistic predictions to be made at remote LWXs. Flood communication can be improved by using metrics in conjunction that focus on both the historical record of a catchment, and the types of infrastructure and vehicles that cross the rivers within that catchment. Visualization techniques look to communicate flood hazards simply, but it is understood that such methods cannot be static in time. The best way to lessen the number of warnings that are ignored by the public is to foster two-way communication between forecasters and the public. The continuation of this work looks to develop these two-way communication channels with decision-makers, so that the visualizations can be tailored to best meet their needs. The HST criteria are vehicle dependent and the inclusion of the automotive industry into these two-way conversations will be beneficial moving forward. By integrating into existing operational modeling environments, this approach is a financially and computationally feasible alternative to current safety measures. Signage, additional USGS gauges, and more advanced warning systems all have installation and operational/maintenance costs. Such implementation is not necessarily practical since LWXs are not highly trafficked by design. Additional proactive information from a modeling approach might be one spoke of many in the wheel of mitigating flood-related fatalities, but we have shown that it has advantages that reactive technology, such as signage and gauges, cannot offer.
Acknowledgments
This work is supported by the U.S. Army Engineer Research and Development Center Construction Engineering Research Laboratory and the Department of Army Management Office Training Ranges and Simulations, Sustainable Range Program, Integrated Training Area Management TATM 121018000. We thank Professor Praveen Kumar and Dr. Andrew Wood, editor of JHM, for insightful discussion.
REFERENCES
Ackerman, C. T., 2009: HEC-GeoRAS, GIS tools for support of HEC-RAS using ArcGIS. Doc. CPD-83, U.S. Army Corps of Engineers Hydrologic Engineering Center, 244 pp., https://www.hec.usace.army.mil/software/hec-georas/documentation/HEC-GeoRAS42_UsersManual.pdf.
Addor, N., S. Jaun, F. Fundel, and M. Zappa, 2011: An operational hydrological ensemble prediction system for the city of Zurich (Switzerland): Skill, case studies and scenarios. Hydrol. Earth Syst. Sci., 15, 2327–2347, https://doi.org/10.5194/hess-15-2327-2011.
Ashley, S. T., and W. S. Ashley, 2008: Flood fatalities in the United States. J. Appl. Meteor. Climatol., 47, 805–818, https://doi.org/10.1175/2007JAMC1611.1.
Balke, K., L. Higgins, S. Chrysler, G. Pesti, N. Chaudhary, and R. Brydia, 2011: Signing strategies for low-water and flood-prone highway crossings. Texas Department of Transportation Research and Technology Implementation Office, 201 pp., https://static.tti.tamu.edu/tti.tamu.edu/documents/0-6262-1.pdf.
Bartholmes, J. C., J. Thielen, M. H. Ramos, and S. Gentilini, 2009: The European Flood Alert System EFAS - Part II: Statistical skill assessment of probabilistic and deterministic operational forecasts. Hydrol. Earth Syst. Sci., 13, 141–153, https://doi.org/10.5194/hess-13-141-2009.
Brunner, G. W., 2016: HEC-RAS river analysis system hydraulic reference manual. Doc. CPD-69, U.S. Army Corps of Engineers Hydrologic Engineering Center, 538 pp., https://www.hec.usace.army.mil/software/hec-ras/documentation/HEC-RAS%205.0%20Reference%20Manual.pdf.
Buizza, R., M. Milleer, and T. N. Palmer, 2007: Stochastic representation of model uncertainties in the ECMWF ensemble prediction system. Quart. J. Roy. Meteor. Soc., 125, 2887–2908, https://doi.org/10.1002/qj.49712556006.
Carstens, R. L., and R. Y.-H. Woo, 1984: Liability and traffic control considerations for low water stream crossings. Tech. Rep., Iowa State University, 55 pp., https://www.semanticscholar.org/paper/Liability-and-Traffic-Control-Considerations-for-Carstens-Woo/ab610c77c286efd347891bba73dfc70fe4b3d24d.
Chow, V. T., 1959: Open-Channel Hydraulics. McGraw-Hill, 680 pp.
Cloke, H., and F. Pappenberger, 2009: Ensemble flood forecasting: A review. J. Hydrol., 375, 613–626, https://doi.org/10.1016/j.jhydrol.2009.06.005.
Demargne, J., J. Brown, Y. Liu, D.-J. Seo, L. Wu, Z. Toth, and Y. Zhu, 2010: Diagnostic verification of hydrometeorological and hydrologic ensembles. Atmos. Sci. Lett., 11, 114–122, https://doi.org/10.1002/asl.261.
Demargne, J., and Coauthors, 2014: The science of NOAA’s operational hydrologic ensemble forecast service. Bull. Amer. Meteor. Soc., 95, 79–98, https://doi.org/10.1175/BAMS-D-12-00081.1.
Demeritt, D., H. Cloke, F. Pappenberger, J. Thielen, J. Bartholmes, and M.-H. Ramos, 2007: Ensemble predictions and perceptions of risk, uncertainty, and error in flood forecasting. Environ. Hazards, 7, 115–127, https://doi.org/10.1016/j.envhaz.2007.05.001.
Demeritt, D., S. Nobert, H. L. Cloke, and F. Pappenberger, 2013: The European flood alert system and the communication, perception, and use of ensemble predictions for operational flood risk management. Hydrol. Processes, 27, 147–157, https://doi.org/10.1002/hyp.9419.
Dowell, D., and Coauthors, 2016: Development of a High-Resolution Rapid Refresh Ensemble (HRRRE) for severe weather forecasting. 28th Conf. on Severe Local Storms, Portland, OR, Amer. Meteor. Soc., JP1.7, https://ams.confex.com/ams/28SLS/webprogram/Paper301555.html.
Duan, Q., F. Pappenberger, A. Wood, H. Cloke, and J. Schaake, 2019: Handbook of Hydrometeorological Ensemble Forecasting. Springer-Verlag, 1528 pp., https://doi.org/10.1007/978-3-642-39925-1.
Emerton, R. E., and Coauthors, 2016: Continental and global scale flood forecasting systems. Wiley Interdiscip. Rev.: Water, 3, 391–418, https://doi.org/10.1002/wat2.1137.
Feldman, A. D., 2000: HEC-HMS, hydrologic modeling system technical reference manual. Doc. CPD-74B, U.S. Army Corps of Engineers Hydrologic Engineering Center, 148 pp., https://www.hec.usace.army.mil/software/hec-hms/documentation/HEC-HMS_Technical%20Reference%20Manual_(CPD-74B).pdf.
Gautam, S., and R. Bhattarai, 2018: Low-water crossings: An overview of designs implemented along rural, low-volume roads. Environments, 5, 22, https://doi.org/10.3390/ENVIRONMENTS5020022.
Gesch, D., M. Oimoen, S. Greenlee, C. Nelson, M. Steuck, and D. Tyler, 2002: The National Elevation Dataset. Photogramm. Eng. Remote Sensing, 68, 5–11.
Hopson, T. M., and P. J. Webster, 2010: A 1–10-day ensemble forecasting scheme for the major river basins of Bangladesh: Forecasting severe floods of 2003–07. J. Hydrometeor., 11, 618–641, https://doi.org/10.1175/2009JHM1006.1.
Jaun, S., and B. Ahrens, 2009: Evaluation of a probabilistic hydrometeorological forecast system. Hydrol. Earth Syst. Sci., 13, 1031–1043, https://doi.org/10.5194/hess-13-1031-2009.
Kumar, P., 2011: Typology of hydrologic predictability. Water Resour. Res., 47, W00H05, https://doi.org/10.1029/2010WR009769.
Lin, P., L. J. Hopper, Z.-L. Yang, M. Lenz, and J. W. Zeitler, 2018: Insights into hydrometeorological factors constraining flood prediction skill during the May and October 2015 Texas hill country flood events. J. Hydrometeor., 19, 1339–1361, https://doi.org/10.1175/JHM-D-18-0038.1.
Lin, Y., 2011: GCIP/EOP surface: Precipitation NCEP/EMC 4km gridded data (GRIB) Stage IV data, version 1.0. UCAR/NCAR Earth Observing Laboratory, accessed 17 August 2018, https://doi.org/10.5065/D6PG1QDD.
Lorenz, E. N., 1969: Atmospheric predictability as revealed by naturally occurring analogues. J. Atmos. Sci., 26, 636–646, https://doi.org/10.1175/1520-0469(1969)26<636:APARBN>2.0.CO;2.
Lyne, V., and M. Hollick, 1979: Stochastic time-variable rainfall-runoff modelling. Institute of Engineers Australia National Conf., Barton, Australia, Institute of Engineers, 89–93.
Martínez-Gomariz, E., M. Gómez, B. Russo, and S. Djordjević, 2018: Stability criteria for flooded vehicles: A state-of-the-art review. J. Flood Risk Manage., 11, S817–S826, https://doi.org/10.1111/jfr3.12262.
McEnery, J., J. Ingram, Q. Duan, T. Adams, and L. Anderson, 2005: NOAA’s advanced hydrologic prediction service. Bull. Amer. Meteor. Soc., 86, 375–386, https://doi.org/10.1175/BAMS-86-3-375.
Mockus, V., 1961: Watershed lag. Doc. ES-1015, U.S. Department of Agriculture, Soil Conservation Service.
Murphy, J., and T. Palmer, 1986: A real time extended range forecast by an ensemble of numerical integrations. Meteor. Mag., 337, 348.
Nathan, R. J., and T. A. McMahon, 1990: Evaluation of automated techniques for base flow and recession analyses. Water Resour. Res., 26, 1465–1473, https://doi.org/10.1029/WR026i007p01465.
NOAA, 2016: High water level terminology. NOAA/National Weather Service, accessed 10 March 2020, https://www.weather.gov/aprfc/terminology.
NOAA, 2019a: River Forecast Centers. NOAA/National Weather Service, accessed 10 March 2020, https://www.weather.gov/jetstream/rfcs.
NOAA, 2019b: Turn Around Don’t Drown. NOAA/National Weather Service, accessed 10 March 2020, http://www.weather.gov/safety/flood-turn-around-dont-drown.
NOAA, 2020a: About AHPS. NOAA/National Weather Service, accessed 10 March 2020, https://water.weather.gov/ahps/about/about.php.
NOAA, 2020b: The National Water Model. NOAA/Office of Water Prediction, accessed 10 March 2020, https://water.noaa.gov/about/nwm.
NOAA, 2020c: Weather related fatality and injury statistics. NOAA/National Weather Service, accessed 10 March 2020, https://www.weather.gov/hazstat/.
NRCS, 2010: Time of concentration. National Engineering Handbook, Natural Resources Conservation Service, 29 pp., https://directives.sc.egov.usda.gov/OpenNonWebContent.aspx?content=27002.wba.
NRCS, 2019: Soil Survey Geographic (SSURGO) database. Soil Survey Staff, Natural Resources Conservation Service, https://sdmdataaccess.sc.egov.usda.gov.
Pagano, T. C., and Coauthors, 2014: Challenges of operational river forecasting. J. Hydrometeor., 15, 1692–1707, https://doi.org/10.1175/JHM-D-13-0188.1.
Palmer, T. N., 2001: A nonlinear dynamical perspective on model error: A proposal for non-local stochastic-dynamic parametrization in weather and climate prediction models. Quart. J. Roy. Meteor. Soc., 127, 279–304, https://doi.org/10.1002/QJ.49712757202.
Palmer, T. N., 2017: The primacy of doubt: Evolution of numerical weather prediction from determinism to probability. J. Adv. Model. Earth Syst., 9, 730–734, https://doi.org/10.1002/2017MS000999.
Pappenberger, F., E. Stephens, J. Thielen, P. Salamon, D. Demeritt, S. J. van Andel, F. Wetterhall, and L. Alfieri, 2013: Visualizing probabilistic flood forecast information: Expert preferences and perceptions of best practice in uncertainty communication. Hydrol. Processes, 27, 132–146, https://doi.org/10.1002/hyp.9253.
Ramos, M.-H., T. Mathevet, J. Thielen, and F. Pappenberger, 2010: Communicating uncertainty in hydro-meteorological forecasts: Mission impossible? Meteor. Appl., 17, 223–235, https://doi.org/10.1002/met.202.
Rawls, W. J., D. L. Brakensiek, and N. Miller, 1983: Green-Ampt infiltration parameters from soils data. J. Hydraul. Eng., 109, 62–70, https://doi.org/10.1061/(ASCE)0733-9429(1983)109:1(62).
Ritter, A., and R. Muñoz-Carpena, 2013: Performance evaluation of hydrological models: Statistical significance for reducing subjectivity in goodness-of-fit assessments. J. Hydrol., 480, 33–45, https://doi.org/10.1016/j.jhydrol.2012.12.004.
Rossmiller, R. L., and Coauthors, 1983: Design manual for low water stream crossings. Tech. Rep., Iowa State University, 55 pp.
Schaake, J. C., T. M. Hamill, R. Buizza, and M. Clark, 2007: Hepex: The hydrological ensemble prediction experiment. Bull. Amer. Meteor. Soc., 88, 1541–1548, https://doi.org/10.1175/BAMS-88-10-1541.
Schaap, M. G., F. J. Leij, and M. T. van Genuchten, 2001: ROSETTA: A computer program for estimating soil hydraulic parameters with hierarchical pedotransfer functions. J. Hydrol., 251, 163–176, https://doi.org/10.1016/S0022-1694(01)00466-8.
Scharffenberg, B., M. Bartles, T. Brauer, M. Fleming, and G. Karlovits, 2018: HEC-HMS, hydrologic modeling system user’s manual. U.S. Army Corps of Engineers Hydrologic Engineering Center, 640 pp., https://www.hec.usace.army.mil/software/hec-hms/documentation/HEC-HMS_Users_Manual_4.3.pdf.
Schwartz, C. S., G. S. Romine, R. A. Sobash, K. R. Fossell, and M. L. Weisman, 2015: NCAR’s experimental real-time convection-allowing ensemble prediction system. Wea. Forecasting, 30, 1645–1654, https://doi.org/10.1175/WAF-D-15-0103.1.
Shand, T., R. Cox, M. Blacka, and G. Smith, 2011: Appropriate safety criteria for vehicles. Australian Rainfall and Runoff Rep. P10/S2/020, 29 pp., http://arr.ga.gov.au/__data/assets/pdf_file/0004/40486/ARR_Project_10_Stage2_Report_Final.pdf.
USGS, 2019: National Hydrography Dataset. U.S. Geological Survey, https://www.usgs.gov/core-science-systems/ngp/national-hydrography.
van Genuchten, M. T., 1980: A closed-form equation for predicting the hydraulic conductivity of unsaturated soils. Soil Sci. Soc. Amer. J., 44, 892–898, https://doi.org/10.2136/sssaj1980.03615995004400050002x.
Xia, Y., and Coauthors, 2012: Continental-scale water and energy flux analysis and validation for the North American Land Data Assimilation System project phase 2 (NLDAS-2): 1. Intercomparison and application of model products. J. Geophys. Res., 117, D03109, https://doi.org/10.1029/2011JD016048.
Yang, L., and Coauthors, 2018: A new generation of the United States National Land Cover Database: Requirements, research priorities, design, and implementation strategies. ISPRS J. Photogramm. Remote Sens., 146, 108–123, https://doi.org/10.1016/j.isprsjprs.2018.09.006.
Zappa, M., and Coauthors, 2010: Propagation of uncertainty from observing systems and NWP into hydrological models: COST-731 working group 2. Atmos. Sci. Lett., 11, 83–91, https://doi.org/10.1002/asl.248.
Zappa, M., F. Fundel, and S. Jaun, 2013: A “peak-box” approach for supporting interpretation and verification of operational ensemble peak-flow forecasts. Hydrol. Processes, 27, 117–131, https://doi.org/10.1002/hyp.9521.
