190 JOURNAL OF HYDROMETEOROLOGY VOLUME 7

The Use of an Automated Nowcasting System to Forecast Flash Floods in an

Urban Watershed
HATIM O. SHARIF*
Lawrence Berkeley National Laboratory, Berkeley, California, National Center for Atmospheric Research, Boulder, Colorado, and
Princeton University, Princeton, New Jersey

DAVID YATES
⫹
National Center for Atmospheric Research, and Department of Civil Engineering, University of Colorado, Boulder, Colorado

RITA ROBERTS AND CYNTHIA MUELLER

National Center for Atmospheric Research, Boulder, Colorado

(Manuscript received 19 January 2005, in final form 16 May 2005)

ABSTRACT

Flash flooding represents a significant hazard to human safety and a threat to property. Simulation and
prediction of floods in complex urban settings requires high-resolution precipitation estimates and distrib-
uted hydrologic modeling. The need for reliable flash flood forecasting has increased in recent years,
especially in urban communities, because of the high costs associated with flood occurrences. Several storm
nowcast systems use radar to provide quantitative precipitation forecasts that can potentially afford great
benefits to flood warning and short-term forecasting in urban settings. In this paper, the potential benefits
of high-resolution weather radar data, physically based distributed hydrologic modeling, and quantitative
precipitation nowcasting for urban hydrology and flash flood prediction were demonstrated by forcing a
physically based distributed hydrologic model with precipitation forecasts made by a convective storm
nowcast system to predict flash floods in a small, highly urbanized catchment in Denver, Colorado. Two
rainfall events on 5 and 8 July 2001 in the Harvard Gulch watershed are presented that correspond to times
during which the storm nowcast system was operated. Results clearly indicate that high-resolution radar-
rainfall estimates and advanced nowcasting can potentially lead to improvements in flood warning and
forecasting in urban watersheds, even for short-lived events on small catchments. At lead times of 70 min
before the occurrence of peak discharge, forecast accuracies of approximately 17% in peak discharge and 10
min in peak timing were achieved for a 10 km2 highly urbanized catchment.

1. Introduction 2003) because the hydraulic properties of these areas,

such as large expanses of impervious areas, smoothed
Hydrologic modeling of urban flood potential has and compacted land surfaces, and modification of natu-
witnessed an upsurge in interest recently (e.g., Ogden ral flow paths, create conditions suitable for reduced
et al. 2000; Lee and Heaney 2003; Zhang and Smith infiltration, storage, and friction losses, creating condi-
tions favorable to high-peak flow responses. The prob-
ability of flooding from a given storm is typically higher
in urban areas (e.g., Konrad and Booth 2002); a striking
* Current affiliation: University of Texas at San Antonio, San
example is that on the same evening of the flash flood
Antonio, Texas.
⫹ The National Center for Atmospheric Research is sponsored that devastated Fort Collins, Colorado, on 27 July 1997,
by the National Science Foundation. a more intense storm occurred in rural Colorado with
no reported injuries or significant damage (M. Kelsch
2003, personal communication). Simulation and predic-
Corresponding author address: Dr. Hatim Sharif, Civil and En-
vironmental Engineering Dept., University of Texas at San An- tion of floods in complex urban settings requires dis-
tonio, 6900 N. Loop 1604 W., San Antonio, TX 78249. tributed precipitation estimates and distributed hydro-
E-mail: [email protected] logic modeling.

© 2006 American Meteorological Society

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC
FEBRUARY 2006 SHARIF ET AL. 191

The need for reliable flash flood forecasting has in- ing (TITAN) algorithm (Dixon and Wiener 1993) and
creased in recent years, especially in urban communi- the Strom Cell Identification and Tracking algorithm
ties, because of the high costs associated with flood used by the National Severe Storm Laboratory’s Warn-
occurrences; in the United States, an average of 100 ing Decision Support System (Johnson et al. 1998).
people lose their lives in floods annually, with flood More complex nowcast systems employ radar data in
damage averaging more than $2 billion (see htpp:// combination with meteorological observations, numeri-
www.noaa.gov/floods.html). A number of flood control cal weather prediction, and feature-detection algo-
districts, including that of Denver, Colorado, use the rithms to nowcast storm evolution. Examples of these
Automated Local Evaluation in Real-Time (ALERT) sophisticated systems that are used operationally in-
method developed in the 1970s by the National clude the Federal Aviation Administration Regional
Weather Service. The ALERT system continues to pro- Convective Weather Forecast and Terminal Weather
vide valuable early flood detection and decision sup- Convective Forecast System (Boldi et al. 2002) and the
port for several urban communities, including the Den- Auto-Nowcaster (ANC) system developed at the Na-
ver area, with more than 140 gauging stations. ALERT tional Center for Atmospheric Research (NCAR)
systems depend mainly on automated rain gauge net- (Mueller et al. 2003). The ANC is the nowcast system
works that can provide rainfall estimates in near–real used in this study, which makes use of multiple data
time by radio telemetry. However, the densities of ingest products (gridded radar fields, surface mesonet
these networks are typically not fine enough for accu- data, sounding data, etc.) to produce, among other
rate flash flood forecasting (Bedient et al. 2003). An- things, 0–60-min quantitative precipitation nowcasts.
other problem with rain gauge networks is that they are The overall objective of this study is to demonstrate
subject to degraded levels of accuracy with increased the potential benefits of high-resolution weather radar
precipitation intensities, such as those associated with data, physically based distributed hydrologic modeling,
flood-producing storms. In addition, gauge-based storm and quantitative precipitation nowcasting for urban hy-
data are less useful in real-time storm tracking com- drology and flash flood forecasting. Radar reflectivity
pared to those of radar. data from WSR-88D radar and forecast fields are used
The Weather Surveillance Radar-1988 Doppler to compute high-resolution rainfall estimates that are
(WSR-88D) Next Generation Weather Radar (NEXRAD) input to a state-of-the-art physics-based distributed-
is capable of detecting precipitation at a resolution and parameter hydrologic model to forecast flooding over a
areal extent previously impossible with traditional rain small, highly urbanized catchment in Denver. Two rain-
gauge networks and has brought unprecedented ad- fall events that occurred on 5 and 8 July 2001 in the
vances in estimating areally distributed, real-time rain- Denver region, during which the ANC system had been
fall data for both hydrologic and hydrometeorological operated, are presented. The observed rainfall and
applications, including flood warning and forecasting streamflow data on 8 July were used to validate the
(National Research Council 1996). In addition to the hydrologic model. Simulations driven by rain gauge and
high temporal and spatial resolution of the radar- radar-rainfall estimates are compared to highlight the
rainfall data, an equally important factor in the future benefit of distributed rainfall information. Nowcasts for
utility of radar-rainfall data for operational use in urban both events were used to evaluate the ability of ANC to
hydrology is the cost savings that may arise from the produce rainfall nowcasts that are useful for urban hy-
use of radars over rain gauge networks (Tilford et al. drologic analysis and forecasting. Hydrologic model
2002). Moreover, several nowcast systems use radar to outputs using nowcast precipitation fields were com-
provide quantitative precipitation forecasts (Wilson et pared to outputs produced using the radar-rainfall es-
al. 1998) that can potentially afford great benefits to timates, which were assumed to be the best available
flood warning and short-term forecasting in urban set- rainfall information.
tings.
Nowcasts are typically defined as short-time and 2. The hydrologic model
space-specific forecasts of periods less than a few hours, The Gridded Surface Subsurface Hydrologic Analy-
and may include storm initiation, growth, dissipa- sis (GSSHA) model (Downer and Ogden 2004) is a
tion, and storm features such as wind speeds and direc- physically based distributed-parameter hydrologic
tion and precipitation rates. Some of these nowcast sys- model. GSSHA is a reformulation and enhancement
tems employ gridded radar data for storm analysis and of the finite-difference Cascade Two-Dimensional
trending and to extrapolate storms positions (Mueller (CASC2D) model (Ogden 2000) that simulates infiltra-
et al. 2003). Examples of this type are the Thunder- tion-excess (Hortonian) runoff. It extends CASC2D to
storm Identification, Tracking, Analysis, and Nowcast- allow the subsurface water components of the hydro-

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

192 JOURNAL OF HYDROMETEOROLOGY VOLUME 7

logic balance, both saturated and unsaturated zones, to in space. Flux updating, as described by Kirkland
be included in watershed simulations. GSSHA simu- (1991), is used to ensure mass conservation. Iterations
lates two-dimensional overland flow, one-dimensional can also be used to improve accuracy and mass balance.
channel routing, rainfall distribution, canopy intercep- Internal time step limitations, as described by Belmans
tion, microtopography, infiltration, and evapotranspira- et al. (1983), also help keep the model stable, accurate,
tion using finite-difference and finite-volume methods. and mass conserving.
Additionally, GSSHA uses a one-dimensional finite- Once ponding occurs on a grid cell, surface water is
difference solution of Richards’ equation to simulate accumulated until the specified retention depth of the
the unsaturated zone. GSSHA operates on a digital cell is exceeded. Thereafter, the overland flow is routed
elevation model of the watershed using square grids in two orthogonal directions using Manning’s equation
that typically range from 10 to 1000 m on a side. with the diffusion waveform of the de St. Venant equa-
Infiltration into the soil is optionally modeled using tions to estimate friction slope. When the overland flow
either the Green and Ampt (1911) method, or the reaches a model grid cell that is specified as a channel
Green and Ampt redistribution method (Ogden and cell, the flow is passed into the channel and routed
Saghafian 1997). This infiltration technique is similar to using a one-dimensional routing technique.
the method described by Smith et al. (1993), with the
assumption of rectangular soil moisture profiles and the
addition of an analytically derived unsaturated capillary 3. The 8 July 2001 flood event
head term (Ogden and Saghafian 1997). The infiltration
The 2001 flood season in Denver was above average
option allows accurate simulation of infiltration when
in terms of the number of flood messages issued by the
there are multiple ponding periods. The original Green
Flood Control District. The week of 14 July was par-
and Ampt (1911) equation is expressed in GSSHA as:
ticularly wet with flash flood warnings issued on 3 days

f⫽K 冋 Hc共␪e ⫺ ␪i兲

F 册
⫹1 , 共1兲
(8, 10, and 13 July) and flash flood watches for three
others. The 8 July storm event produced the worst
flooding of the year because flash flooding was ob-
where f is the infiltration rate (L/T), Ks the soil satu- served on several watersheds, including the Harvard
rated hydraulic conductivity (L/T), Hc the Green and Gulch, and along a major highway. Annual peaks were
Ampt capillary head term (L), ␪e the soil effective po- recorded by 16 stream gauges, with 5 breaking their
rosity (dimensionless), ␪i the soil initial water content historic marks. Harvard Gulch experienced a record
(dimensionless), and F the cumulative infiltrated depth flood with much of the upper basin receiving more than
(L). 3 in. of rain. In this study, the flooding of Harvard
GSSHA also provides detailed modeling of the soil Gulch is simulated to validate the distributed modeling
water profile in the unsaturated zone. It solves the one- approach for this type of watershed. Several specific
dimensional (in the vertical direction) head-based form tasks are required: acquiring and processing radar-
of Richards’ equation, rainfall data for the event from the Denver WSR-88D

冋 冉 冊册
radar archived level II data, acquiring rainfall observa-
⭸␺ ⭸ ⭸␺ tions from five U.S. Geological Survey (USGS) rain
C共␺兲 ⫺ K共␺兲 ⫺1 ⫺ W, 共2兲
⭸t ⭸z ⭸z gauges located in or very close to Harvard Gulch, and
using the rainfall data as input to GSSHA to simulate
where C is the specific moisture capacity, ␺ soil capil-
the 8 July flood.
lary head (L), z vertical coordinate (downward posi-
tive) (L), t time (T), K(␺) effective hydraulic conduc-
tivity (L/T), and W a source/sink term (L/T). The head-
a. Rainfall data
based formulation allows for the solution of Richards’ Rainfall observations from five tipping-bucket rain
equation in both saturated and unsaturated conditions gauges operated by the USGS were obtained (bucket
(Haverkamp et al. 1977). Either the Brooks and Corey size of 0.01 in.; see Fig. 1).
(1964) equations, as extended by Hutson and Cass The Denver WSR-88D radar is located about 40 km
(1987), or the Haverkamp et al. (1977) equations, as from the watershed and operated continuously during
modified by Lappala et al. (1987), can be used to define the 5 and 8 July cases. Rainfall was estimated from the
the soil pressure (␺)–water content (␪), ␺–hydraulic three-dimensional volume scan reflectivity fields using
conductivity (K ), and ␺–water capacity (C) relation- the National Weather Service default relationship be-
ships. The implicit solution to Richards’ equation is tween the radar reflectivity Z (mm6 m⫺3) and the rain-
first-order accurate in time and second-order accurate fall rate R (mm h⫺1):

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

FEBRUARY 2006 SHARIF ET AL. 193

FIG. 1. The Harvard Gulch watershed, storm-total radar-estimated rainfall, and location of
USGS rain gauges. The street networks of the urban area are shown in the background.

R ⫽ aZb. 共3兲 http://landcover.usgs.gov, respectively). Land use/cover

features were modified based on information obtained
The values of a and b in Eq. (3) are a ⫽ 0.017 and b ⫽ from the Denver Urban Drainage and Flood Control
0.714. These parameters correspond to the relation Z ⫽ District and a site visit. General terrain slope of the
300R1.4 (Battan 1973). This is the relationship typically catchment is from east to west with local slope values
applied by the National Weather Service for this radar ranging from 0.5 to 2%. Base flow is typically less than
(Fulton 1999). Also, for this radar a rain-rate threshold 0.12 m3 s⫺1. Impervious areas cover about 40% of the
(also called the hail cap) of 74.7 mm h⫺1 is specified by watershed, while soils in the watershed are classified as
the National Weather Service (Fulton et al. 1998; Ful- the Soil Conservation Service Group B. The Harvard
ton 1999). Rain rates are capped at this value to prevent Gulch is heavily channelized and is drained by a com-
hail contamination associated with convective storms, bination of trapezoidal, rectangular, and closed-conduit
which can significantly enhances radar reflectivity val- channels. Details of the channel cross sections and
ues. All reflectivity values above 51 dBZ (which corre- properties were also provided by the Flood Control
sponds to the rain rate of 74.7 mm h⫺1) were assumed District.
to have been a result of hail presence and were reset to
51 dBZ [radar reflectivity is usually expressed in deci- c. Hydrologic model validation
bels of Z, that is, dBZ, where dBZ ⫽ 10 log10(Z )]. On
A number of studies discussed the effects of spatial
the other hand, all reflectivity values below 25 dBZ are
variability in watershed characteristics on the hydro-
assumed to be the result of clear-air return not associ-
logic response. Woolhiser (1996) showed that infiltra-
ated with rainfall and were eliminated. Radar-rainfall
tion-excess runoff, typically associated with urban and
data are then used without further adjustments. The
semiarid catchments, is strongly influenced by the spa-
radar-rainfall field has a 5–6-min temporal resolution
tial variability of soil hydraulic conductivity. Merz and
and a special resolution of 1 km ⫻ 1°.
Plate (1997) demonstrated that the effect of spatial
variability, which is typically large for midsized events,
b. Watershed characteristics
decreases for very small and very large events. Ogden
The Harvard Gulch (Fig. 1) has an area of approxi- et al. (2000) reached similar conclusions when they ana-
mately 10.2 km2 of mixed urban land use and is twice as lyzed the July 1997 flash flood in Fort Collins.
long as it is wide. A 30-m grid size was found sufficient The GSSHA model was used to simulate the spatially
to describe the topography and land surface features of varied hydrologic response of Harvard Gulch on 8 July.
the watershed. Topography and land use/cover data The simulation was simplified for two reasons. First, it
were obtained from the USGS digital database (Earth is assumed that runoff from the storm is essentially gen-
Resources Observing Systems Data Center, National erated by the infiltration-excess mechanism. Because of
Elevation Data and National Land Cover Data infor- the semiarid climate and the low soil hydraulic conduc-
mation available online at http://ned.usgs.gov and tivity (the dominant soil type is silty loam), lateral sub-

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

194 JOURNAL OF HYDROMETEOROLOGY VOLUME 7

surface flow and saturation from below resulting from TABLE 1. Model parameters for different land use/cover types.
the rising of the water table for this short, high-intensity
Manning roughness Retention
storm is likely insignificant. Second, no complex hy- Land use/cover type coefficients storage (mm)
draulic structures exist in the simulated catchment.
Impervious areas 0.02 1.3
Nonetheless, we included as much detail of the land Industrial 0.15 1.4
surface features and the drainage system as was avail- Public areas 0.20 1.4
able. Residential (lawns) 0.20 5.0
Soil hydraulic properties are available from the Natu-
ral Resources Conservation Service county map. Be-
cause those are approximate values, we used values wet initial conditions for actually extremely dry initial
computed through calibration for the Fort Collins flood conditions. Merz and Plate (1997) also downplayed the
(Ogden et al. 2000) where similar soil properties exist. impact of soil moisture variability for the case of large
We used the same parameters used in Fort Collins for events. There were some rainfall events in the weeks
identical soil types because the two watersheds are lo- before this event but no events since 5 July. Certainly,
cated in the same region and topography and the cli- the watershed was not very dry, and given the tempera-
matology are similar. Moreover, sensitivity experiments ture of 5–8 July it is very unlikely that the watershed
indicate that even a factor of 2 change in soil hydraulic was still extremely wet on 8 July. We used an interme-
conductivity does not have a significant effect on the diate value of initial soil moisture to minimize the effect
runoff hydrograph driven by heavy precipitation for on our results.
this catchment (H. O. Sharif et al. 2003, unpublished Rainfall data from the WSR-88D radar and five
manuscript). There are two dominant soil types: silty USGS rain gauges were used as input to the hydrologic
loam covers about 60% of the watershed and the re- model. No calibration was done at this stage. Unfortu-
mainder is impervious. Silty loam is assigned a mean nately, the USGS stream gauge at the watershed outlet
saturated hydraulic conductivity (K) of 0.34 cm h⫺1, a stopped recording 15 min after the beginning of the 8
mean effective porosity of 0.5, and a mean capillary July event because of the discharge intensity. Conse-
head of 16.7 cm. The values of the hydraulic parameters quently, there are only three points on the actual outlet
are randomly varied around the means to account for hydrograph that are comparable with the hydrograph
soil heterogeneity. Computed percent imperviousness simulated by the model.
for 33 subareas within the Harvard Gulch was obtained Information from the stream gauge operated by the
from the Denver Flood Control District. Impervious Flood Control District was not of much help because of
model grid cells within these subareas are distributed the effects of a detention basin and the inflow from
randomly to represent the specified percentages. In another watershed (Dry Creek) that merges with the
some cases, impervious cells are aligned along major Harvard Gulch channel between the two stream
streets – the size of the model cell is 30 m. Values for gauges. Figure 2 shows that simulation results based on
overland flow roughness coefficients were obtained
from published sources (Soil Conservation Service
1986; Ogden et al. 2000) and are listed in Table 1. Re-
tention depth is assumed to be a function of land use,
and values were assigned following Tholin and Keifer
(1960). Retention depths for different land use/cover
types are also listed in Table 1. Previous applications of
the model showed that the value of the retention depth
has some effect on the runoff volume and hydrograph
shape (e.g., Ogden et al. 2000). Unlike hydraulic prop-
erties, which were stochastically distributed, we as-
signed constant values of the roughness coefficient (or
retention depth) for each land use/cover type. After
vigorous sensitivity testing on a watershed where the
Hortonian runoff was dominant, Ogden and Dawdy
(2003) concluded that for big rainfall events the initial
soil moisture state can have a significant effect, but only FIG. 2. Simulated runoff hydrographs driven by radar and rain
if the soils are extremely wet or dry, that is, a significant gauge estimates. Available USGS stream gauge measurements
difference between results if one assumes extremely are shown (open circles).

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

FEBRUARY 2006 SHARIF ET AL. 195

FIG. 3. Simplified schematic of the ANC system.

radar and rain gauge data match discharge values re- schematic of the ANC system is provided in Fig. 3. The
corded by the USGS stream gauge both in terms of steps in producing the nowcast are briefly reviewed
timing and magnitudes. The inverse-distance method is here to give an overview of the entire system.
used to construct rain fields from rain gauge data. The The system receives operational meteorological data
peak discharge computed using rain gauge data is from several sources, including radar (mainly WSR-88D),
higher than the one computed using radar data by 20%, satellite, surface stations (including special mesonets),
although there is no difference in timing. lightning, profilers, numerical weather model, and ra-
The authors learned that USGS subsequently esti- diosondes. The first step in producing the nowcast is
mated a peak discharge of 2040 cfs at the Harvard running a group of analysis algorithms on these datasets
Gulch Park gauge based on their high-water marks field to calculate predictor storm fields. Analysis algorithms
survey, which is less than our estimated peak discharge include data quality control routines, TITAN, and a
based on radar rainfall by about 9% (K. Stewart, Den- Tracking Radar Echoes by Correlation algorithm
ver Urban Drainage and Flood Control District official, (Tuttle and Foote 1990) that retrieves the three-dimen-
2004, personal communication). This makes us more sional wind speed and direction, and a numerical
confident about the simulation results, especially when boundary layer model and its adjoint (Variational
radar data are used, having accurate simulation of the Doppler Radar Analysis System; Sun and Crook 2001).
rising limb of the hydrograph reasonable estimate of A major feature of ANC is its ability to use this
the peak discharge. The total accumulated precipitation boundary layers model and other feature-detection al-
from the event is virtually the same using radar or gorithms to identify and characterize the boundary
gauge information. One has to be cautious about sug- layer. It is possible for a forecaster to interact with
gesting that radar-rainfall data are better for this case
ANC by manually inputting the position of boundaries
because radar estimates can be biased for one reason or
in an optional intermediate step. The predictor fields
another. The radar estimates have spatially and tempo-
developed in the first step are combined using a fuzzy
rally variable errors even if the mean bias is eliminated;
logic approach. The fuzzy logic approach uses member-
Sharif et al. (2002) demonstrated examples of such
ship functions to map the predictor fields to the likeli-
cases.
hood of storms (likelihood fields). The dimensionless
likelihood fields are meant to represent the relationship
4. The NCAR ANC between the predictor fields and the existence of con-
vective storms at validation time. The likelihood fields
The Auto-Nowcaster used in this study is a software are weighted and summed to produce a combined like-
system that combines output from feature-detection al- lihood field. The combined likelihood field is filtered
gorithms and thunderstorm extrapolation/trending soft- and thresholded in a third step to generate the nowcast
ware in a “data fusion” system to produce short-term areas of convective activity.
0–1-h thunderstorm forecasts (nowcasts). A simplified In ANC deployments to date, the automated now-

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

196 JOURNAL OF HYDROMETEOROLOGY VOLUME 7

casts are used as guidance by the forecasters. The major a. 5 July 2001 case
advantage of ANC is its ability to forecast storm, ini-
Mueller et al. (2003) provided a detailed review of
tiation, growth, and dissipation. For this application,
this case. Here, a short summary is presented. On 5
the simulated new reflectivity fields are converted to July, synoptic-scale forcing over Colorado was weak.
rain rates using the WSR-88D Z–R relationship. The steering-level winds were approximately 5 m s⫺1
As with virtually all automated nowcast systems, from the southeast. Surface dewpoints were ⬃10°C,
the existence of mountains imposes a serious challenge which is typical of values observed on days when thun-
on utilizing the full capabilities of the ANC system. derstorms occur in the Denver area in the summer con-
TITAN is less affected by the complex terrain and is vective season. Storms initiated, grew, and dissipated
used to detect and extrapolate the position of existing within the radar domain over a 2.5-h period. The storms
thunderstorms. TITAN uses three-dimensional Carte- were located approximately 40 km east of the foothills
sian radar data as input. It employs a centroid-based of the Rocky Mountains near the Denver International
methodology for identifying storms as objects and Airport. In this region, the Denver WSR-88D has a
matching these objects to those at a subsequent time to strong signal from clear-air return and the boundary
produce storm-track information. Based on past storm layer is well defined. Figure 4b shows an example a
trends, TITAN predicts future storm location and size. 60-min ANC precipitation rate nowcast. The radar re-
Products from TITAN are input into the Auto-Now- flectivity at nowcast time is shown in Fig. 4a. The 60-
caster and provide information on storm size, move- min nowcast predicts storm initiation and growth (as
ment, and trend. The TITAN algorithm does not rely indicated by an increased area of convection in the
on boundary layer winds and can be used in mountain- nowcast). In this case, boundary layer convergence
ous terrain. This algorithm, along with satellite-based fields associated with the collision of multiple bound-
algorithms, is the major component of the Auto- aries were the main contributors to the nowcast. Figure
Nowcaster that are typically employed in complex ter- 4c shows the radar reflectivity at validation time. The
rain when boundary layer winds are not obtainable. ANC nowcast compares well with the observations.
Outputs from all of the various algorithms described Later in the time period, the dissipation of the system is
above are combined by the Auto-Nowcaster system to also handled well by the ANC. This is a case demon-
produce nowcasts every 5–6 min of thunderstorm ini- strating the performance of the ANC when information
tiation, growth, decay, and movement. Nowcasts using about the early growth of storm was available. The per-
the full capacity of the ANC and nowcasts based on formance is typical for cases when all data needed are
TITAN extrapolations will be used in this study to high- available and full capabilities of the ANC can be uti-
light the differences between the two approaches. The lized.
hydrologic model will be used to simulate runoff hy-
drographs based on the ANC nowcast for 2 days in July b. 8 July 2001 case
2001. This is a case when topography limits the amount of
information available for the ANC especially early in
the growing stage of the storm. On 8 July 2001, storms
5. Examples of rainfall nowcasting developed in the early afternoon over the Rocky
Mountains, located approximately 60 km west of the
The ANC nowcast of two storm events that produced center of the Denver metropolitan area. The more in-
significant rainfall in the Denver metropolitan area are tense thunderstorms, gust fronts, and associated con-
demonstrated. The 5 July 2001 storm represents a case vergence boundaries moved off the foothills and prop-
where the ANC produced reasonably accurate now- agated toward Denver at 2200 UTC. One thunderstorm
casts. On 8 July 2001, flash flooding occurred from outflow located SW of Denver was particularly intense,
storms over the Denver urban area, but the perfor- initiating new storm development by 2230 UTC as it
mance of the ANC was limited because of the presence moved toward Denver. The new storms (see Fig. 5a)
of complex terrain located approximately 60 km west of grew rapidly over a 30-min period and produced heavy
Denver. Given the difference in the performance of precipitation, hail, and flash flooding over portions of
ANC during these two events, the main motivation for the Denver urban area and at several of the Denver
studying these cases is to evaluate the hydrologic fore- streams by 2300 UTC (Fig. 5b). The areal extent of
casts based on these nowcasts. A brief description of precipitation rain rates ⬎100 mm h ⫺1 essentially
the meteorological synoptics of these storms and a dis- doubled during the 30-min period shown in Figs. 5a,b.
cussion of the ANC results are provided below. Both TITAN and the Auto-Nowcaster ran in real time

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

FEBRUARY 2006 SHARIF ET AL. 197

on 8 July. The Auto-Nowcaster produced 30- and 60-

min nowcasts of rain rate at 6-min intervals during the
evolution of this event.
The 30-min storm extrapolation (TITAN) rainfall
rates compared relatively well to the gauge rain rates
during the first 30 min of the flash flood event, with
differences of 20%–30% observed in peak magnitudes
between the gauge measurements and extrapolated
storm rain rates. The ANC 30-min rain-rate nowcasts
showed some improvement over the extrapolation rain
rates throughout the flash flood event. The difference
in performance between ANC nowcasts and TITAN
extrapolations can be seen in Figs. 5c,d at the time of
peak rainfall at 2300 UTC. While there is not a major
difference in the two rain-rate plots, it can be seen that
the ANC nowcasts provide a better representation of
the larger areal extent of the ⬎100 mm h⫺1 rain rates
that actually occurred (see Fig. 5b). This is because the
Auto-Nowcaster is able to provide nowcasts of storm
growth and decay, in addition to nowcasting the initia-
tion of new storms—a capability that does not exist
within extrapolation techniques such as TITAN. The
Auto-Nowcaster system relies on the timely detection
of both a surface convergence boundary and the detec-
tion of clouds growing aloft, above the boundary, in
order to produce accurate 30-min nowcasts of rainfall
rate. The complex terrain of the Rocky Mountains lo-
cated upstream of Denver prevented early observation
of the thunderstorm outflows and of the rapid growth
of the storms 15 min prior to their impact on the Har-
vard Gulch basin. Thus, longer-period nowcasts failed
to capture storm initiation and predict the intensifica-
tion of outflows, which is crucial to nowcasting the
rapid intensification of storms.

6. Runoff predictions based on rainfall nowcasts

Rainfall estimates and nowcasts for the 5 and 8 July
events were used as inputs to the GSSHA model to
compute corresponding runoff nowcasts. The nowcast
precipitation fields have roughly the same spatial and
temporal resolutions as the radar estimates used to pro-
duce them—1-km grids at 5–6-min intervals. The now-
cast rainfall fields are generated at lead times of 30 and
60 min. Hydrographs for the events computed using
radar estimates directly are assumed to be the refer-
ences against which hydrographs driven by the nowcast
precipitation fields are compared; note that the former
FIG. 4. An example ANC precipitation rate nowcast for the 5 have a 5–6-min temporal resolution and a 1 km ⫻ 1°
Jul event: (a) reflectivity at forecast time (dBZ ), (b) 60-min fore- spatial resolution.
cast of rainfall rate (mm h⫺1), and (c) reflectivity at valid time for The hydrologic model was run first with rainfall in-
60-min forecast (dBZ ).
puts from the ANC for the 5 July event to evaluate the
accuracy of the simulated hydrologic forecasts. As men-

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

198 JOURNAL OF HYDROMETEOROLOGY VOLUME 7

FIG. 5. An example ANC precipitation rate nowcast for the 8 Jul event: (a) precipitation rate at forecast time
(mm h⫺1) (2230 UTC), (b) observed precipitation rate (mm h⫺1) (2300 UTC), (c) TITAN forecast of (b) made at
2230 UTC, and (d) ANC forecast of (b) made at 2230 UTC.

tioned earlier, 5 July represents a case where the ANC errors in peak discharge and runoff volume do not de-
produced reasonable nowcasts. Every 5 min the ANC crease with time and have multiple spikes. One would
generated discrete 30- and 60-min nowcasts (two val- expect that the error plotted in Fig. 6 would decrease
ues). The hydrologic model uses precipitation fields with time if the skill of the ANC is consistent, and
from radar estimates up to the nowcast time and the 30- actually the latter is true for this experiment. But, the
and 60-min nowcasts. Because the rainfall inputs to the error values fluctuate because of the interpolation be-
hydrologic model have a 5–6-min resolution, estimates tween the forecast values. Even if the forecasts are per-
for intermediate time steps were interpolated linearly fect, two instantaneous values cannot represent a rain-
from the 30- and 60-min nowcast fields. As the ANC fall process with high temporal variability. For some
produces new precipitation nowcasts (every 5–6 min), cases, for example, 55 min prior to the peak occurrence,
previous nowcast values are discarded. In Fig. 6, the the two values represented the rainfall variability quite
lead time represents the difference between the now- well, while the two forecast values at 35 min before the
cast time “now” and the time when the peak discharge peak occurrence represent two spikes in the rainfall
occurred, as estimated using radar-observed precipita- hyetograph, and when interpolation was solely based
tion. The error in estimating the peak discharge 70 min on these two points the forecast rainfall volume was
before the peak occurrence is about 65%. The error in much larger than the estimates. The spatial variability
runoff volume is smaller than the error in peak dis- also played a role in this result. We actually performed
charge, but the shapes of the two plots look similar. The careful inspection of precipitation values on every radar

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

FEBRUARY 2006 SHARIF ET AL. 199

FIG. 6. Errors in the forecasted runoff as a function of the

forecast lead time for the 5 Jul event (based on ANC precipitation
nowcasts); only 60- and 30-min nowcast precipitation is used, val-
ues for intermediate 5-min intervals are interpolated.

pixel for each time step to confirm this fact. This prob-
lem is similar to the problem resulting from coarse tem-
poral resolution of radar precipitation estimates (e.g.,
Wilson and Brandes 1979).
In a second experiment, for the same event on 5 July,
we use radar information up to the forecast time 30-
and 60-min nowcast, while intermediate periods be- FIG. 7. Same as in Fig. 6, but previous precipitation nowcast
values are also used to fill intermediate 5-min values; error in peak
tween “now” and the 30- and 60-min nowcast are filled timing is shown in the middle panel (see text for details).
with 30- and 60-min nowcast values from previous pe-
riods. No previous nowcast is discarded unless observed
radar data become available to replace it. Use of pre- slightly as the lead time becomes smaller and the de-
vious nowcast values, as opposed to interpolation, con- crease becomes sharper 15 min before the peak occur-
siderably improves the hydrologic forecasts. This high- rence.
lights the importance of temporal resolution of precipi- Results of simulations using nowcast fields based
tation for this type of event. As seen in Fig. 7, the error only on TITAN extrapolation for the 5 July event,
in estimating the peak discharge, the most important shown in Fig. 8, are quite different. The errors in peak
hydrologic variable in flash flood nowcasting, 70 min discharge and runoff volume from the flood hydro-
before the peak occurrence is only about 17%. The graphs are about 4 times as large compared to errors
error in peak timing is around 10 min at that time, while from the ANC fields. Errors in peak timing are twice as
the error in runoff volume is about 25%. For such a large compared to the ANC results with a peak at 35
small catchment the concentration time is very short min before the peak occurrence. The extrapolation er-
and the performance of ANC would likely be better for rors drop to values comparable to the ANC maximum
a larger watershed where the lag time between the rain- errors only 10 min before the peak occurrence. The
fall peak and the peak discharge is much longer. The results highlight the benefits of nowcasting storm ini-
errors in peak discharge and runoff volume decrease tiation, growth, and dissipation provided by the ANC.

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

200 JOURNAL OF HYDROMETEOROLOGY VOLUME 7

FIG. 9. Same as in Fig. 7, but results are for the 8 Jul event.
FIG. 8. Same as in Fig. 7, but precipitation nowcasts are
computed using extrapolation only.
longer time. This extended the time over which the
hydrologic forecasts were useful.
On 8 July the quality of ANC nowcasts is poorer,
compared to the 5 July event, because of factors de-
7. Summary and conclusions
scribed in section 5. The hydrologic nowcasts are very
poor at the beginning of the ANC runs, as shown in Fig. A preliminary attempt to simulate runoff nowcasts in
9, but the errors drop very sharply for lead times less a highly urbanized small catchment is presented to
than 80 min. Interestingly, the errors in peak timing are demonstrate the utility of advanced nowcasting tech-
comparable to the 5 July errors. The timing error curve niques in urban hydrology. A physically based distrib-
shows a sharp decrease and then a sharp increase 80 uted-parameter hydrologic model is used to simulate
min prior to the peak discharge, which is an indication runoff generation driven by precipitation nowcasts. The
of the complex relationship between hydrographs pro- hydrologic model was first validated on the watershed
duced using different rainfall fields. As in Figs. 7 and 8, using radar-rainfall estimates. Based on the limited ob-
the shapes of the peak discharge and runoff volume servations available, validation indicates that there is
errors are similar. The errors in Fig. 9 are smaller than good agreement in simulation of the rising limb of the
the 5 July case for lead times shorter than 55 min due hydrograph with a difference of 9% between simulated
the difference in the watershed response to the storms. peak discharge and peak discharge estimate based on
The watershed response to the 8 July storm took a high water marks.

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

FEBRUARY 2006 SHARIF ET AL. 201

Results using the full capabilities of the ANC storm use of probabilistic precipitation nowcasts, and conse-
forecasting system clearly demonstrate that advanced quently probabilistic hydrologic forecasts, is a research
nowcasting can lead to significant improvements in thrust that seems to be promising.
flood warning and forecasting in urban watersheds,
even for short-lived events on small catchments. For Acknowledgments. The authors acknowledge the as-
the case study event typical of the performance of the sistance of Dr. Fred Ogden of the University of Con-
ANC when needed data are available, at lead times of necticut in providing the hydrologic model GSSHA and
about 70 min before the occurrence of peak discharge, assisting in the simulation setup. The constructive com-
forecast accuracies of approximately 17% in peak dis- ments of David Gochis of the National Center for At-
charge, 10 min in peak timing, and 25% in hydrograph mospheric Research are greatly appreciated. We ac-
volume were achieved for a 10 km2 highly urbanized knowledge the insightful comments by an anonymous
catchment. Significantly larger forecast errors resulted reviewer, which significantly improved the quality of
when only storm extrapolation was used to produce the paper.
precipitation forecasts. The use of nowcast fields made
at earlier time periods to fill values between the present REFERENCES
time observations and 30- and 60-min nowcasts greatly
Battan, L. J., 1973: Radar Observation of the Atmosphere. Uni-
improved the hydrologic forecasts. Even for the 8 July versity of Chicago Press, 324 pp.
case, where the ANC nowcasts were not good because Bedient, P. B., A. Holder, J. A. Benavides, and B. E. Vieux, 2003:
of a lack of meteorological information, hydrologic pre- Radar-based flood warning system applied to Tropical Storm
dictions driven by the ANC nowcasts were reasonably Allison. J. Hydrol. Eng., 8, 308–318.
accurate at short lead times. Belmans, C., J. G. Wesseling, and R. A. Feddes, 1983: Simulation
model of the water balance of a cropped soil: SWATRE. J.
Two facts make us believe the methodology used in
Hydrol., 63, 271–286.
this study can be feasible in real-time flood forecasting. Boldi, R. A., M. M. Wolfson, R. J. Johnson Jr., K. E. Theriault,
First, Harvard Gulch is a small catchment where the B. E. Forman, and C. A. Wilson, 2002: An automated, opera-
response to convective storms is very fast while for tional two hour convective weather forecast for the corridor
larger catchments the hydrologic response will be integrated weather. Preprints, 10th Conf. on Aviation, Range,
and Aerospace Meteorology, Portland, OR, Amer. Meteor.
slower, allowing for a greater lead time of reliable pre-
Soc., 116–119.
cipitation forecast; the case of 8 July where the re- Brooks, R. H., and A. T. Corey, 1964: Hydraulic properties of
sponse was slower is a good example. For large catch- porous media. Colorado State University Hydrology Paper 3,
ments even precipitation forecasts based on simple 27 pp.
storm translation may be very useful. Second, the com- Dixon, M., and G. Wiener, 1993: TITAN: Thunderstorm Identi-
plex terrain affects the performance of the ANC in the fication, tracking, Analysis and Nowcasting—A radar-based
methodology. J. Atmos. Oceanic Technol., 10, 785–797.
Denver area; Mueller et al. (2003) and Roberts and
Downer, C. W., and F. L. Ogden, 2004: GSSHA user’s manual:
Rutledge (2003) presented examples where the ANC Gridded surface subsurface hydrologic analysis, version 1.43
performed much better in other areas. for WMS 6.1. U.S. Army Corps of Engineers, Engineer Re-
More studies are required before implementing an search and Development Center Tech. Rep., in press.
urban flash flood nowcasting system based on the tools Fulton, R. A., 1999: Sensitivity of WSR-88D rainfall estimates to
the rain-rate threshold and rain gauge adjustment: A flash
described in this study. The hydrologic uncertainty as-
flood case study. Wea. Forecasting, 14, 604–624.
sociated with ANC outputs needs to be quantified ——, J. P. Breidenbach, D. J. Seo, D. A. Miller, and T. O’Bannon,
through Monte Carlo simulations or a similar method- 1998: The WSR-88D rainfall algorithm. Wea. Forecasting, 13,
ology. In many cases the ANC precipitation fields are 377–395.
reasonable in term of their magnitudes, but not so in Green, W. H., and G. A. Ampt, 1911: Studies on soil physics: 1.
term of their placement or orientation; the same is true Flow of air and water through soils. J. Agric. Sci., 4, 1–24.
Haverkamp, M. V., J. Vauclin, J. Touman, P. J. Wierenga, and G.
for any nowcasting technique. The effects of such errors
Vachaud, 1977: A comparison of numerical simulation mod-
need to by analyzed, especially for small watersheds. els for one-dimensional infiltration. Soil Sci. Soc. Amer. J., 41,
However, radar- and ANC-based flood warning and 285–294.
nowcast system have the potential to provide accurate Hutson, J. L., and A. Cass, 1987: A retentivity function for use in
and easy-to-understand information that can be avail- soil-water simulation models. J. Soil Sci., 38, 105–113.
able in real time using Geographic Information Systems Johnson, K., and Coauthors, 1998: Warning decision support sys-
tem: The next generation. Preprints, 14th Int. Conf. on Inter-
(GIS) and Internet technology. This information, to-
active Information and Processing Systems (IIPS) for Meteo-
gether with distributed hydrologic outputs from a rology, Oceanography, and Hydrology, Phoenix, AZ, Amer.
model similar to GSSHA, can be an invaluable re- Meteor. Soc., J25–J28.
source to decision makers during flooding events. The Kirkland, M.R., R. G. Hills, and P. J. Wierenga, 1992: Algorithms

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC

202 JOURNAL OF HYDROMETEOROLOGY VOLUME 7

for solving Richards’ equation for variably saturated soils. the Fort Collins, Colorado, flash flood of 1997. J. Hydrol.,
Water Resour. Res., 28, 2049–2058. 228, 82–100.
Konrad, C. P., and D. B. Booth, 2002: Hydrologic trends associ- Roberts, R. D., and S. Rutledge, 2003: Nowcasting storm initia-
ated with urban development for selected streams in western tion and growth using GOES-8 and WSR-88D data. Wea.
Washington. U.S. Geological Survey Water-Resources Inves- Forecasting, 18, 562–584.
tigations Rep. 02-4040, 40 pp. [Available online at http:// Sharif, H. O., F. L. Ogden, W. F. Krajewski, and M. Xue, 2002:
water.usgs.gov/pubs/wri/wri024040.] Numerical studies of radar-rainfall error propagation. Water
Lappala, E. G., R. W. Healy, and E. P. Weeks, 1987: Documen- Resour. Res., 38, 1140, doi:10.1029/2001WR000525.
tation of computer program VS2D to solve the equations of Smith, R. E., C. Corradini, and F. Melone, 1993: Modeling infil-
fluid flow in variably saturated porous media. U.S. Geologi- tration for multistorm runoff events. Water Resour. Res., 7,
cal Survey Water Resources Investigations Rep. 83-4099, 184 1219–1227.
pp. Soil Conservation Service, 1986: Urban hydrology for small wa-
Lee, J. G., and J. P. Heaney, 2003: Estimation of urban impervi- tersheds. UDSA Technical Release 55 (TR55), 164 pp.
ousness and its impact on storm water systems. J. Water Re- Sun, J., and N. A. Crook, 2001: Real-time low-level wind and
sour. Plann. Manage., 129, 419–426. temperature analysis using single WSR-88D data. Wea. Fore-
Merz, B., and E. J. Plate, 1997: An analysis of the effects of spatial casting, 16, 117–132.
Tholin, A. L., and C. J. Keifer, 1960: The hydrology of urban run-
variability of soil and soil moisture on runoff. Water Resour.
Res., 33, 2909–2922. off. Trans. ASEC, 125, 1308–1379.
Tilford, K. A., N. I. Fox, and C. G. Collier, 2002: A weather radar
Mueller, C., T. Saxen, R. Roberts, J. Wilson, T. Betancourt, S.
system for urban hydrology. Meteor. Appl., 9, 95–104.
Dettling, N. Oien, and J. Yee, 2003: NCAR Auto-Nowcast
Tuttle, J. D., and G. B. Foote, 1990: Determination of boundary
System. Wea. Forecasting, 18, 545–561.
layer airflow from a single Doppler radar. J. Atmos. Oceanic
National Research Council, 1996: Toward a New National
Technol., 7, 218–232.
Weather Service: Assessment of Hydrologic and Hydrometeo-
Wilson, J. W., and E. A. Brandes, 1979: Radar measurement of
rological Operations and Services. National Academy Press, rainfall—A summary. Bull. Amer. Meteor. Soc., 60, 1048–
51 pp. 1058.
Ogden, F. L., 2000: CASC2D reference manual, version 2.0. De- ——, N. A. Crook, C. K. Mueller, J. Sun, and M. Dixon, 1998:
partment of Civil and Environmental Engineering U-37, Uni- Nowcasting thunderstorms: A status report. Bull. Amer. Me-
versity of Connecticut, 83 pp. teor. Soc., 79, 2079–2099.
——, and B. Saghafian, 1997: Green and Ampt infiltration with Woolhiser, D. A., 1996: Search for physically based runoff
redistribution. J. Irrig. Drain. Eng., 123, 386–393. model—A hydrologic El Dorado? ASCE J. Hydraul. Eng.,
——, and D. R. Dawdy, 2003: Peak discharge scaling in a small 122, 122–129.
Hortonian watershed. J. Hydrol. Eng., 8, 64–73. Zhang, Y., and J. A. Smith, 2003: Space–time variability of rainfall
——, H. O. Sharif, S. U. S. Senarath, J. A. Smith, M. L. Baeck, and extreme flood response in the Menomonee River basin,
and J. R. Richardson, 2000: Hydrometeorological analysis of Wisconsin. J. Hydrometeor., 4, 506–517.

Unauthenticated | Downloaded 01/13/24 09:12 AM UTC