NWP and Radar Extrapolation: Comparisons and Explanation of Errors

JAMES WILSON,a DAN MEGENHARDT,a AND JAMES PINTOa

aNational Center for Atmospheric Research, Boulder, Colorado

(Manuscript received 13 July 2020, in final form 28 September 2020)

ABSTRACT: This paper examines nowcasts of precipitation from the High-Resolution Rapid Refresh (HRRRv2) model

from the summer of 2017 along the Colorado Front Range. It was found that model nowcasts (2 h or less) of precipitation

amount were less skillful than extrapolation of the KFTGWSR-88-D data at a spatial scale of 120 km. It was also found that

local-scale (mesoscale) influences on rainfall intensity and amount have a much greater impact on rainfall intensity than

large-scale (synoptic) influences. Thus, large-scale trends are not useful for modifying extrapolation nowcasts on the local

scale. Errors in the HRRRnowcasts are attributed to an inability of the model and data assimilation to resolve convergence

along outflow boundaries and other terrain-influenced mesogamma-scale flows that contribute to storm formation and

evolution. While the HRRRv2 1-h nowcasts were strongly correlated with observed precipitation events, the nowcast

precipitation amounts were in error by more than a factor of 2 about 50% of the time, with half of the cases being over-

estimates and half being underestimates. A large fraction of theHRRRv2 overestimateswere associatedwith stratiform rain

events. It is speculated that this was a result of misinterpretation of the radar bright band as more intense precipitation aloft

by the data assimilation scheme. A large fraction of the HRRRv2 underestimates occurred when the data assimilation and

model were unable to fully resolve the low-level convergence along small-scale, narrow boundaries that led to new storm

initiation and/or storm growth.

KEYWORDS: Convective storms; Radars/Radar observations; Nowcasting; Numerical weather prediction/forecasting

1. Introduction

This paper is motivated by the desire to improve 0–2-h

nowcasts of rainfall amount to support flash flood forecast-

ing, agriculture, recreation and other sectors of the economy.

There are several methods for nowcasting precipitation

amounts with the two most common being extrapolation and

numerical weather prediction. Extrapolation of radar re-

flectivity coupled with using a reflectivity–rainfall-rate rela-

tionship or polarimetric parameter–rainfall-rate relationship

generally has greater accuracy than model nowcasts of rainfall

amounts for nowcasts of 2 h or less. This paper quantifies the

relative skill of radar extrapolation and numerical weather

prediction (NWP) for nowcasting precipitation amount in the

Front Range of Colorado and assesses the potential reasons

why radar extrapolation forecasts tend to be more skillful

than present-day operational NWP with advanced data

assimilation.

Weather radar is often used to anticipate precisely where

and when precipitation will occur. This might be for life-

threatening situations or simply for keeping dry while run-

ning errands. The typical radar method used for anticipating

where the precipitation will be in the future was originally

proposed as early as 1953 (Ligda 1953). This method consists of

moving radar precipitation echoes forward in time using their

present intensity and recent past motion. This procedure is

referred to as Lagrangian extrapolation or here simply as ex-

trapolation. Bellon and Austin (1978) were likely the first to

produce precipitation nowcasts based on this procedure. Their

technique digitally matched two radar scans 1 h apart to obtain

the movement. The accuracy of radar extrapolation now-

casts decreases rapidly with time particularly for smaller-

scale convective precipitation. This paper does not review

the long history of radar extrapolation nowcasting tech-

niques; the interested reader can find appropriate refer-

ences in Sun et al. (2014) and WMO (2017) as well as Li et al.

(1995), Wilson et al. (1998), Germann and Zawadzki (2002, 2004),

Bowler et al. (2006), and Mandapaka et al. (2012) for examples.

Attempts to improve extrapolation nowcasts have been

largely unsuccessful. Tsonis and Austin (1981) investigated

the use of time trends in echo size and intensity to improve

extrapolation nowcasts of cells that already lived at least

30min, but they found negligible improvement in skill even in

elaborate nonlinear time-trending schemes. Wilson et al.

(1998) also tested similar trending techniques along the

Rocky Mountain Front Range region in Colorado and found

similar results. Tsonis and Austin (1981) concluded that es-

sential physical processes that dictate the change in rainfall

with time are not necessarily observable in the history of a

particular echo development. In the case of convective

storms, these physical processes are often events occurring in

the boundary layer such as wind convergence (Garstang and

Cooper 1981). Nowcasting systems that include human input

(expert systems) have shown improved skill over extrapola-

tion techniques by predicting storm initiation, growth and

decay based on knowing the location of boundary layer

convergence lines (Mueller et al. 1993; Wilson et al. 2004;

Roberts et al. 2012; Sharif et al. 2006). Satellite-based

methods that incorporate model analyses of stability have

also shown some promise in nowcasting convection initiation

under some conditions particularly when the satellite view is

unobstructed by high clouds (Iskenderian et al. 2010). Huang

et al. (2012) combined multiple NWPmodels and observationsCorresponding author: James W. Wilson, jwilson@ucar.edu;

James Pinto, Pinto@ucar.edu

DECEMBER 2020 W I L SON ET AL . 4783

DOI: 10.1175/MWR-D-20-0221.1

� 2020 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS CopyrightPolicy (www.ametsoc.org/PUBSReuseLicenses).

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

to improve nowcasts of temperature, relative humidity, wind

speed and wind gusts.

As summarized in Sun et al. (2014) and WMO (2017), ex-

trapolation of radar echoes tends to be more skillful than NWP

in predicting precipitation intensity up to 2 h out. Inmost cases,

NWP skill exceeds that of extrapolation after 2 or 3 h if

radar data (reflectivity, radial velocities) are assimilated.

Predictability is dependent on scale and forcing (Wilson et al.

1998; Germann et al. 2006; Kiel et al. 2014) with larger-scale

strongly forced regions having increased rainfall predictability.

NWP handles synoptic-scale forcing features, like frontal sys-

tems, better than local-scale features. On the other hand, radar

is particularly effective in observing small-scale features like

gust fronts, which can initiate new storms or intensify existing

storms. Early nowcasting techniques to combine NWP and

extrapolation consisted of applying lead-time-varying weights

to the extrapolation and NWP fields. At the shorter nowcast

time periods (1–2 h) full weight is given to extrapolation. That

weight is then gradually decreased with time while at the same

time increasing the weight given to NWP resulting in full

weight to NWPby 4–8 h (Browning 1980; Doswell 1986; Austin

et al. 1987; Golding 1998; Bowler et al. 2006; Pinto et al. 2010;

Seed et al. 2013). More advanced techniques for combining

model- and extrapolation-based forecasts have been devel-

oped that use probabilistic information from the model to

blend model data with extrapolation starting as early as 1 h

(Pinto et al. 2018). In addition, machine learning techniques

have also been used to nowcast radar imagery (Han et al. 2017;

Foresti et al. 2019; Franch et al. 2020).

As models and data assimilation techniques continue to

improve, it is expected that short-term forecasts and eventually

nowcasts will match or exceed that of extrapolation at all but

the shortest lead times (Sun et al. 2014). These increases in skill

afforded by advanced data assimilation (DA) techniques is a

critical aspect of the NOAA’s Warn-on-Forecast system

(Stensrud et al. 2009). A decade later, this area of intense re-

search to assimilated submesoscale observations is still very

active (Wheatly et al. 2015) but is moving toward ensemble

assimilation techniques, which better captures uncertainties in

the assimilation process. A full dissertation of the recent state

of mesoscale data assimilation at operational and research

centers around the world is given by Gustafsson et al. (2018).

Progress continues to be made in the assimilation of radar

observations using variational, ensemble, and hybrid tech-

niques (Duda et al. 2019). The assimilation of radar reflectivity

either via a latent heat nudging (e.g., Benjamin et al. 2016) or

directly relating reflectivity to the cloud hydrometeor proper-

ties of precipitating clouds (e.g., Dowell et al. 2011), has been

shown to improve the accuracy of analyses and subsequent

model nowcasts. A number of difficulties of assimilating radar

data have been discussed. For example, Janjic et al. (2018) have

shown that a mismatch between the scales of motion and

thermodynamic variability represented in the radar observa-

tions and that which can be resolved for a given model grid

can lead to significant errors. Also of particular relevance are

the uncertainties associated with nonlinear moist physical

processes and unbalanced flows that are characteristic of

convective-scale processes (e.g., Pagé et al. 2007).

This paper (i) examines the skill of an operational

convection-permitting NWP model with radar DA in now-

casting rainfall, (ii) assesses whether trend information avail-

able from the model can be used to improve radar-based

nowcasts obtained via extrapolation, and (iii) examines error

sources in 1-h nowcasts by NWP. Section 2 describes the

datasets used in this study and the experimental methods.

Section 3 provides a relative comparison of the model and

extrapolation performance and an assessment of whether or

not trends obtained from the model can be used to improve

extrapolation nowcasts. Section 4 examines the processes that

contribute to errors in the model 1-h nowcast of rainfall at a

scale of 120 km.

2. Experiment procedure

The experiment was conducted along the Rocky Mountain

Front Range for the summer (June, July, and August) of 2017.

The precipitation during this time period was dominated by

afternoon and evening thunderstorms that were largely initi-

ated by mesoscale boundaries and small-scale variations in

stability rather than via synoptic-scale forcing. Terrain also

played a significant role on the evolution of thunderstorms that

were observed; see Fig. 1 for map of the terrain. Thus, the re-

sults obtained in this study may not be transferable to other

regions of the United States, and especially not to areas where

synoptic-scale forcing may dominate.

a. Experimental area and time period

Figure 1 shows the Rocky Mountain Front Range. The

red-outlined box defines the region (108 000 km2) where

the large-scale rainfall intensity trend was obtained from

HRRRv2 and compared with the trend from a much smaller

FIG. 1. Region of the experiment. The yellow circle shows the 60-

km-radius circle about the KFTG NEXRAD radar, which is the

rainfall nowcast area. The red-outlined rectangle shows the area

over which themodel large-scale trend is obtained. The small black

polygon is the East Denver Rain Gauge Network. The background

is the elevation of the topography, scale on the right in thousands of

feet (1000 ft 5 ;300m).

4784 MONTHLY WEATHER REV IEW VOLUME 148

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

area (11 304 km2). This smaller area is defined by a 60-km-ra-

dius circle centered on the KFTG WSR-88-D where the

HRRRv2 and radar extrapolation rainfall nowcasts were made

and the radar rainfall estimates were made. The experiment

covered all hours during June, July, and August of 2017. Of the

possible 2208 h over the 3-month period, 2140 were used. The

remaining 68 h were times for which radar data were missing.

b. Rainfall nowcast area

The 60-km-radius circle about the Denver WSR-88D

(KFTG) was chosen on the basis of the desire to reduce to-

pography influences on precipitation amount by keeping the

region on the plains. At the same time, the quality of polari-

metric parameters and rainfall estimation is greater at these

closer radar ranges. The quality of the rainfall estimates de-

creases with range because of radar beam broadening and in-

creased beam height with increasing range.

c. Rainfall estimation

Rainfall estimates were derived fromKFTG shown in Fig. 1.

The KFTG radar is part of the national network ofWSR-88Ds,

which are polarimetric S-band radars with a 18 beamwidth.

These radars are considered state-of-the-art for estimating

rainfall. The radar rainfall estimates were computed using

the National Center for Atmospheric Research (NCAR)

Environmental Operations Laboratory (EOL) ‘‘HYBRID’’

algorithm (Dixon et al. 2015). Using the polarimetric radar

parameters, the algorithm uses a fuzzy logic hydrometeor

classification technique that selects a correspondingly appro-

priate precipitation rate relationship for each hydrometeor

type. In addition to horizontal radar reflectivity, the polari-

metric radar parameters (differential reflectivity and specific

differential phase) are used in estimating rainfall rate. In

addition, a relationship based on specific differential phase is

used in the particle identification algorithm to identify hail.

The rainfall-rate algorithm also selects the best elevation angle

in the case where radar beam blocking is a possibility. Figure 2

shows a radar versus rain gauge comparison of the 1-h rainfall

estimates for the 3 months from the East Denver Rain Gauge

Network (see Fig. 1). This group of rain gauges covers a 230-

km2 area. The area was chosen because it had a particularly

dense uniform distribution of rain gauges. Editing of the rain

gauge observations was performed to remove spurious gauge

tips that were not associated with rain. The coefficient of cor-

relation between the radar and gauge estimates is 0.89. Based

on close examination of the two estimates it is not clear if one

technique is more accurate than the other. At times, it was noted

that the scale of the rainfall core was sufficiently small that it

occurred between gauges, and thus, evenwith a dense network, a

heavy rain area might not be detected. The 3-month total of

gauge measured area-averaged rainfall over the area covered by

the network of rain gauges was only 62mm.1 This was far below

the average Denver rainfall for the 3 months of 147mm.

The observed rainfall exhibited a strong diurnal pattern

resulting from afternoon and evening thunderstorms (Fig. 3)

with a peak occurring between 1600 and 1700 LT. Figure 3

shows the diurnal cycle of total rainfall accumulation obtained

with HRRRv2, extrapolation and radar observed. It is inter-

esting to note that the model tends to overestimate the rain

during the night andmorning and underestimate in the afternoon,

which is the most convective period. Reasons for this are ad-

vanced in section 3. Not surprisingly, radar extrapolation and ra-

dar observed agree closely except for the 1-h delay in nowcast

amounts during the period of storm initiation and growth.

FIG. 2. Comparison of rain gauge and radar estimates of the

average 1-h rainfall amount for the East Denver Rain Gauge

Network. All hourly estimates for the 3 months of the experiment

are shown.

FIG. 3. Total radar estimated rainfall, by hour of the day, for

June, July, and August 2017 for the area 60 km in radius about the

KFTG NEXRAD. The rainfall is the area average in millimeters.

Green is the radar estimate, cyan is the radar extrapolation 1-h

nowcast, and red is the HRRRv2 1-h nowcast.

1 The HRRRv2 1-h nowcast was 66mm for the same area

and time.

DECEMBER 2020 W I L SON ET AL . 4785

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

d. NWP model

The NWP dataset used here is from version 2 of the High

Resolution Rapid Refresh (HRRRv2). As described in

Benjamin et al. (2016) and C. R. Alexander et al. (2020, per-

sonal communication) ‘‘the HRRR is a NOAA real-time 3-km

resolution, hourly updated, convection-allowing atmospheric

model run using a horizontal grid spacing of 3 km. Radar data

are assimilated into the HRRR model every 15min over a 1-h

period.’’ The 3-km HRRR forecasts are forced at the lateral

boundaries by the 13-km Rapid Refresh (RAP) model, which

is a continental-scale hourly updated assimilation/modeling

system run operationally at NCEP. The HRRRmodel and the

data assimilation system continue to be developed, with

HRRRv4 becoming operational at NCEP in late 2020. It is

noted that the data assimilation to be included with the

HRRRv4 upgrade will be a hybrid variational ensemble

scheme that will use 36 members run at convection-permitting

resolution to assimilate radar reflectivity and radial winds

which could help to improve initial conditions and ultimately

nowcasts of precipitation intensity in the years to come.

e. Precipitation extrapolation

The rainfall-rate field estimated from reflectivity was ex-

trapolated with motion vectors obtained using Thunderstorm

Identification Tracking and Nowcasting (TITAN) software

developed by Dixon and Wiener (1993). Note that this now-

casting system does not attempt to account for initiation,

growth or dissipation of precipitation areas. The field of motion

vectors was obtained in a six-step process. First, TITANwas used

to identify radar reflectivity objects with reflectivities .35 dBZ.

Second, TITAN determined the motion of these objects by

observing their past locations. Third, zonal and meridional

components (u, y) of the motion vector for each storm object

was assigned to a 1-km grid. That is, each grid point within the

storm object has the same u, ymotion values. Fourth, the u and

y values for each object were extended 40 km from the centroid

location of each 35 dBZ precipitation object. When grid points

contain more than one set of u, y values, distance weighted

averaging was applied. Fifth, grid points with no values were

assigned the 700-hPa wind from the HRRRv2. Sixth, temporal

smoothing was applied that limits how much a vector can

change in speed and direction from one time to the next.

Figure 4 is an example of the above TITAN-based field of

motion vectors superimposed on the rainfall-rate field.

The hourly estimated rainfall accumulation field was ob-

tained by extrapolating in 1-min time steps the above rainfall-

rate field with the TITAN-based extrapolation vectors. The

individual 1-min rainfall-rate values at each 1-km grid point

were then summed to obtain a field of 1-h rainfall amounts at

each grid point. Then the grid points within the 60-km-radius

circle were summed (11 305 points) and the area-average 1-h

rainfall accumulation was obtained.

3. HRRRv2: Extrapolation comparisons, blending, anderrors

The rapid assimilation of radar data and the rapid cycling

(15min) by the HRRRv2 greatly reduces the model spin up

period, which was often associated with low skill by NWP

during the first few hours. In addition, radar data assimilation

provides the HRRRv2 the distribution of precipitation at

forecast time. Improved nowcasts beyond a few tens ofminutes

requires predicting storm initiation, growth and decay, which

are dependent on small-scale wind, humidity and temperature

variations in the boundary layer (e.g., Weckwerth 2000; Ge

et al. 2013). Current observational networks, primarily de-

signed for synoptic-scale forecasting, generally cannot provide

the environmental conditions with the temporal and spatial

resolution required for nowcasting. However, convergence

lines are often evident in the radar reflectivity field (Wilson

et al. 1994). These convergence lines (boundaries) often ap-

pear as thin lines in the ‘‘clear air’’ radar reflectivity field.

However, as will be discussed later, the convergence occurs

along narrow zones along these convergence lines that are

often not as well observed in the Doppler velocity field and not

well resolved by the 3-km grid spacing of HRRRv2. Thus,

NWP data assimilation of Doppler velocity can at best only

coarsely represent these boundaries. In addition, there pres-

ently is no method for assimilating the location of convergence

boundaries that are indicated by radar reflectivity thin lines.

The importance of boundaries with respect to HRRRv2 rain-

fall underestimation is examined in section 4.

In the following four subsections of section 3 the HRRRv2

hourly rainfall nowcasts for the 60-km area will be 1) compared

with extrapolation, 2) investigated as a means to improve radar

extrapolation nowcasts, 3) compared with radar observations,

and 4) examined for skill in nowcasting the start of a precipi-

tation event and the skill in nowcasting the hourly rainfall

amounts. Section 4 examines issues that cause errors.

a. Comparison of model and extrapolation nowcasting skill

Figure 5 compares the skill of radar extrapolation and

HRRRv2 as a function of nowcast time. The nowcast is for the

FIG. 4. Example of vectors derived from the extrapolation of the

radar reflectivity field as derived from TITAN. These vectors are

superimposed on the rainfall-rate field derived from the EOL

HYBRID rainfall estimation algorithm. The rainfall rates are

shown in the color bar on the right. The 60-km-radius circle about

KFTG is shown by the yellow circle.

4786 MONTHLY WEATHER REV IEW VOLUME 148

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

hourly average rain accumulation over the 60-km-radius circle

of KFTG. The extrapolation skill was higher at 1 and 2 h with a

cross over point where the HRRRv2 becomes superior just

over 2 h. A similar comparison was made for the much smaller

East Denver rain gauge network with almost the same results;

however, the number of cases is relatively small.

Several studies have examined the crossover point between

the skill of NWP and extrapolation. Pinto et al. (2010) and Sun

et al. (2014) used the HRRR without radar data assimilation

and found the cross over point near 4 h. Later studies with

HRRR data from 2012 and 2013 showed increased skill with

the HRRR using radar data assimilation (Pinto et al. 2015).

Hwang et al. (2015) using 2013HRRR data with radar data

assimilation to forecast the maximum altitude of 18-dBZ echo

top heights within a 20-km-radius circle found the crossover

points between 2 and 3 h. As mentioned above the HRRRv2

data used in our study were from 2017 where radar data as-

similation included both radar reflectivity and Doppler veloc-

ity and the crossover point was just over 2 h.

b. Blending HRRRv2 and extrapolation

Radar-based 0–2-h extrapolation nowcasts often have skill

in nowcasting flash floods, particularly in the first hour (Vivoni

et al. 2006). These nowcasts make use of extrapolating radar

reflectivity and only a few techniques include nowcasting in-

tensity changes (Wilson et al. 1998). The premise being tested

here is whether predicted larger-scale trends obtained from the

HRRRv2 could be used to improve 1- and 2-h radar extrapo-

lation nowcasts on the local scale to support nowcasting flash

floods and the hydrology of urban areas.

An area 300 km 3 360 km (108 000 km2) was chosen for

obtaining large-scale rainfall intensity changes. This area is

referred to locally as the ANC (AutoNowcaster) domain and is

an area for which we traditionally conduct convective storm

studies. The ANC and 60-km-circle domains are shown in Fig. 1.

The procedure was to obtain the change in the area average

rainfall amount from one hour to the next for 1) the model

nowcast for the ANC domain and 2) the radar observed for the

60-km circle. Correlating the changes/trends for these two

different size areas for all hours that any rain occurred in the

ANC domain for the 3-month period resulted in a correlation

coefficient of 20.003. By restricting the correlation to only

occasions when the average hourly rainfall accumulation,

within the 60-km-radius circle, was at least 0.5mm (top 12% of

the 412 cases) the correlation was only 0.28. From these very

low correlations it is apparent that small-scale (mesoscale)

influences on rainfall intensity dominate over large-scale

(synoptic) influences along the Front Range of Colorado.

In addition, a similar comparison was made for the 60-km

circle where the radar and model change/trend was compared

for the 60-km circle. The correlation including all hours was

0.47; restricting the correlation to observed rain amounts

of.0,.0.01,.0.05,.0.10, and.0.25mm gave correlations of

0.32, 0.48, 0.50, 0.52, and 0.51, respectively. While these rainfall

values seem very low, they are averages for an area of

11 304 km2, where much of the area often receives no rain but

amounts can be high over a small portion of the area. These low

correlations together with the above results indicate that there

is no scale at which model trends can be used to improve 1-h

extrapolation nowcasts. This will be reinforced further in a

discussion below of the HRRRv2 ability to nowcast the be-

ginning of a rain event.

Since there is, a very large diurnal cycle in rainfall (Fig. 3) a

test was made to determine if the 1-h extrapolation nowcasts

could be improved by just increasing/decreasing the amount of

rainfall based on trends evident in the diurnal composite of

precipitation accumulation. There is a large increase in rainfall

amount from 1400 to 1600 mountain standard time (MST) and

subsequent decrease from 1700 to 1900MST.Using the rates of

increase from Fig. 3 the rainfall for hour 1400–1500 MST was

multiplied by 2.98, 1500–1600 MST by 2.2, 1700–1800 MST by

0.58, and 1800–1900 MST by 0.76. However, this simple cli-

matological method to modify the 1-h extrapolation nowcasts

did not improve the extrapolation nowcasts (not shown). The

various results all indicate that the local-scale meteorological

factors contribute more to rainfall-rate changes from hour to

hour than any large-scale or diurnal influences.

c. HRRRv2 compared with radar observation

Figure 6 compares the HRRRv2 1-h nowcasts of hourly rain

accumulation for the 60-km circle with that observed by the

radar. The diagonal lines are the factor-of-2 ratio (0.5 and 2.0)

betweenmodel nowcast and radar observed. The correlation of

HRRRv2 nowcasts with the radar observed—that is, all points

in Fig. 6 including those with no rain—is 0.75. This correlation

decreases to 0.61 for the 2-h nowcast (plot not shown). For the

1-h nowcast, excluding the hours where both techniques have

no rain, the correlation is 0.70. The correlation decreases as the

observed accumulation threshold is increased; for example, the

correlations for thresholds of 0.01, 0.05, 0.10, and 0.25mm are

0.67, 0.62, 0.58, and 0.53, respectively. The average ratio of

HRRRv2 divided by radar observed is 1.01 for the 1-h nowcast

and 1.19 for the 2-h nowcast, indicating a slight tendency to

predict more precipitation than was observed.

For the entire summer, there are 1627 h with no rain from

either technique of 2140 nowcast hours. Thus, 515 h remain

FIG. 5. Comparison of the skill (correlation coefficient) of the

HRRRv2 (blue) and radar extrapolation (green) as a function of

nowcast length.

DECEMBER 2020 W I L SON ET AL . 4787

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

where at least one of the techniques is nonzero. Table 1 is a

contingency table comparing all the HRRRv2 1-h nowcasts for

rain with those observed by the KFTG radar. The occurrence of

rain is defined as an average rainfall amount of$0.01mm for the

60-km-radius circle of KFTG. From Table 1 the HRRRv2

1-h nowcasts for rain/no rain were correct 91% of the time.

Considering only the hours for which at least onemethod had rain,

62%of thenowcasts for rainwere correct. The above indicates that

theHRRRoften nowcasts the presence of rain; however, aswill be

shown later, it has difficulty nowcasting the amount.

d. Nowcasting precipitation events

The question arises as to how well the HRRRv2 nowcasts

the start of a precipitation event, as well as the hourly rainfall

amounts during the precipitation event. A precipitation event

was defined as a period at least 3 h long with at least 1 h having

an area-average rain accumulation of 0.10mm within the 60-

km-radius circle. There were 41 such events observed. The

start of an event was classified into three categories: translation

of rain into the 60-km circle (11 events), rain initiation within

the circle (25 events), and a combination of translation and rain

initiation (5 events). The classification was based on examining

the time series of radar reflectivity, HRRRv2 and extrapola-

tion using the ‘‘CIDD’’ display ((https://ral.ucar.edu/CIDD/

user_manual/cidd_users.html). For the sake of simplicity, we

will call rain initiating within the 60-km circle CI (convection

initiation) since convective rain was always the case. There

was a wide variety of skill in the HRRRv2 nowcasts of the start

time of rain and the hourly rainfall amounts after the start time.

For CI events the start time was correctly nowcast by themodel

for 7 of the 25 events, for translation events it was correct 7 of

11 times, and for a combination CI and translation it was cor-

rect for 3 of 5 events.

Figure 7 provides six examples of the performance of the

HRRRv2 and radar extrapolation in the prediction of 1-h rain

amounts within the 60-km circle. The figure shows, for six

cases, a sequence of 1-h nowcasts by both HRRRv2 and radar

extrapolation. Also shown is the verifying radar estimated 1-h

rainfall amounts. Figures 7a and 7b are contrasting examples of

the HRRRv2 skill in nowcasting the start time and amount of

rainfall for two CI events. In Fig. 7a the model essentially

misses an observed CI event with a delayed start time for

nowcasts issued at 1400, 1500, and 1600 MST and greatly un-

derestimated amounts in subsequent 1-h nowcasts. This bias in

the HRRRv2 nowcasts occurred for several of the observed CI

events; however, as Fig. 7b shows, just the opposite can be the

case. Figure 7b shows the model significantly overpredicts the

hourly amounts. In this event, the observed convective storms

are moving from the northwest. Figure 7c is a particularly in-

teresting CI case; the model, as is typical, has delayed CI and

underestimates the rainfall amount until 0700 MST.2 Later in

the morning, the HRRRv2 appears to try to ‘‘catch up’’ but

then overestimates the rainfall amount and duration. Further

investigation into this case revealed that the HRRRv2 nowcast

overestimates coincided with the appearance of the radar

bright band (American Meteorological Society 2020) in the

radar reflectivity data within the stratiform rain area of the

convection. The linkage between nowcast overestimates and

the bright band, which was apparent in several of the cases, is

discussed further in section 4. While the radar bright band was

linked to overforecasting in many cases, the HRRRv2 also

overestimated the rainfall associated with stratiform events

whether or not there was a bright band present.

Figures 7d–f are translation events. Figures 7d and 7e are

examples of both over- and underestimating the rainfall. The

HRRRv2 precipitation nowcasts were delayed in both of these

cases. Later in the event shown in Fig. 7d, HRRRv2 nowcasts

greatly overestimate the amounts. While occasionally the

bright band was observed, it was not deemed to be the primary

reason for the overestimation. The rain amounts associated

with event shown in Fig. 7e were generally underestimated

except for at the end of the event. In this case there were waves

TABLE 1. Contingency table comparing the HRRRv2 1-h now-

casts of occurrence of area-average rainfall within a 60-km-radius

circle of KFTGwith that estimated by radar. A threshold of 0.1mm

is used to identify an event.

Radar

No Yes

HRRR No 1627 107

Yes 90 318

FIG. 6. Comparison of HRRRv2 and radar estimates of area

average rainfall for the 60-km-radius circle of KFTG. The

HRRRv2 are 1-h nowcasts. The black lines represent the factor-of-

2 ratio, 0.5 and 2.0, betweenHRRRnowcast and radar observation.

The red points are the hours that are examined in detail and are the

subject of the discussion in section 4. The orange point in the upper

right is examined in Fig. 8, below.

2 Interestingly both of the cases in Figs. 7b and 7c are nocturnal

events, which are unusual and likely driven by synoptic-scale

features.

4788 MONTHLY WEATHER REV IEW VOLUME 148

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

of intense storms moving from northwest to southeast across

the domain with a squall line near the end. A bright band in the

stratiform rain at the rear of the squall line coincided with the

HRRRv2 overestimate of rain at hour 2100 MST. Figure 7f

starts as a relative weak stratiform area with some imbedded

convection, the start of which is nicely nowcast by HRRRv2.

The big peak in the radar observed rain at hour 1600MST is the

result of a large area of CI associated with a convergence line.

This is a nice example of the HRRRv2 not initiating storms

associated with mesoscale convergence lines, which will dis-

cussed further in section 4.

In summary, the HRRRv2 had little skill in nowcasting the

start of a rain event that was the result of CI within the circle;

however, it had more skill when the rain translated into the

circle. There was a tendency for the rain amount to be

overestimated by the HRRRv2 in stratiform events and

underestimated during more intense periods of convective

precipitation, particularly when the storms initiated within the

60-km-radius verification circle. The presence of a radar bright

band often coincided with overestimates in the 1-h HRRRv2

nowcasts of rain amount. Also, the model did not anticipate CI

associated with mesogamma-scale convergence lines. These

last two items are discussed further below.

4. NWP 1-h nowcast errors

Figure 8 and the following figures provide representative

examples of the performance of 1-h rainfall accumulation

nowcasts obtained from HRRRv2. The case shown in Fig. 8

corresponds with the heaviest rainfall case, which is indicated

FIG. 7. Time lines of area average hourly rainfall amounts for the 60-km circle for 1-h nowcasts by HRRRv2 (red

line) and extrapolation (blue line) and the radar observed (green line) at verification time for the (a) CI event on 24

Aug, (b) CI event on 23 Jun, (c) CI event on 15 Jul, (d) translation event on 7 Aug, (e) translation event on 5 Aug,

and (f) translation event on 8 Aug. See the text for a discussion of the features.

DECEMBER 2020 W I L SON ET AL . 4789

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

by the large orange dot in Fig. 6 (far upper right). In this case,

the 1-h rain accumulation from theHRRRv2 (2.23mm) almost

perfectly matched the observed radar estimated accumulation

(2.28mm). The HRRRv2 nicely nowcasts the movement of the

two most intense storms observed in Fig. 8a. It is clear that the

HRRRv2 assimilation technique captures the initial spatial

extent and intensity of the rainfall at analysis time and thus,

produces a 1-h nowcast that captures the observed precipita-

tion pattern within the 60-km circle. The HRRRv2 shows a

third area of rainfall accumulation near the center of the

60-km-radius circle (Fig. 8c) that was not observed. This is

likely the result of incorrectly initiating a new storm within the

first hour of the simulation. Another notable aspect of the

HRRRv2 1-h nowcast is that the rainfall fields are much more

smoothly varying in space (Fig. 8c vs Fig. 8d) than observed. This

is because the HRRRv2 cannot resolve the smaller-scale more

intense updrafts and corresponding heavier precipitation.

Additional hours (68) examined in detail are shown in Fig. 6

as colored red dots. For 36 of the 68 h being examined in detail

theHRRRv2 rainfall nowcast was significantly greater than the

radar observed rain amount; these cases were grouped into

four categories. 1) Stratiform echo was dominant in the radar

observations and the model 1-h rainfall nowcast was for in-

tensification (19 cases). This included cases in which convective

echo was embedded in the stratiform echo. Typically, the

stratiform precipitation was dissipating and not intensifying.

FIG. 8. Examination of the orange point in Fig. 6, which has the highest rainfall amount with the highest HRRR

accuracy. It is the 1-h nowcast for the hour 2200–2300 UTC 10 Aug 2017. Shown are (a) radar reflectivity at 2200

UTC, (b) radar reflectivity at 2300 UTC (the echo near the tail of the orange arrow is a new storm), (c) HRRR

rainfall accumulation between 2200 and 2300 UTC, and (d) the radar estimate of rainfall accumulation between

2200 and 2300UTC. The white arrows represent themovement of the two strongest storms from 2200 to 2300UTC,

which are responsible for the heaviest rainfall tracks in (d). The short orange arrow in (a) and (b) is the movement

and quick dissipation of the weak storm in (a) indicated by the small orange circle in (a) and (b). The color bar at the

bottom of (a) and (b) is for radar reflectivity (dBZ). The color bar at the bottom of (c) and (d) is hour rainfall

amount (mm). The yellow circle is the 60-km radius around the KFTG radar.

4790 MONTHLY WEATHER REV IEW VOLUME 148

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

2) Short-lived, small, weak storms were observed and theHRRRv2

grouped the storms into more organized, intense lines or larger

areas (9 cases). 3) The HRRRv2 nowcast had considerable

error in storm location and/or storm motion was too slow

(5 cases). 4) The HRRRv2 significantly over forecast the

rainfall or nowcast storms that did not occur (3 cases). Any

single hour could have fit into more than one of these cate-

gories, but the most dominant situation was chosen when

placing an hour into a category.

For 31 of the 68 1-h rainfall nowcasts, the HRRR was sig-

nificantly less than the radar-estimated amount. These cases

occurred during conditions in which storm initiation and/or

growth was observed and are grouped into two categories: 1)

storm growth and/or initiation was associated with significant

increase in radar Doppler velocity convergence without dom-

inant reflectivity thin lines (16 cases) and 2) storm growth

and/or initiation was clearly associated with reflectivity thin

lines (15 cases).

As mentioned above, the most dominant situations leading

to HRRRv2 overnowcasting precipitation accumulation was

associated with the presence of stratiform precipitation and a

tendency for HRRRv2 to intensify the stratiform rain. This is

FIG. 9. Example from 15 Jul 2017 of an HRRR 1-h nowcast that intensifies stratiform rain. Shown are radar re-

flectivity at (a) 1400 and (b) 1500UTC, rainfall accumulation between 1400 and 1500UTCby (c) HRRRand (d) radar,

and brightband (arc of high reflectivity) (e) at an elevation angle of 88 at 1400 UTC and (f) at a radar elevation angle of

108 at 1500 UTC. The color scale for reflectivity (dBZ) is at the bottomof (a), (b), (e), and (f). The color scale for hourly

rainfall amount (mm) is at the bottom of (c) and (d).

DECEMBER 2020 W I L SON ET AL . 4791

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

demonstrated in Fig. 9 with an example from 15 July at 1400

UTC. A likely contributing cause is the treatment of the radar

bright band in the data assimilation. A radar bright band was

present in nearly all the cases in the stratiform overestimation

category. It is speculated that the HRRRv2 reflectivity di-

agnostic is interpreting the bright band as an area of more

intense rainfall aloft. In this case, the top of the bright band

was at about 3 km, near the freezing level. Figures 9e and 9f

nicely show the radar beam intersecting the bright band at

radar elevation angles of 88 at 1400UTC and 108 at 1500UTC,

respectively. All other radar elevation angles in the figures

are at 0.58.The second-highest number of cases in which HRRRv2

overestimates the rain is when it wrongly intensifying short-

lived, small, weak, convective storms. Figure 10 is such an

example from 1100 UTC 7 August. The cause for the erro-

neous formation of small, weak, short-lived storms in the

HRRRv2 nowcast is not known and is in need of further

exploration.

The two most dominant underforecasting situations were

associated with the HRRRv2 missing storm initiation/growth

associated with Doppler velocity-indicated convergence or

reflectivity thin lines. Figure 11 is an example of significant

storm intensification and initiation associated with a well-

defined area of increasing radar radial velocity convergence

where radar reflectivity thin lines are not dominant. This in-

crease in convergence can be seen in the Doppler velocity in

Figs. 11c and 11d as indicated by the area within the white

contour. Note that most of the convergence occurs over a

distance of only a few kilometers and thus would be under-

resolved and likely greatly smoothed by amodel with 3-km grid

spacing. Because of inadequate surface convergence, the

model only produces modest rainfall that was more than 50%

less than observed within the verification radius.

FIG. 10. Example from 7 Aug 2017 of HRRR 1-h nowcast that wrongly overintensifies and organizes short-lived

weak storms. Shown are radar reflectivity at (a) 1100 and (b) 1200 UTC, (c) the HRRR1-h nowcast of rainfall

accumulation between 1100 and 1200 UTC, and (d) the radar estimate of rainfall accumulation between 1100 and

1200 UTC. The color bar for reflectivity (dBZ) is at the bottom of (a) and (b). The color bar for rain accumulation

(mm) is at the bottom of (c) and (d). The yellow circle is the 60-km radius around KFTG.

4792 MONTHLY WEATHER REV IEW VOLUME 148

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

The use of radar reflectivity thin lines to detect boundary

layer convergence lines and their use in nowcasting convective

storms are well-documented (Wilson and Schreiber 1986;

Wilson andMueller 1993; Roberts et al. 2012). However, there

is no method to ingest or assimilate thin line information into a

numerical model. Figure 12 provides an example in which

convergence lines (reflectivity thin line) moving rapidly south

initiate new storms and enhances existing storms. Figure 13 is

an example of storm initiation and intensification associ-

ated with colliding convergence lines. While radar Doppler

velocities also show corresponding lines of convergence, the

radial velocities are often not as reliable as reflectivity for

detecting a thin line/boundary because of several factors. First

the thin line is the result of a greatly increased concentration of

insects that occurs within the updraft associated with the low-

level convergence. This concentration of insects results in a

FIG. 11. Example from 12Aug 2017 of the HRRR1-h nowcast underestimating the rainfall in a situation with the

development of strong convergence. Shown are radar reflectivity at (a) 2200 and (b) 2300UTC,Doppler velocity at

(c) 2200 and (d) 2230 UTC [the white outline in (c) and (d) shows the region of strong convergence], (e) the

HRRRv2 1-h nowcast of hourly rain accumulation between 2200 and 2300 UTC, and (f) the radar-observed ac-

cumulation between 2200 and 2300 UTC. The color bars are for reflectivity (dBZ), Doppler velocity (m s21), and

rain accumulation (mm), respectively.

DECEMBER 2020 W I L SON ET AL . 4793

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

vertical column of enhanced reflectivity that extends above the

low-level convergence (Wilson et al. 1994). Thus, the thin line

is not as sensitive to the radar beam overshooting the shallow

convergence. Second, much of the convergence may be below

the radar beam and not available for assimilation. Third, the

convergence line velocity component may be far from normal

to the radar, and thus the radar may only be able to observe a

small component of the total convergence. Fourth, the con-

vergence may be horizontally narrow, greatly reducing the

ability of the radar to resolve, particularly at longer ranges. In

addition theHRRRv2 and other numerical models are not able

to fully resolve the convergence line since the grid spacing is

on a scale similar to the convergence line.

5. Summary

Local-scale influences (mesoscale) on storm evolution were

far greater than large-scale (synoptic) influences. Thus using

FIG. 12. Example from 30 Jul 2017 of the HRRR 1-h nowcast underestimating the rainfall as an east–west

convergence line (white arrows) moves south, interacting with a north–south convergence line (yellow arrows).

Shown are radar reflectivity at (a) 2200 and (b) 2300 UTC, Doppler velocity at (c) 2200 and (d) 2300 UTC, and

rainfall accumulation between 2200 and 2300 UTC for the (e) HRRR 1-h nowcast and (f) radar observation. The

color bars are as in Fig. 11.

4794 MONTHLY WEATHER REV IEW VOLUME 148

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

large-scale time trends in HRRRv2 nowcast rainfall amounts is

not useful for improving radar extrapolation for 1- and 2-h

rainfall amount nowcasts. Note that results presented herein

are specific to the Front Range of Colorado and may not be

applicable in other climatic regions across the country, par-

ticularly in areas influenced by stronger synoptic-scale forcing.

Comparison of numerical model (HRRRv2) and radar ex-

trapolation nowcasts of rainfall amount for nowcast lead times

of 1, 2, and 3 h were conducted for the 60-km-radius circle

around the KFTG radar. It was found that, for this region, the

HRRRv2 is more skillful at predicting rain amounts than radar

extrapolation at lead times greater than 2 h for the 60-km-

radius circle. This lead time, when the skill of the model exceeds

that of extrapolation, is similar to that found by Hwang et al.

(2015) who were using an earlier version of the HRRR (2011

version) for aviation purposes and not for nowcasting rainfall.

FIG. 13. Example from 25 Aug 2017 of the HRRR 1-h nowcast underestimating the rainfall as two north–south

convergence lines collide. Shown are radar reflectivity at (a) 2300 and (b) 0000 UTC, Doppler velocity at (c) 2300

and (d) 0000 UTC, and rainfall accumulation between 2300 and 0000 UTC for the (e) HRRR 1 and (f) radar

observation. The white and orange arrows in (a) indicate the colliding convergence lines. The color bars are as

in Fig. 11.

DECEMBER 2020 W I L SON ET AL . 4795

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

Comparison of themodel 1-h nowcasts with the radar observed

amounts for the 3-month total rainfall showed no bias (i.e., the

ratio HRRRv2/radar rain amounts was 1.01 for 1-h nowcasts

and 1.19 for the 2-h nowcast. TheHRRRv2 nowcasts of rain/no

rain over the 60-km-radius circle were correct 91% of the time.

For hours when rainfall was observed (i.e., 24% of the time)

the correlation coefficient was 0.70.

Comparison of individual HRRRv2 1-h nowcasts with ob-

served 1-h rain amounts revealed that the HRRRv2 generally

reproduces the broadscale observed rainfall pattern in and

around the 60-km-radius verification circle, although the

modeled rainfall patterns tended to be much more smoothly

varying than observed rain amounts. The HRRRv2 was often

delayed in nowcasting the beginning of CI events within the 60-

km circle. That is, the precipitation had to be present in the

observations for the HRRRv2 assimilation to produce rainfall

nowcasts. There were many hours when the HRRRv2 hourly

rain amounts were too large or too small by more than a factor

of 2. Examination of these cases showed a variety of reasons for

the errors as summarized below.

The HRRRv2 nowcasts tended to overforecast the rainfall

amount in stratiform rain events. The HRRRv2 also had a

tendency to intensify and not dissipate stratiform rain areas.

These errors coincided with the presence of the radar bright

band within the stratiform rain regions. It is believed that these

areas of bright band were interpreted as areas of enhanced

rainwater and thus increased latent heat release within the data

assimilation system. It is speculated that further east of the

Front Range this could be an even greater problem since there

tends to be more stratiform rainfall associated with the pres-

ence of stronger large-scale forcing.

The HRRRv2 tended to underforecast the rainfall amount

observed during convective rain events. HRRRv2 nowcasts

was often delayed in both initiation and growth of convective

precipitation. In particular, the model significantly under-

estimated rainfall in situations where there was a large increase

in low-level convergence and/or boundary layer convergence

lines were apparent in the radar data. It is believed that the grid

spacing of the model/data assimilation of HRRRv2 was inad-

equate for resolving mesogamma-scale convergence features

responsible for initiating storms and thus missed many cases of

CI and rapid growth that were clearly associated with these

low-level features. Moreover, these low-level convergence

areas are often not well observed in the Doppler velocity field

despite being clearly apparent in radar reflectivity field as thin

lines of enhanced reflectivity that delineated areas of upward

vertical motions (Wilson et al. 1994). These updrafts may ex-

tend far above the low-level convergence and thus is much

more likely to be evident in radar reflectivity than in the low-

level wind field. While thin lines are easy to detect in radar

imagery, NWP does not have the ability to interpret them as

updrafts.

Since no corrections for removing radar bright band con-

tamination or assimilating boundary layer convergence lines

have been added to the HRRR up through 2020, versions of

the HRRR through 2020 will likely have the same errors that

are discussed in this study. Implementing an algorithm to

remove the bright band in the reflectivity field, prior to data

assimilation, would be an important first step to improve the

HRRR 1- and 2-h rainfall nowcasts.

Acknowledgments. The authors thank two presubmission

reviewers who were instrumental in improving the draft of this

paper. Dr. Stan Benjamin of NOAA, a primary developer of

the HRRR model, was particularly useful for making certain

we accurately represented the HRRR throughout the paper.

Dr. Tammy Weckwerth of NCAR was instrumental in im-

proving the organization and clarity of the paper. Author

Wilson (retired emeritus) thanks NCAR’s Research Applications

Program and Earth Observing Laboratory for office space,

computer resources, and publication funds. Author Megenhardt

was partially supported by the Short Term Explicit Prediction

program, which is supported by theNational Science Foundation

funds of the U.S. Weather Research Program. The contributions

of author Pinto were funded by the National Science

Foundation, which sponsors much of the research at NCAR.

Titan and the data viewer called CIDD are available online

(https://github.com/NCAR/lrose-core). Standard plotting soft-

ware (MicrosoftOffice Professional Plus 16)was used to prepare

the figures. The National Center for Atmospheric Research

(NCAR) is sponsored by the National Science Foundation.

REFERENCES

American Meteorological Society, 2020: Bright band. Glossary of

Meteorology, http://glossary.ametsoc.org/wiki/bright band.

Austin, G. L., P. Dionne, and M. Roch, 1987: On the interaction

between radar and satellite image nowcasting systems and

mesoscale numerical models. Proc. Mesoscale Analysis and

Forecasting, Vancouver, BC, Canada, European Space

Agency, 225–228.

Bellon, A., and G. L. Austin, 1978: The evaluation of two years

of real time operation of a short-term precipitation fore-

casting procedure (SHARP). J. Appl. Meteor., 17, 1778–

1787, https://doi.org/10.1175/1520-0450(1978)017,1778:

TEOTYO.2.0.CO;2.

Benjamin, S. G., and Coauthors, 2016: A North American hourly

assimilation and model forecast cycle: The Rapid Refresh.

Mon. Wea. Rev., 144, 1669–1694, https://doi.org/10.1175/

MWR-D-15-0242.1.

Bowler, N. E., C. E. Pierce, and A. W. Seed, 2006: STEPS: A

probabilistic precipitation forecasting scheme which merges

an extrapolation nowcast with downscaled NWP. Quart.

J. Roy. Meteor. Soc., 132, 2127–2155, https://doi.org/10.1256/

qj.04.100.

Browning, K. A., 1980: Review Lecture: Local weather forecasting.

Proc. Roy. Soc. London, 371A, 179–211, https://doi.org/

10.1098/rspa.1980.0076.

Dixon, M., and G. Wiener, 1993: TITAN: Thunderstorm identifi-

cation, tracking, analysis and nowcasting—A radar-based

methodology. J. Atmos. Oceanic Technol., 10, 785–797,

https://doi.org/10.1175/1520-0426(1993)010,0785:TTITAA.2.0.CO;2.

——, J. W. Wilson, T. M. Weckwerth, D. Albo, and E. J.

Thompson, 2015: A dual-polarization QPE method based on

the NCAR particle ID algorithm—Description and prelimi-

nary results. 37th Conf. on Radar Meteorology, Norman, OK,

Amer. Meteor. Soc., 9A.1, https://ams.confex.com/ams/

37RADAR/webprogram/Paper275705.html.

4796 MONTHLY WEATHER REV IEW VOLUME 148

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

Doswell, C. A., III, 1986: Short-range forecasting. Mesoscale

Meteorology and Forecasting, P. Ray, Ed., Amer. Meteor.

Soc., 689–719.

Dowell, D. C., L. J. Wicker, and C. Snyder, 2011: Ensemble

Kalman filter assimilation of radar observations of the 8 May

2003 Oklahoma City supercell: Influences of reflectivity ob-

servations on storm-scale analysis. Mon. Wea. Rev., 139, 272–

294, https://doi.org/10.1175/2010MWR3438.1.

Duda, J. D., X. Wang, Y. Wang, and J. R. Carley, 2019: Comparing

the assimilation of radar reflectivity using the directGSI-based

ensemble–variational (EnVar) and indirect cloud analysis

methods in convection-allowing forecasts over the continental

United States. Mon. Wea. Rev., 147, 1655–1678, https://doi.org/

10.1175/MWR-D-18-071.1.

Foresti, L., I. V. Sideris, D. Nerini, L. Beusch, and U. Germann,

2019: Using a 10-year radar archive for nowcasting precipi-

tation growth and decay: A probabilistic machine learning

approach. Wea. Forecasting, 34, 1547–1569, https://doi.org/

10.1175/WAF-D-18-0206.1.

Franch, G., D. Nerini, M. Pendesini, L. Coviello, G. Jurman, and

C. Furlanello, 2020: Precipitation nowcasting with orographic

enhanced stacked generalization: Improving deep learning

predictions on extreme events. Atmosphere, 11, 267, https://

doi.org/10.3390/atmos11030267.

Garstang, M., and H. J. Cooper, 1981: The role of near surface

outflow in maintaining convective activity. Proc. Int. Symp.,

Hamburg, Germany, ESA SP, 161–168.

Ge, G., J. Gao, and M. Xue, 2013: Impact of assimilating mea-

surements of different state variables with a simulated supercell

storm and three-dimensional variational method.Mon.Wea. Rev.,

141, 2759–2777, https://doi.org/10.1175/MWR-D-12-00193.1.

Germann, U., and I. Zawadzki, 2002: Scale-dependence of the

predictability of precipitation from continental radar images.

Part I: Description of the methodology. Mon. Wea. Rev., 130,

2859–2873, https://doi.org/10.1175/1520-0493(2002)130,2859:

SDOTPO.2.0.CO;2.

——, and ——, 2004: Scale dependence of the predictability

of precipitation from continental radar images. Part II:

Probability forecasts. J. Appl. Meteor., 43, 74–89, https://doi.org/

10.1175/1520-0450(2004)043,0074:SDOTPO.2.0.CO;2.

——, ——, and B. Turner, 2006: Predictability of precipitation

from continental radar images. Part IV: Limits to predic-

tion. J. Atmos. Sci., 63, 2092–2108, https://doi.org/10.1175/

JAS3735.1.

Golding, B. W., 1998: Nimrod: A system for generating automated

very short range forecasts. Meteor. Appl., 5, 1–16, https://

doi.org/10.1017/S1350482798000577.

Gustafsson, N., and Coauthors, 2018: Survey of data assimilation

methods for convective-scale numerical weather prediction at

operational centres. Quart. J. Roy. Meteor. Soc., 144, 1218–

1256, https://doi.org/10.1002/qj.3179.

Han, L., J. Sun, W. Zhang, Y. Xiu, H. Feng and Y. Lin, 2017: A

machine learning nowcasting method based on-real-time re-

analysis data. J. Geophys. Res. Atmos., 122, 4038–4051, https://

doi.org/10.1002/2016JD025783.

Huang, L. X., G. A. Isaac, and G. Sheng, 2012: Integrating NWP

forecasts and observation data to improve nowcasting accu-

racy. Wea. Forecasting, 27, 938–953, https://doi.org/10.1175/

WAF-D-11-00125.1.

Hwang, Y., A. J. Clark, V. Lakshmanan, and S. E. Koch, 2015:

Improved nowcasts by blending extrapolation and model

forecasts. Wea. Forecasting, 30, 1201–1217, https://doi.org/

10.1175/WAF-D-15-0057.1.

Iskenderian, H., and Coauthors, 2010: Satellite data applications

for nowcasting of convection initiation. 14th Conf. on

Aviation, Range and Aerospace Engineering, Atlanta, GA,

Amer. Meteor. Soc., 5.2, https://ams.confex.com/ams/90annual/

techprogram/paper_161354.htm.

Janjic, T. N., and Coauthors, 2018: On the representation error in

data assimilation.Quart. J. Roy. Meteor. Soc., 144, 1257–1278,

https://doi.org/10.1002/qj.3130.

Kiel, C., F. Heinlein, and G. C. Craig, 2014: The convective ad-

justment time-scale as indicator of predictability of convective

precipitation. Quart. J. Roy. Meteor. Soc., 140, 480–490,

https://doi.org/10.1002/qj.2143.

Li, L., W. Schmid, and J. Joss, 1995: Nowcasting of motion and

growth of precipitation with radar over a complex orography.

J. Appl. Meteor., 34, 1286–1300, https://doi.org/10.1175/1520-

0450(1995)034,1286:NOMAGO.2.0.CO;2.

Ligda, M. G., 1953: The horizontal motion of small precipitation

areas as observed by radar. MIT Department of Meteorology

Tech. Rep. 21, 60 pp.

Mandapaka, P. V., U.Germann, L. Panziera, andA.Hering, 2012: Can

Lagrangian extrapolation of radar fields be used for precipitation

nowcasting over complex alpine orography?Wea. Forecasting, 27,

28–49, https://doi.org/10.1175/WAF-D-11-00050.1.

Mueller, C. K., J. W. Wilson, and N. A. Crook, 1993: The utility of

sounding and mesonet data to nowcast thunderstorm initia-

tion. Wea. Forecasting, 8, 132–146, https://doi.org/10.1175/

1520-0434(1993)008,0132:TUOSAM.2.0.CO;2.

Pagé, C., L. Fillion, and P. Zwack, 2007: Diagnosing summertime

mesoscale vertical motion: Implications for atmospheric data

assimilation. Mon. Wea. Rev., 135, 2076–2094, https://doi.org/

10.1175/MWR3371.1.

Pinto, J., W. Dupree, S. Weygandt, M. Wolfson, S. Benjamin, and

M. Steiner, 2010: Advances in the Collaborative Storm

Prediction for Aviation (CoSPA). 14th Conf. on Aviation,

Range, and Aerospace Meteorology, Atlanta, GA, Amer.

Meteor. Soc., J11.2, https://ams.confex.com/ams/90annual/

webprogram/Paper163811.html.

——, J.Grim, andM.Steiner, 2015:Assessment of theHigh-Resolution

Rapid Refresh model’s ability to predict mesoscale convective

systems using object-based evaluation. Wea. Forecasting, 30, 892–

913, https://doi.org/10.1175/WAF-D-14-00118.1.

——, M. Steiner, K. Stone, and D. Albo, 2018: Quantifying the

relationship between storm size and forecast uncertainty for

enhanced blending. Sixth Aviation, Range, and Aerospace

Meteorology Special Symp., Austin, TX, Amer. Meteor.

Soc., 3.6, https://ams.confex.com/ams/98Annual/webprogram/

Paper335500.html.

Roberts, R. D., A. R. S. Anderson, E. Nelson, B. G. Brown, J. W.

Wilson, M. Pocernich, and T. Saxen, 2012: Impacts of fore-

caster involvement on convective storm initiation and evolu-

tion nowcasting. Wea. Forecasting, 27, 1061–1089, https://

doi.org/10.1175/WAF-D-11-00087.1.

Seed, A. W., C. E. Pierce, and K. Norman, 2013: Formulation and

evaluation of a scale decomposition-based stochastic precipi-

tation nowcast scheme. Water Resour. Res., 49, 6624–6641,

https://doi.org/10.1002/wrcr.20536.

Sharif, H. O., D. Yates, R. Roberts, and C. Mueller, 2006: The use

of an automated nowcasting system to forecast flash floods in

an urban watershed. J. Hydrometeor., 7, 190–202, https://

doi.org/10.1175/JHM482.1.

Stensrud, D. J., and Coauthors, 2009: Convective-Scale Warn-on-

Forecast System: A vision for 2020. Bull. Amer. Meteor. Soc.,

90, 1487–1500, https://doi.org/10.1175/2009BAMS2795.1.

DECEMBER 2020 W I L SON ET AL . 4797

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC

Sun, J., and Coauthors, 2014: Use of NWP for nowcasting con-

vective precipitation: Recent progress and challenges. Bull.

Amer. Meteor. Soc., 95, 409–426, https://doi.org/10.1175/

BAMS-D-11-00263.1.

Tsonis, A. A., and G. L. Austin, 1981: An evaluation of extrapo-

lation techniques for the short-term prediction of rain

amounts.Atmos.–Ocean, 19, 54–65, https://doi.org/10.1080/

07055900.1981.9649100.

Vivoni, E. R., D. Entekhabi, R. Bras, V. Ivanov, M. Van Home,

C.Grassoti, andR.Hoffman, 2006: Extending the predictability of

hydrometeorological flood events using radar rainfall nowcasting.

J. Hydrometeor., 7, 660–677, https://doi.org/10.1175/JHM514.1.

Weckwerth, T. M., 2000: The effect of small-scale moisture vari-

ability on thunderstorm initiation. Mon. Wea. Rev., 128,

4017–4030, https://doi.org/10.1175/1520-0493(2000)129,4017:

TEOSSM.2.0.CO;2.

Wheatley, D.M., K.H. Knopfmeier, T.A. Jones, andG. J. Creager,

2015: Storm-scale data assimilation and ensemble forecasting

with the NSSL Experimental Warn-on-Forecast System: Part

I: Radar data experiments. Wea. Forecasting, 30, 1795–1817,

https://doi.org/10.1175/WAF-D-15-0043.1.

Wilson, J. W., and W. E. Schreiber, 1986: Initiation of convective

storms by radar-observed boundary layer convergent lines.

Mon. Wea. Rev., 114, 2516–2536, https://doi.org/10.1175/

1520-0493(1986)114,2516:IOCSAR.2.0.CO;2.

——, and C. K. Mueller, 1993: Nowcasts of thunderstorm initi-

ation and evolution. Wea. Forecasting, 8, 113–131, https://doi.org/10.1175/1520-0434(1993)008,0113:NOTIAE.2.0.

CO;2.

——, T. M. Weckwerth, J. Vivekanandan, R. M. Wakimoto, and

R. W. Russell, 1994: Boundary layer clear-air radar echoes:

Origin of echoes and accuracy of derived winds. J. Atmos.

Oceanic Technol., 11, 1184–1206, https://doi.org/10.1175/1520-

0426(1994)011,1184:BLCARE.2.0.CO;2.

——, N. A. Crook, C. K. Mueller, J. Sun, and M. Dixon, 1998:

Nowcasting thunderstorms: A status report.Bull. Amer.Meteor.

Soc., 79, 2079–2099, https://doi.org/10.1175/1520-0477(1998)

079,2079:NTASR.2.0.CO;2.

——,E. Ebert, T. Saxen, R. Roberts, C.Mueller,M. Sleigh, C. Pierce,

and A. Seed, 2004: Sydney 2000 Forecast Demonstration

Project: Convective storm nowcasting.Wea. Forecasting, 19,

131–150, https://doi.org/10.1175/1520-0434(2004)019,0131:

SFDPCS.2.0.CO;2.

WMO, 2017: Guidelines for nowcasting techniques. WMO Doc.

WMO-1198, 82 pp., https://library.wmo.int/doc_num.php?

explnum_id53795.

4798 MONTHLY WEATHER REV IEW VOLUME 148

Brought to you by University of Colorado Libraries | Unauthenticated | Downloaded 03/19/21 06:55 PM UTC