Intercomparison of In-Flight Icing Algorithms. Part I: WISP94 Real-Time Icing Prediction and Evaluation Program

Abstract

The purpose of the Federal Aviation Administration's Icing Forecasting Improvement Program is to conduct research on icing conditions both in flight and on the ground. This paper describes a portion of the in-flight aircraft icing prediction effort through a comprehensive icing prediction and evaluation project conducted by the Research Applications Program at the National Center for Atmospheric Research. During this project, inflight icing potential was forecast using algorithms developed by RAP, the National Weather Service's National Aviation Weather Advisory Unit, and the Air Force Global Weather Center in conjunction with numerical model data from the Eta, MAPS, and MM5 models. Furthermore, explicit predictions of cloud liquid water were available from the Eta and MM5 models and were also used to forecast icing potential. To compare subjectively the different algorithms, predicted icing regions and observed pilot reports were viewed simultaneously on an interactive, real-time display. To measure objectively the skill of icing predictions, a rigorous statistical evaluation was performed in order to compare the different algorithms (details and results are provided in Part II). Both the subjective and objective comparisons are presented here for a particular case study, whereas results from the entire project are found in Part II. By statistically analyzing 2 months worth of data it appears that further advances in temperature and relative-humidity-based algorithms are unlikely. Explicit cloud liquid water predictions, however, show promising results although still relatively new in operational numerical models.

Full text

Current Icing Potential: Algorithm Description and Comparison with
Aircraft Observations
BEN C. BERNSTEIN,FRANK MCDONOUGH,MARCIA K. POLITOVICH, AND BARBARA G. BROWN
Research Applications Program, National Center for Atmospheric Research,* Boulder, Colorado
THOMAS P. RATVASKY AND DEAN R. MILLER
NASA Glenn Research Center, Cleveland, Ohio
CORY A. WOLFF AND GARY CUNNING
Research Applications Program, National Center for Atmospheric Research,* Boulder, Colorado
(Manuscript received 14 April 2004, in final form 5 December 2004)
ABSTRACT
The “current icing potential” (CIP) algorithm combines satellite, radar, surface, lightning, and pilot-
report observations with model output to create a detailed three-dimensional hourly diagnosis of the
potential for the existence of icing and supercooled large droplets. It uses a physically based situational
approach that is derived from basic and applied cloud physics, combined with forecaster and onboard flight
experience from field programs. Both fuzzy logic and decision-tree logic are applied in this context. CIP
determines the locations of clouds and precipitation and then estimates the potential for the presence of
supercooled liquid water and supercooled large droplets within a given airspace. First developed in the
winter of 1997/98, CIP became an operational National Weather Service and Federal Aviation Adminis-
tration product in 2002, providing real-time diagnoses that allow users to make route-specific decisions to
avoid potentially hazardous icing. The CIP algorithm, its individual components, and the logic behind them
are described.
1. Introduction
In-flight icing is a significant threat to aircraft, result-
ing in loss of lift, reduced airspeed, and, in some cases,
loss of control. This threat has been underscored by two
recent accidents. An ATR-72 (all acronyms in this pa-
per are defined in the appendix) holding in icing con-
ditions crashed near Roselawn, Indiana, in 1994, and an
EMB-120 descending through icing conditions crashed
on final approach to Detroit, Michigan, in 1997, result-
ing in the deaths of all 68 and 29 people on board,
respectively. Following a brief icing encounter in 2001,
another commuter aircraft lost control and plunged
8000 ft (2440 m), but regained control and made an
emergency landing.
Following the ATR-72 crash, the National Transpor-
tation Safety Board (NTSB) recommended that the
Federal Aviation Administration (FAA) “continue to
sponsor the development of methods to produce
weather forecasts that both define specific locations of
atmospheric icing conditions (including freezing drizzle
and freezing rain) and produce short-range forecasts
(‘nowcasts’) that identify icing conditions for a specific
geographic area with a valid time of 2 hours or less”
(National Transportation Safety Board 1996). To that
end, the National Center for Atmospheric Research
(NCAR), under the FAA’s Aviation Weather Re-
search Program, expanded their in-flight icing research
efforts on the diagnosis and forecasting of supercooled
large droplets (SLD; droplet diameters greater than 50
␮
m). Through case studies (e.g., Politovich and Bern-
stein 1995; Rasmussen et al. 1995), examination of ic-
* The National Center for Atmospheric Research is sponsored
by the National Science Foundation.
Corresponding author address: Ben C. Bernstein, National Cen-
ter for Atmospheric Research, Research Applications Program,
P.O. Box 3000, Boulder, CO 80307-3000.
E-mail: bernstei@rap.ucar.edu
V
OLUME 44 JOURNAL OF APPLIED METEOROLOGY JULY 2005
© 2005 American Meteorological Society
969
JAM2246

ing-related aircraft accidents (e.g., Marwitz et al. 1997),
and directing research aircraft into icing, including SLD
(Rasmussen et al. 1992; Miller et al. 1998; Ryerson et al.
2000; Isaac et al. 2001), NCAR meteorologists have
identified robust signatures in observational and model
datasets that are linked to the presence and intensity of
icing conditions.
One result of this work has been the development of
improved icing diagnosis and forecast techniques, in-
cluding the “current icing potential” (CIP). CIP com-
bines satellite, radar, surface, lightning, and pilot-report
observations with numerical model output to create an
hourly three-dimensional diagnosis of the potential for
icing and SLD. First developed during the winter of
1997/98, it became an official FAA and National
Weather Service (NWS) product in 2002 (available on-
line at http://adds.aviationweather.gov). Since that
time, numerous upgrades have been made that will be
implemented in the operational system in 2005. In this
paper, the upgraded version of CIP is described, in-
cluding the information extracted from each dataset it
employs, the application of fuzzy-logic membership
functions and a decision tree, and how the data are
integrated to diagnose icing. CIP’s abilities and short-
comings are demonstrated using example cases
sampled by a research aircraft and a brief verification
using pilot reports. A more in-depth verification of CIP
was completed as part of its assessment for FAA and
NWS approval (Brown et al. 2001).
2. Recent in-flight icing diagnosis and forecast
techniques
Forecasters have used rules of thumb, observations,
and numerical weather prediction model output to di-
agnose and forecast in-flight icing for many years (e.g.,
Jensen 1963; Air Weather Service 1980). The advent of
modern observation networks and improved numerical
models has led to recent advances in this arena. Schultz
and Politovich (1992) found combinations of model
temperature T and relative humidity RH that were co-
incident with icing by comparing pilot reports (PIREPs)
with Nested Grid Model output. They created a simple,
two-level icing product based on T and RH thresholds.
Forbes et al. (1993) and Thompson et al. (1997a) fur-
thered this approach by characterizing four meteoro-
logical situations for icing, based on combinations of T,
RH, and vertical thermodynamic structure. In an at-
tempt to find more intense icing, Carriere et al. (1997)
combined upward vertical motion with favorable T and
RH ranges. Although these purely model-based algo-
rithms capture a large percentage of PIREPs, they tend
to overforecast icing, even indicating it in cloud-free
areas (Brown et al. 1997; Thompson et al. 1997a,b).
Another purely model-based approach to forecasting
icing is the use of microphysics schemes to forecast su-
percooled liquid water (SLW) explicitly (e.g., Reisner
et al. 1998; Tremblay and Glazer 2000; Thompson et al.
2004). Verification of the first two schemes showed that
they captured less than one-half of the icing observed
but were very efficient because of the small volume of
airspace they covered (Guan et al. 2001; Brown et al.
2001).
Several purely observation-based icing diagnosis
techniques have also been developed. For example, Lee
et al. (1997), Ellrod (1996), and Smith et al. (2002) used
multispectral satellite data to discriminate between
cloudy and cloud-free areas and to identify cloud tops
likely to contain supercooled liquid water. These tech-
niques work well for single-layer clouds that are illumi-
nated by sunlight, but alternative techniques used at
night are not as robust, and both break down near the
solar terminator (the transition between areas that are
and are not illuminated by sunlight). The techniques
also miss potential icing clouds obscured from satellite
view by higher cloud layers. Radar-based techniques
have used polarization signals in attempts to differen-
tiate between water droplets and ice crystals (e.g.,
Reinking et al. 1997; Vivekanandan et al. 1999). It is
clear that a variety of data sources can be used to di-
agnose icing, each of which has its strengths and weak-
nesses. Rather than using a single data source, a system
that uses multiple data sources is likely to diagnose and
forecast icing conditions with greater accuracy (Carri-
ere et al. 1997).
Subsequent efforts attempted to combine datasets to
mitigate the problems associated with single-data-
source approaches to detection and forecasting of icing.
Bernstein (1996) used a gridded analysis of surface ob-
servations to limit model-based icing diagnoses to
cloudy and precipitating areas. Some situations condu-
cive to the presence of SLD were identified, based on
the principle that surface observations of freezing
drizzle, freezing rain, and ice pellets provide direct in-
dications of the existence of SLD aloft (e.g., Hanesiak
and Stewart 1995; Politovich and Bernstein 1995).
Thompson et al. (1997b) compared multispectral satel-
lite data with model temperature grids to eliminate
cloud-free areas and altitudes from their T/RH-based
icing scheme. Tafferner et al. (2003) used a version of
the four-category Thompson et al. (1997a) scheme and
then confirmed or corrected the first-guess icing mecha-
nism through comparison with surface observations and
radar data. Le Bot (2004) combined satellite and radar
data with model output, associated more severe icing
with warm cloud tops, and related radar reflectivity to
droplet size. The latter two systems demonstrate that
970 JOURNAL OF APPLIED METEOROLOGY VOLUME 44

the intelligent combination of data from multiple data
sources can be useful in the diagnosis of icing.
All of the techniques described above apply hard
thresholds to create their diagnoses, indicating icing at
T ⫽⫺14.9°C but not at T ⫽⫺15.1°C, for example.
Thus, small changes in temperature, relative humidity,
or other parameters could result in abrupt changes in
the icing field. Such an approach is not representative
of the more gradual transitions from icing to nonicing
environments that exist in nature. Like many weather
phenomena, icing has both discontinuous and continu-
ous aspects. Abrupt discontinuities occur at cloud
boundaries, where icing can change from significant to
nonexistent over a few hundred meters, especially in
the vertical direction. Icing conditions can also change
gradually in time and space. An example is a gradual
increase in temperature from ⫺6° to 0°C as an aircraft
flies at a constant altitude within an otherwise similar
cloud. The nature of icing is suited well to a hybrid
approach to its diagnosis. Black-and-white decisions
make sense for some aspects, such as the presence or
absence of clouds, whereas other aspects such as tem-
perature lend themselves well to shades of gray.
To maximize the value of multiple data sources and
to represent better the hybrid nature of icing condi-
tions, CIP merges satellite, surface, radar, lightning,
and PIREP observations with model forecasts of T,
RH, SLW, and vertical velocity and then uses fuzzy-
logic and decision-tree techniques to determine the
likelihood of icing and SLD at each location. This ap-
proach is designed to maximize the strengths and to
minimize the weaknesses of each dataset while mimick-
ing manual techniques used by NCAR meteorologists
to direct research aircraft into icing and SLD. The goal
is to indicate the maximum potential for icing and SLD
conditions in each part of the 3D model domain. The
icing and SLD potential values range from 0.0 to 1.0.
Although they are not calibrated as true probabilities,
high (low) values indicate a relatively high (low) chance
for icing and SLD to be present.
3. The CIP technique
CIP determines icing and SLD potentials in a step-
wise fashion (see Fig. 1 for a conceptual diagram and
Fig. 2 for a flowchart of the process) and will be de-
scribed in this way. In step 1, the datasets are placed
onto a common grid. In step 2, the 3D locations of
clouds and precipitation are found using satellite, sur-
face, and radar observations. In step 3, fuzzy-logic
membership functions are applied to icing-related fields
to create interest maps. In step 4, the physical icing
situation is determined by using a decision tree. In step
5, the initial icing and SLD potentials are calculated by
situationally combining interest maps from basic fields
(e.g., T, RH). In step 6, the final icing potential is cal-
culated by increasing or decreasing the initial icing po-
tential using the vertical velocity, SLW, and PIREP in-
terest maps.
a. Step 1: Place the datasets onto a common grid
The first step in the process is to map current satel-
lite, surface, radar, lightning, and PIREP observations
FIG. 1. CIP conceptual diagram. Precipitation types: snow (asterisks), rain (large open
circles), and freezing drizzle (small gray circles).
J
ULY 2005 B E R N STEIN ET AL. 971

and explicit model forecasts of supercooled liquid water
content to the Rapid Update Cycle (RUC) model
(Benjamin et al. 2004) pressure grid. The RUC pressure
grid supplies the temperature, relative humidity, verti-
cal velocity, and geopotential height fields with 25-hPa
vertical and 20-km horizontal grid spacing. Fields from
the RUC 3-h forecast are typically used, but CIP could
be run using other forecast lengths and even different
models. The 0-h RUC diagnosis is not used because
moisture parameters, including cloud microphysics,
typically need several hours to spin up.
1) SATELLITE DATA
To determine cloud locations and cloud-top tempera-
tures (CTT), CIP uses observations from the following
NWS Geostationary Operational Environmental Satel-
lite (GOES) imager channels and combinations
thereof: visible, shortwave and longwave infrared,
shortwave reflectance (Turk et al. 1998), and the
NCAR satellite-icing algorithm (Thompson et al.
1997b). The high-resolution satellite data are projected
to a Lambert-conformal 5-km grid. Sixteen pixels are
matched to each 20-km RUC model grid box. The vis-
ible and shortwave infrared reflectance fields are de-
pendent on solar reflection, and so the day is broken
into three distinct periods (day, night, and solar termi-
nator; see Table 1) to apply appropriate cloud detection
tests that will be described in section 3b.
2) SURFACE OBSERVATIONS
Surface observations (METARs) are used to deter-
mine ceiling height, precipitation occurrence, and pre-
cipitation type. Although generally accurate, these
point observations are made at irregularly spaced loca-
tions, and the number of stations reporting in the vi-
cinity of a grid box may vary. For this reason, CIP
incorporates information from surface stations using a
concentric-circle approach, assuming that data from the
surface observations closest to a given grid box are
most representative of the conditions present within it.
CIP initially searches for METARs within 40 km of the
center of the grid box for at least one observation of
FIG. 2. Flowchart of the CIP process.
972 JOURNAL OF APPLIED METEOROLOGY VOLUME 44

cloud cover and each of six precipitation types. If such
data are not found within 40 km, then the search con-
tinues using incrementally larger circles up to a maxi-
mum radius of 125 km.
3) RADAR MOSIAC
Mosaics of NWS Next-Generation Weather Radar
(NEXRAD) data are used to refine the precipitation
fields. They are available at 4-km resolution, and 25
pixels are mapped to each RUC grid box. The percent-
age of the grid box filled with echoes exceeding 18 dBZ
(light precipitation) is saved. The 18-dBZ threshold is
used because until recently it was the minimum reflec-
tivity available for use in the development code. Future
versions of CIP will take advantage of the full range of
reflectivity, which can be useful in the diagnosis of icing
when applied correctly.
4) EXPLICIT SLW FROM THE RUC NATIVE GRIDS
RUC’s explicit microphysics fields, including SLW,
are not included in the pressure grids, and so they must
be extracted from the native, “hybrid b” grids. Hori-
zontal spacing and valid times are the same for both
grids, and so the SLW amounts are linearly interpo-
lated in the vertical direction to the pressure grid.
5) PILOT REPORTS AND LIGHTNING
PIREPs indicate point observations of the presence
(or absence) and severity of icing. Each report contains
the geographic location, altitude, and time of the icing.
If a PIREP occurred within the last hour, then its hori-
zontal and vertical distances from nearby RUC grid
points (within 150 km), as well as the reported icing
severity, are saved. The influence of a given PIREP
within this range decreases with increasing distance.
This fact will be described in greater detail later.
Lightning observations provide the geographic loca-
tion and time of the strikes. Those strikes that occurred
in close proximity to an RUC grid point (within 25 km)
during the 15 min prior to CIP valid time are mapped to
the grid. The relatively small space and time scales used
for lightning are appropriate for deep convection, be-
cause it typically results in relatively small scale icing
conditions.
b. Step 2: Find the 3D locations of clouds and
precipitation
Once the datasets have been mapped to the model
grid, the matched data are examined to determine
whether clouds are present in each model grid box. If
adequate cloudiness is found, then cloud-base and
cloud-top heights, as well as the presence of precipita-
tion and its type, are assessed.
1) CLOUDINESS
The presence of clouds is determined using an exten-
sion of the Lee et al. (1997) and Thompson et al.
(1997a) technique, which applies time-of-day-specific
thresholds on data from the various satellite channels
(see Table 1). During the day and night, the model grid
box is considered to be a candidate for icing if at least
40% of the 16 matched satellite pixels are cloudy, be-
cause icing rarely occurs with less than broken cloud
cover (Bernstein et al. 1997). In the solar terminator,
satellite data are not used alone to determine whether
grid boxes are cloudy because several fields that are
dependent on solar radiation are contaminated. During
these narrow time windows around sunrise and sunset,
the model grid box is considered to be cloudy if one or
more of the following are true: the satellite-measured
infrared temperature is greater than ⫺35°C and surface
observations mapped to it indicate at least broken sky
conditions, the infrared temperature was less than
TABLE 1. Satellite cloud diagnosis (adapted from Thompson et
al. 1997b). Channel 1 ⫽ 0.65
␮
m, channel 2 ⫽ 3.9
␮
m, and channel
4 ⫽ 10.7
␮
m for GOES imagers.
Daytime (solar zenith angle ⬍70°)
Cloudy if channel-1 albedo* ⱖ 20% and any of the following
are true:
1) NCAR Research Applications Program satellite icing
algorithm indicates icing
2) 20°C ⬎ channel 4 ⱖ ⫺35°C and (channel 2 ⫺ channel
4 ⬎ 10°C)
3) ⫺35°C ⬎ channel 4
4) ⫺10°C ⱖ channel 4 ⱖ ⫺35°C and surface-observed
cloud cover at least broken
5) Channel 4 ⱖ 15°C colder than the model temperature
at topography
6) Channel 2 reflectance ⱖ 20%
Nighttime [solar zenith angle ⱖ 90° (sun is below the horizon)]
Cloudy if any of the following are true:
1) Channel 4 ⬍ 5°C and [(channel 2 ⫺ channel 4 ⱕ
⫺2.1°C) or (channel 2 ⫺ channel 4 ⱖ 4.1°C)]
2) Channel 4 ⬍⫺35°C
3) 5°C ⱖ channel 4 ⱖ ⫺35°C and surface-observed cloud
cover is at least broken
4) 20°C ⬎ channel 4 ⬎ 5°C and (channel 2 ⫺ channel 4 ⬍
⫺2.1°C)
Terminator (70° ⬍ solar zenith angle ⬍ 90°)
Cloudy if any of the following are true:
1) ⫺35°C ⬎ channel 4
2) 20°C ⱖ channel 4 ⱖ ⫺35°C and surface-observed
cloud cover is at least broken
3) Channel 4 ⱖ 15°C colder than the model temperature
at the ground
* Channel-1 albedo is normalized by dividing by the cosine of the
solar zenith angle.
J
ULY 2005 B E R N STEIN ET AL. 973

⫺35°C, or the satellite-measured infrared temperature
was at least 15°C colder than the RUC model–
predicted surface temperature. The latter two tests in-
fer cloud presence based on temperatures that are very
different than those expected at the surface within the
continental United States. The ⫺35°C threshold would
not apply in particularly cold climates. Grid boxes that
do not meet the cloudiness parameters described above
are considered to be ice free (icing and SLD potentials
⫽ 0.0 throughout the column).
2) CLOUD-TOP HEIGHT
Cloud-top height is estimated by finding the coldest
satellite-measured infrared temperature among the
cloudy pixels and comparing it with the profile of model
temperature in a top-down manner [similar to Thomp-
son et al. (1997b)]. Once a model temperature greater
than the satellite-measured temperature is found, cloud
top is set to the next model level above, because clouds
are likely to exist somewhere between these two levels.
All altitudes above the highest cloud top are considered
to be ice free. When strong inversions are present in the
model at temperatures close to the satellite-measured
cloud-top temperature, this method can overestimate
the cloud-top height and, in some cases, the top of the
icing layer.
3) CLOUD-BASE HEIGHT, PRECIPITATION
PRESENCE
, AND PRECIPITATION TYPE
Using the concentric-circle approach described
above, CIP searches for METARs that provide obser-
vations of ceiling height and precipitation type. Ceiling
height is typically measured well by all weather station
platforms, and so cloud-base height is set to the lowest
value (m MSL) of those reported by all of the stations
within the first circle in which ceiling observations are
found. In the absence of precipitation, all altitudes be-
low cloud base are considered to be ice free.
It is important to determine the presence of precipi-
tation and its type, because subfreezing liquid precipi-
tation within and below the clouds can also pose an
icing hazard. To make this determination, CIP searches
the concentric circles for information about the pres-
ence or absence of six categories of precipitation: freez-
ing drizzle, freezing rain, ice pellets, rain, drizzle, and
snow. A report of any of the first five precipitation
types means that altitudes below cloud base need to be
considered for possible icing and SLD, because sub-
freezing liquid precipitation may be present.
The determination of precipitation occurrence and
type from METARs is somewhat complicated, because
the precipitation information available varies among
station types. Automated stations, even of the same
type, may have very different capabilities. This situa-
tion is accounted for by determining whether informa-
tion on each of the six precipitation types is available
within each concentric circle. Once good information
on a given precipitation type is found within a given
circle, the search for that type ceases. For example,
information on the presence or absence of freezing rain
is found if a manual station or an automated station
with a working freezing-rain sensor is present within
the circle. That same automated station may not be
capable of reporting freezing drizzle (Wade 2003), and
so the search for that precipitation type must continue.
In the absence of METARs reporting precipitation,
the subcloud layer may also be further considered for
possible icing based on the presence of radar echoes.
Because the radar cannot directly provide information
on precipitation type, the potential for supercooled liq-
uid precipitation below cloud base cannot be negated.
c. Step 3: Apply fuzzy-logic membership functions
to icing-related fields
An important element of the CIP technique is the use
of fuzzy-logic membership functions to develop interest
maps for the T, RH, CTT, vertical velocity, SLW, and
PIREP fields. Rather than applying thresholds, CIP at-
tempts to handle uncertainties evident in the datasets it
employs and to mimic the gradual transition from icing
to nonicing environments associated with each field,
based on cloud physics principles, experience gained
from in-flight icing field programs, and distributions of
icing PIREPs relative to these parameters.
1) TEMPERATURE
Both research flight observations and basic cloud
physics concepts support the notion that clouds and
precipitation are more likely to contain SLW at certain
temperatures. SLW is most common at temperatures
close to freezing, becomes less frequent with decreasing
temperature, and is relatively rare at temperatures less
than ⫺25°C (Korolev et al. 2003), except in deep con-
vection and isolated “clean” clouds (e.g., Cober et al.
2001). Conversely, ice crystals are unlikely to form at
temperatures close to freezing and become increasingly
common as temperature decreases beyond ⫺10°C
(Rogers and Yau 1989; Rauber et al. 2000; Cober et al.
2001; Korolev et al. 2003). Data collected during sev-
eral field programs indicate that icing occurred most
often at temperatures between ⫺15° and ⫺3°C (Sand et
al. 1984; Schultz and Politovich 1992; Cober et al. 1995).
Based on cloud physics principles, observations, and
forecaster experience, the temperature interest map
974 JOURNAL OF APPLIED METEOROLOGY VOLUME 44

(T
map
) is designed to indicate the likelihood of SLW
that may freeze onto an aircraft, given only tempera-
ture from the model. Here, T
map
is maximized at tem-
peratures at which SLW is most frequently expected
and drops off as glaciation becomes more likely on the
cold side and as SLW becomes less likely to freeze onto
an aircraft because of compressional heating on the
warm side (Fig. 3a). The convective T
map
curve allows
for a much greater chance for icing on the cold end of
the spectrum, down to ⫺30°C, because strong upward
motion associated with convection allows SLW to exist
at relatively cold temperatures (Rosenfeld and Wood-
ley 2000).
Comparisons of RUC temperatures with 19 057 icing
PIREPs made between November of 2002 and March
of 2003 indicated that peak frequencies were centered
on ⫺7°C, with a gradual decrease on the cold side and
a sharp decrease on the warm side of that peak. About
4.6% of the PIREPs occurred at temperatures colder
than ⫺25°C, but indicating icing at such cold tempera-
tures would result in a gross overforecast, except in
areas of deep convection. This fact is borne out by di-
viding the PIREP counts in each temperature bin (e.g.,
⫺8°C ⬍ T ⬍⫺6°C) by the overall frequency of occur-
rence of model forecasts of that temperature range.
This calculation results in a tiny value (on the order of
1 ⫻ 10
⫺4
) for each bin. The distribution of these rela-
tive PIREP frequency values by temperature is then
normalized to a 0–1 range. The result is the “normal-
ized PIREP curve,” which provides an indication of the
likelihood of icing for a forecast of a given temperature
(or other variable; see sections that follow). It shows
that icing is most likely for a model forecast of ⫺8°C ⬍
T ⬍⫺6°C but does not imply that icing will always be
found at such temperatures. Icing is relatively infre-
quent, but is not completely absent, when relatively
cold temperatures are forecast (Fig. 3a). The curve
nicely matches the independently developed T
map
curve, except at temperatures for which glaciation is
expected to be likely (T ⬍⫺20°C) and at above-
freezing temperatures, for which icing should not occur.
The fact that 3.5% of the PIREPs occurred at above-
FIG. 3. (a) Standard and convective temperature interest maps (T
map
, T
map-convective
), and normalized icing PIREP ratio (frequency
of PIREPs at a given temperature divided by the frequency of occurrence of that temperature in the model, normalized to a range of
0–1). (b) As in (a), but for CTT. (c) As in (a), but for RH. (d) As in (a), but for VV.
J
ULY 2005 B E R N STEIN ET AL. 975

freezing temperatures is likely due to errors in reported
icing altitudes (Brown et al. 1999) and to incorrect
model forecasts.
2) CLOUD-TOP TEMPERATURE
The phase of hydrometeors at cloud top can affect
the composition of the cloud below. Relatively warm
cloud tops imply that the cloud layer is likely to be
dominated by liquid water (Rauber et al. 2000). Con-
versely, if cloud tops are cold enough to produce ice
crystals, these crystals can grow and fall through the
clouds below, sometimes leading to complete glaciation
[as in Hill (1980) and Politovich and Bernstein (1995)].
Geresdi et al. (2005) found observational evidence of a
gradual transition from liquid- to ice-dominated pre-
cipitation as cloud-top temperature decreased.
The cloud-top temperature interest map (CTT
map
)
was developed to estimate the likelihood that clouds
contain liquid water rather than being glaciated for a
given CTT (Fig. 3b). Values are maximized for CTT ⬎
⫺12°C, because liquid water is likely to dominate in
such warm clouds, and drop off gradually with decreas-
ing CTT but never reach zero. Though cold cloud tops
imply that ice crystals are likely to dominate, liquid
water can still exist within such clouds if its production
rate exceeds the depletion rate. The normalized distri-
bution of CTT for 7855 icing PIREPs made in single-
layer clouds (as diagnosed by CIP) shows that the nor-
malized frequency of icing peaks around ⫺12°C (Fig.
3b). It drops off sharply as CTT increases beyond about
⫺8°C, because it becomes increasingly difficult to have
sufficiently deep layers with good icing temperatures
beneath such warm tops. CTT
map
does not decrease at
CTT ⬎⫺8°C because such warm cloud tops imply liq-
uid-dominated cloud processes. The distribution drops
off gradually with decreasing cloud-top temperatures
between ⫺12° and about ⫺30°C and then flattens out.
CTT
map
appears to drop off a little too much on the
cold end of the spectrum, but many of the PIREPs
associated with these cold CTTs are associated with
unresolved multilayered situations in which the icing
occurred in a lower deck with higher CTT.
3) RELATIVE HUMIDITY
CIP also employs a relative humidity map (RH
map
;
Fig. 3c) in many situations to assess the likelihood that
clouds exist between the observed cloud top and base
given the model RH forecast. Icing PIREPs matched to
RUC RH forecasts indicate that they occur most fre-
quently with high RH, as expected, and decrease gradu-
ally with decreasing RH. The relatively low RH values
matched to some PIREPs are mostly due to poor RH
forecasts, but some may be caused by errors in PIREP
locations (Brown et al. 1999). Overall, 74.9% of all
PIREPs occurred with RUC RH in excess of 70%, and
only 1.7% occurred with RH less than 25%. The RH
map
roughly matches the normalized PIREP distribution,
though it is perhaps a little too generous at moderately
high RH values (70%–90%). The relationship between
model RH and icing is mostly a function of model ac-
curacy rather than icing physics. Note that the meaning
of a given model RH forecast is highly dependent on
the model’s handling of moisture. Adjustments to
RH
map
and, perhaps, to other interest maps (e.g., ver-
tical velocity and explicit SLW) may be needed to
handle differences among models.
4) VERTICAL VELOCITY
Given that a cloud is present, upward vertical veloc-
ity can aid the production of liquid water, even in what
might normally be an ice-dominated environment.
Downward motion may indicate that liquid water is
waning. The CIP vertical velocity membership function
(VV
map
; Fig. 3d) is designed with these concepts in
mind, and it is used to adjust the initial estimates of
positive icing potential upward or downward by as
much as 25%, because upward and downward motion
similarly affect forecaster confidence that icing will be
present at ideal temperatures within a cloud. Distribu-
tions of PIREP occurrence with vertical velocity (VV)
forecasts from the RUC show that 65.6% of all icing
occurs in areas of forecast upward motion, but the nor-
malized distribution shows that forecasts of moderately
strong upward motions (⬍⫺0.5
␮
bar s
⫺1
) are much
more likely to be associated with icing than forecasts of
weak upward or any downward motions, as expected.
The stronger upward motions are relatively uncommon
in the model, but when they are present, icing is likely
to occur. The shape of the VV
map
curve roughly
matches the normalized VV distribution, though it was
developed independently, using forecaster experience
and cloud physics principles. The slight increase in the
normalized PIREP distribution at strong downward
motions (⬎0.5
␮
bar s
⫺1
) is associated with very few
data points.
5) EXPLICIT SUPERCOOLED LIQUID WATER
PREDICTIONS
The interest map for explicitly predicted supercooled
liquid water content (cloud and rain water content at
subfreezing temperatures) from the model is designed
somewhat differently because of the characteristics of
this field in the RUC. Brown et al. (2001) showed that
the RUC liquid water predictions captured ⬃40% of all
976 JOURNAL OF APPLIED METEOROLOGY VOLUME 44

icing PIREPs while warning for a small volume of air-
space. Practical experience has shown that icing is likely
to be present when the model predicts SLW, but the
lack of an SLW prediction has not proven to work well
as a negative indicator of icing. Because the CIP icing
potential field is intended to diagnose the presence of
SLW, rather than its amount, the SLW interest map
(SLW
map
, not shown) is set to unity when any SLW is
predicted and to zero when it is not. The SLW
map
is
used to boost but not to decrease the initial icing po-
tential, because the lack of SLW in a forecast is not a
robust indicator of a lack of icing. Planned upgrades to
the RUC microphysics package described by Thomp-
son et al. (2004) may allow for better use of the SLW
field, including forecasts of a lack of SLW.
6) PILOT REPORTS
A valuable piece of information for any icing fore-
caster is an actual icing report. PIREPs provide the
forecaster with the approximate time, location, and al-
titude of icing. PIREP shortcomings are documented
well and include nonuniformity in time or space and
contamination by errors in location, altitude, and time
(Brown et al. 1997; Kelsch and Wharton 1996). When a
positive icing PIREP is found in a location at which
icing is expected, however, it supplies increased confi-
dence that the diagnosis is correct. The relevance of a
PIREP decreases with increasing distance (horizontally
and vertically) and time since it was made. The PIREP
membership function considers the horizontal and ver-
tical distances from the center of each model grid box
to the nearest PIREP. In CIP, the influence of a given
PIREP can only extend to 150 km horizontally, 300 m
vertically, and 1 h temporally. Its influence decreases
rapidly with increasing distance and height differences
(see Fig. 4). The choices of range and height difference
are somewhat arbitrary but are reasonable, based on
forecaster experience. A PIREP’s influence could be
related to the uniformity of the clouds that surround it,
but this concept is yet to be implemented. The age of a
PIREP would ideally also be a factor, but all PIREPs
made within the last hour are treated equally in this
version of CIP.
Because the icing potential field is intended to show
the potential for any icing conditions, rather than a
specific icing severity, all reported severity levels are
treated equally and PIREPs of “no icing” are not used.
Negative-icing PIREPs often exist within areas of icing
and are sometimes indicative of embedded icing-free
pockets. An example of this situation is the group of
reports from aircraft that flew similar approach paths
within minutes of an EMB-120 aircraft that crashed on
its descent into Detroit on 9 January 1997. The pilots
reported that the icing was “not present,”“light,”
“moderate,” and “extremely heavy to severe.” The ap-
parent disagreement was attributed to small-scale
variations in the clouds and how the aircraft traversed
them (National Transportation Safety Board 1998).
This example demonstrates that small variations in lo-
cation, time, and approach can be the difference be-
tween no icing and icing that contributed to an accident
that killed 29 people.
d. Steps 4 and 5: Determine the physical icing
scenario and calculate the initial icing and SLD
potentials
Icing and SLD conditions result from many pro-
cesses, and so the meteorological structure that is
present must be identified and the data and interest
maps need to be applied in an appropriate manner.
This situational approach is critical, because the mean-
ing of an individual piece of data can be very different
FIG. 4. The pilot-report interest map, showing change in horizontal and vertical distance.
An “x” marks the location of a pilot report of icing.
J
ULY 2005 B E R N STEIN ET AL. 977

TABLE 2. Icing and SLD potential equations. Precipitation abbreviations are defined in the appendix.
Situation Subsituation Initial icing potential Final SLD potential
No clouds observed Not applicable 0.00 0.00
Above highest cloud top Not applicable 0.00 0.00
Below lowest cloud base No precipitation at the surface 0.00 0.00
Below lowest cloud base Only snow observed at the surface 0.00 0.00
Single-layer cloud with uniform tops
Single-layer clouds No precipitation observed at surface
or DZ/RA with CTT ⬍⫺12°C
T
map
CTT
map
RH
map
⫺9.9
Single-layer clouds SN is only precipitation observed at
the surface
{0.9 ⫺ [0.6(%radar)]}
CTT
map
T
map
RH
map
⫺9.9
Single-layer clouds Freezing precipitation (FZDZ, FZRA,
PE) at the surface; CTT ⬎⫺30°C
T
map
CTT
map
RH
map
{T
map
⫹ [0.2(%radar)]}
[(CTT
map
⫺ 0.4)/0.6] ⫹
(0.25T
map
)
Single-layer clouds Freezing precipitation (FZDZ, FZRA,
PE) at the surface; CTT ⬍⫺30°C
T
map
CTT
map
RH
map
0.25T
map
Single-layer clouds CTT ⱖ ⫺12°C; RA and/or DZ at the
surface or radar ⱖ 20% with
T ⬎⫺2°C and no SN
T
map
CTT
map
RH
map
T
map
[(CTT ⫺ 261.0)/14.0]
0.5
CTT gradient [Note: CTT
map
varies with altitude, based on T (see text)]
Single-layer cloud, with CTT
gradient
No precipitation observed at surface
or DZ/RA with CTT ⬍⫺12°C
T
map
CTT
map
RH
map
⫺9.9
Single-layer cloud, with CTT
gradient
SN is only precipitation observed at
the surface
{0.9 ⫺ [0.6(%radar)]}
CTT
map
T
map
RH
map
⫺9.9
Single-layer cloud, with CTT
gradient
Freezing precipitation (FZDZ, FZRA,
PE) at the surface; CTT ⬎⫺30°C
T
map
CTT
map
RH
map
{T
map
⫹ [0.2(%radar)]}
[(CTT
map
⫺ 0.4)/0.6] ⫹
(0.25T
map
)
Single-layer cloud, with CTT
gradient
Freezing precipitation (FZDZ, FZRA,
PE) at the surface; CTT ⬍⫺30°C
T
map
CTT
map
RH
map
0.25T
map
Single-layer cloud, with CTT
gradient
CTT ⱖ ⫺12°C; RA and/or DZ at the
surface or radar ⱖ 20% with
T ⬎⫺2°C and no SN
T
map
CTT
map
RH
map
T
map
[(CTT ⫺ 261.0)/14.0]
0.5
Multiple cloud layers (Note: CTT
map
is calculated using CTT of each layer)
Multiple cloud layers, lowest
deck
No precipitation observed at surface
or DZ/RA with CTT ⬍⫺12°C
T
map
CTT
map
RH
map
⫺9.9
Multiple cloud layers, lowest
deck
SN is only precipitation observed at
the surface
{0.9 ⫺ [0.6(%radar)]}
CTT
map
T
map
RH
map
⫺9.9
Multiple cloud layers, lowest
deck
Freezing precipitation (FZDZ, FZRA,
PE) at the surface; CTT ⬎⫺30°C
T
map
CTT
map
RH
map
{T
map
⫹ [0.2(%radar)]}
[(CTT
map
⫺ 0.4)/0.6] ⫹
(0.25T
map
)
Multiple cloud layers, lowest
deck
Freezing precipitation (FZDZ, FZRA,
PE) at the surface; CTT ⬍⫺30°C
T
map
CTT
map
RH
map
0.25T
map
Multiple cloud layers, lowest
deck
CTT ⱖ ⫺12°C; RA and/or DZ at the
surface or radar ⱖ 20% with
T ⬎⫺2°C and no SN
T
map
CTT
map
RH
map
T
map
[(CTT ⫺ 261.0)/14.0]
0.5
Multiple cloud layers, higher
decks
Treated as nonprecipitating, because
of presence of dry layer
T
map
CTT
map
RH
map
⫺9.9
Multiple cloud layers,
between decks
Conditions related to deck above; use
CTT
map
from it
T
map
CTT
map
RH
map
⫺9.9
Classical FZRA structure with FZRA, PE, RA, FZDZ, and/or DZ at the surface
Classical warm nose present Above the warm nose (above first
k level with T ⬎ 0°C)
T
map
CTT
map
RH
map
[T
map
⫺ 0.2(%radar)]
[(CTT
map
⫺ 0.7)/0.3]
Classical warm nose present Below the warm nose (at/below first
k level with T ⬎ 0°C)
T
map
⫹ [0.2(%radar)T
map
] T
map
⫹ [0.2(%radar)T
map
]
Deep convection
Deep convection Lightning observed within 25 km in
last 15 min
T
map
(convective) T
map
(convective)
978 JOURNAL OF APPLIED METEOROLOGY VOLUME 44

for different situations. CIP identifies five distinct icing
situations: single-layer clouds, multiple-layer clouds,
cloud-top temperature gradients, classical freezing rain,
and deep convection. The methods used to assess the
presence of icing for each situation are described be-
low. The equations used can be found in Table 2.
1) SINGLE-LAYER CLOUDS
The simplest of all icing situations is a single-layer,
nonprecipitating cloud. A single-layer cloud is consid-
ered to be present when high RH is found at all levels
between the satellite-derived cloud top and the
METAR-derived cloud base. An initial assessment of
cloud phase is made using the CTT
map
, and this assess-
ment is combined with the T
map
and RH
map
values at
each level to calculate the initial icing potential. When
there is no indication of the presence of precipitation-
sized liquid water droplets in radar and surface obser-
vations, the SLD potential is set to “unknown” (⫺9.9)
for all altitudes within cloud, because it is uncertain
whether SLD exists there.
The icing situation is more complex when precipita-
tion is observed beneath a cloud layer. Liquid precipi-
tation in combination with warm cloud tops implies that
the collision–coalescence process is active (Cober et al.
1996; Rauber et al. 2000). Thus, the potential for icing
should be high, both within the lowest cloud layer and
beneath its base, if the temperatures are in the proper
range. Large droplets are clearly present from the sur-
face up to at least cloud base and often extend upward
well into the cloud, sometimes to its top [as in Pobanz
et al. (1994)].
In a similar situation in which only snow is reported
at the surface, ice crystals are clearly present beneath
and within the lowest cloud layer. These crystals scav-
enge SLW through riming and may completely glaciate
the cloud (Geresdi et al. 2005). In such cases, CIP de-
creases the maximum possible icing potential somewhat
by including a snow factor in the equation (see Table 2).
When the snow is associated with widespread radar
echoes of greater than 18 dBZ, there is likely to be an
abundance of large ice crystals aloft, implying more
riming, and the icing potential is further lowered. As
more of the grid box is filled with snow echoes, this
factor becomes stronger, further decreasing the poten-
tial for icing.
2) MULTIPLE CLOUD LAYERS
When multiple cloud layers are present and no pre-
cipitation from the upper layer falls into the lower
layer, the icing situation for each layer should be con-
sidered separately. An example of this is a cirrus layer
passing over a low stratus layer. The top of the upper
cloud layer is observed by satellite, and so satellite data
can be used to determine its icing potential. The lower
layer is obscured from view, and so its location and
cloud top must be determined by other means. CIP
infers the presence of an intervening dry layer with
model RH of less than 50% over at least 75 hPa of
depth and infers a lower cloud layer by RH of more
than 70% beneath the dry layer. The lower cloud layer
has its own cloud-top temperature, and CIP separately
calculates the icing and SLD potentials for that layer.
When precipitation is present at the surface, its attrib-
utes are only applied to the lowest cloud layer and the
altitudes beneath it.
3) CLOUD-TOP TEMPERATURE GRADIENT
When a CTT gradient is present within a grid box, it
implies that a transition may be occurring. One portion
of the grid box may contain clouds with relatively warm
tops (e.g., ⫺12°C), for which supercooled liquid water
is expected, while another portion may contain clouds
with relatively cold tops (e.g., ⫺25°C), for which par-
tially or completely glaciated conditions are expected.
An aircraft flying at constant altitude across such a grid
TABLE 2. (Continued)
Maps and factors Situation for application Value
T
map
, SLW
map
, PIREP
map
,VV
map
All See text and figures
T
map
(convective) Deep convection—lightning observed
within 25 km in last 15 min
See text and figures
CTT
map
,RH
map
All but classical freezing rain and
deep convection
See text and figures
Adjusted icing potential Situation for application Equation
Icing potential adjustment using SLW,
PIREP, and VV
All with upward or zero VV Final ⫽ Initial ⫹ (1 ⫺ Initial)(0.4SLW
map
⫹
0.35PIREP
map
⫹ 0.25VV
map
)
Icing potential adjustment using SLW,
PIREP, and VV
All with downward VV Final ⫽ Initial ⫹ [(1 ⫺ Initial)(0.4SLW
map
⫹
0.35PIREP
map
)] ⫹ (Initial ⫻ 0.25VV
map
)
J
ULY 2005 B E R N STEIN ET AL. 979

box at a typical icing temperature (e.g., ⫺8°C) is likely
to find liquid water in portions with warm cloud tops
and mixed-phase or glaciated conditions in portions
with cold cloud tops. An aircraft traversing the grid box
at a higher, colder (e.g., ⫺20°C) level may encounter
clear air in the portions where it is above the warm
clouds and glaciated conditions where it enters deeper
clouds with relatively cold tops.
The icing potential field is designed to find the maxi-
mum potential for supercooled liquid water expected at
each altitude anywhere across the grid box. With a
good chance for icing at low altitudes in some portions
and little chance for icing at higher altitudes across the
entire grid box, CIP treats each altitude differently. For
each altitude, the temperature is compared with the
CTT distribution present within the grid box and an
appropriate CTT and temperature combination is
found. In the example described above, a CTT of
⫺12°C is applied to all altitudes with temperatures
greater than or equal to ⫺12°C. It is in this portion of
the grid box where icing is most likely, because it has
the most ideal combination of temperature and CTT.
Thus, a large icing potential is indicated. At altitudes
with colder temperatures, correspondingly colder CTT
values are used and lower icing potentials result.
4) CLASSICAL FREEZING RAIN
The classical freezing-rain structure consists of a
layer of above-freezing temperatures (the warm layer)
between two layers of subfreezing temperatures (cold
layers). Snow formed in the upper cold layer falls into
the warm layer, where it melts to form rain, and sub-
sequently falls into the lower cold layer to form freezing
rain. Depending on the temperature and depth of these
layers, the resulting precipitation at the surface is typi-
cally freezing rain, ice pellets, or rain (Hanesiak and
Stewart 1995).
Using the model temperature profile, the column is
separated into three layers: above, below, and within
the warm layer. Because no melting has taken place
above the warm layer, any cloud residing there is
treated like a single-layer cloud. Cloud tops are typi-
cally cold (⬍⫺20°C), and so the potential for ice crys-
tals to dominate is large and the resulting icing poten-
tial is often small. If the cloud tops are relatively warm,
the icing potential is larger.
Altitudes beneath the warm layer are treated sepa-
rately, because melting is involved. When liquid pre-
cipitation or ice pellets are observed at the surface in
the presence of this temperature structure, icing and
SLD are likely to be present anywhere in the lower cold
layer, regardless of the relative humidity and cloud-top
temperature. Thus, CIP does not use those two interest
maps to calculate the icing or SLD potentials at these
altitudes. The presence of widespread radar echoes
⬎18 dBZ further suggests the potential for larger and/
or more droplets falling through the lower cold layer.
This information is used to increase both icing and SLD
potentials there (see Table 2). Temperatures within the
warm layer are too warm for icing, and so the icing
potential is zero.
5) DEEP CONVECTION
Icing is one of many hazards associated with deep
convective situations, and CIP uses observations of
lightning from the national lightning network to iden-
tify them. Strong upward motion allows thunderstorms
to produce large amounts of supercooled liquid water
(e.g., Knight and Squires 1982) and, in many cases,
SLD. Icing conditions in convective turrets typically
have small horizontal but large vertical extents from the
freezing level up to unusually cold temperatures
(Rosenfeld and Woodley 2000). Because of this, CIP
uses a special temperature map that allows for icing at
temperatures as cold as ⫺30°C, rather than the ⫺25°C
lower limit employed for other clouds (Fig. 3a). This
map also dramatically increases the interest at tempera-
tures between ⫺30° and ⫺12°C. The lower bound could
be extended to as cold as ⫺40°C to capture the rela-
tively rare icing at these temperatures, but it would also
result in more false alarms. Cloud-top temperature is
not considered to be an important factor in assessing
the icing potential in deep convection, because the
strong lift can allow supercooled liquid water produc-
tion to exceed depletion normally expected with cold
cloud tops. Relative humidity is also not used because
numerical models often underestimate the moisture as-
sociated with subgrid-scale convection.
e. Step 6: Calculate the final icing potential using
boosting factors
In the final step, the situationally derived initial icing
potential (0.0–1.0 scale) is adjusted using recent pilot
reports of icing and model forecasts of vertical velocity
and supercooled liquid water. As described earlier,
these fields are used in manual forecasting to increase
or decrease confidence that icing will be present. Re-
cent reports of icing and forecasts of upward motion
and SLW can all increase the icing potential, while only
a forecast of downward motion can decrease it. The
maximum amount of increase is the difference between
the initial icing potential and unity (e.g., 0.6 for an ini-
tial icing potential of 0.4). SLW
map
, PIREP
map
, and
VV
map
can contribute boosts of as much as 40%, 35%,
and 25% of this value, respectively. When downward
980 JOURNAL OF APPLIED METEOROLOGY VOLUME 44

motion is forecast, the maximum decrease from VV
map
is 25% of the difference between the initial icing po-
tential and zero (see Table 2). These factors are not
applied to the SLD potential field, because they have
not been shown to be well correlated with the presence
of SLD.
4. Examination of four sample cases using
research aircraft data
To demonstrate CIP’s abilities and shortcomings in a
variety of environments, icing and SLD diagnoses from
real-time runs are compared with observations made by
the National Aeronautics and Space Administration
(NASA) Glenn Research Center Twin Otter research
aircraft (Miller et al. 1998) during flights through 1)
freezing rain, 2) a single-layer precipitating cloud, and
3) a cloud with a cloud-top temperature gradient. A
flight through a single-layer, nonprecipitating cloud
that CIP diagnosed poorly is also discussed.
a. Freezing rain
At ⬃2100 UTC 15 February 2003, the Twin Otter
flew through freezing rain over Huntington, West Vir-
ginia (station identifier KHTS). A classical freezing-
rain temperature structure was present in the RUC,
and surface observations indicated freezing rain and
0.2-km ceilings while satellite-measured cloud-top tem-
peratures were near ⫺45°C. CIP indicated high icing
and SLD potentials in the lower cold layer, and zero
and low icing potentials were diagnosed within and
above the warm (above freezing) layer, respectively.
The aircraft encountered freezing rain from the base of
the warm layer (1.1 km) down to the minimum vector-
ing altitude of 0.8 km. Given the temperature structure
that was present and the KHTS surface observation of
freezing rain, SLD is likely to have also existed between
0.8 km and the surface. Rough, bumpy ice formed on
the aircraft, extending back to 20% chord, and a ridge
of ice formed at the trailing edge of the ice protection
boot, leading the pilot to report moderate-to-severe
clear icing. Horizontal and vertical cross sections
through this location show that CIP diagnosed the icing
and SLD conditions well (Fig. 5).
b. Single-layer, precipitating cloud
At ⬃1420 UTC 21 February 2002, the NASA Twin
Otter descended through precipitating clouds north of
Mansfield, Ohio. The satellite-measured cloud-top tem-
perature was ⫺11°C and (above freezing) drizzle was
reported at nearby Cleveland, Ohio. CIP’s initial icing
and SLD potentials were high for this case of collision–
coalescence SLD aloft. RUC predicted slight upward
vertical velocities throughout the layer and SLW from
just above cloud base (0.3 km) to 1.4 km, and recent
PIREPs of icing were found between ⬃0.9 and 2.1 km.
These factors boosted the high initial icing potentials to
near unity. The Twin Otter found a single-layer, all-
water cloud, with drops as large as 300
␮
m in diameter
near 1.6 km. The aircraft accreted mixed icing that ex-
tended beyond the ice protection boots. The cloud-top
temperature measured by the Twin Otter was ⫺9°C,
which was somewhat warmer than that indicated by
satellite. This measurement error, in combination with
RUC temperature forecasts that were too warm near
cloud top, caused CIP to indicate that the icing ex-
tended well above the actual cloud top.
c. Cloud-top temperature gradient
At ⬃1815 UTC 26 February 2002, the Twin Otter
flew across a single RUC model grid box in which
cloud-top temperatures changed from ⫺17° to ⫺22°C.
No precipitation was falling from the warmer clouds at
the eastern end of the grid box, while snow and radar
echoes were present beneath the colder tops at the
western end. As the aircraft traversed the grid box from
east to west at ⬃1.3 km (T ⫽⫺5.5°C level), there was
a clear transition from a mixed-phase icing situation to
a glaciated/icing-free one. As described in section
3d(3), CIP applied the warmest CTT (⫺17°C) to the
⫺5.5°C flight level. This result was combined with at
least one recent PIREP that indicated icing at this level
and RUC predictions of both SLW and upward motion
there, resulting in an icing potential of ⬃0.9. The high
icing potential agrees with Twin Otter observations of
icing in the eastern part of the grid box.
d. Poorly diagnosed, single-layer cloud
At ⬃1922 UTC 6 March 2004, the Twin Otter en-
countered icing conditions between 0.8 and 1.7 km to
the southwest of Cleveland. Peak water contents of 0.6–
0.8gm
⫺3
were found below cloud top at 1.6 km MSL
and ⫺4°C. Satellite data indicated that cloud-top tem-
peratures were near ⫺7°C, resulting in an estimated
cloud top of ⬃2.8 km. Though the top of the layer
sampled on this flight was well below this level, on-
board researchers indicated that a second cloud layer
was present above the sampled clouds. CIP indicated
icing from cloud top to cloud base, with an icing poten-
tial of 0.82 at 1.2 km but only 0.28 at 1.6 km. The
relatively low icing potential resulted from forecasts of
moderate relative humidity (74%), temperatures that
were slightly too warm (⫺2.9°C) in a critical tempera-
JULY 2005 B E R N STEIN ET AL. 981

ture range (between ⫺4° and 0°C; see Fig. 3a), weak
downward motion, and a lack of both model-predicted
SLW and recent PIREPs. Such a case demonstrates the
value of the fuzzy-logic approach. One field (CTT) pro-
vided a good indication of icing, but several others were
less favorable. When presented with conflicting infor-
mation, the CIP technique still indicated that icing was
possible but was relatively unlikely at this level.
FIG. 5. (a) CIP plot for icing potential at 3000 ft (915 m) MSL for 2100 UTC 15 Feb 2003. Color bar shows the
range of icing potentials from 0.05 to 1.00. NASA Twin Otter location is marked with an “x.” (b) CIP flight-route
cross section from Louisville, KY, (SDF) to Washington—Dulles International Airport (IAD) for 2100 UTC 15
Feb 2003. SLD icing conditions were encountered by the NASA Twin Otter near Huntington, WV, (HTS) below
2500 ft (760 m) MSL (marked with an “x”). Color bar shows the range of icing and SLD potentials from 0.05 to
1.00. RUC topography is filled with black. Gray areas are those where icing was diagnosed but there was no direct
indication of SLD [SLD potential is unknown (UKN)].
982 JOURNAL OF APPLIED METEOROLOGY VOLUME 44
Fig 5 live 4/C

5. Verification using PIREPs
Sample cases provide useful demonstrations, but it is
important to assess the overall ability of CIP to capture
icing occurrences and to differentiate between icing and
nonicing environments. To accomplish these goals, 44
376 positive (yes; 7638 with moderate or greater re-
ported severity) and 10 057 negative (no) icing PIREPs
collected during January–March of 2003 are compared
with CIP diagnoses valid at 0000, 0300, 1200, 1500,
1800, and 2100 UTC (late-night hours are excluded be-
cause of a lack of PIREPs). The basic verification ap-
proach is described in Brown et al. (1997) and extended
in Brown et al. (1999, 2001). Probability of detection
(POD) is computed for both yes and no PIREPs, with
the resulting statistics denoted PODy and PODn, re-
spectively. PODy can be interpreted as the proportion
of yes PIREPs that are correctly diagnosed to be in
regions with icing; PODn is the proportion of no-icing
PIREPs that are correctly diagnosed to be in regions
with no icing. Only PIREPs with moderate-or-greater
icing severity are used to compute PODy here, but tests
using PIREPs of all severities yielded very similar re-
sults.
Because CIP values cover a continuous range from
0.0 to 1.0, the verification analyses are based on apply-
ing an array of thresholds to create yes/no icing diag-
noses. That is, a yes diagnosis is inferred at a grid box if
the CIP value equals or exceeds the threshold; a no
diagnosis is inferred if the value at a grid box is less than
the threshold. The relationship between PODy and 1 ⫺
PODn for different algorithm thresholds is the basis for
the verification approach known as the “signal detec-
tion theory” (e.g., Mason 1982). The curve joining the
points for different thresholds is known as the “relative
operating characteristic” (ROC) curve; the area under
this curve is a measure of skill (0.5 indicates no skill). In
the ideal case, the ROC curve will lie above the diago-
nal no-skill line, toward the upper left corner of the
diagram (see Fig. 6).
The verification system compares each PIREP with
the largest icing potential value found within the nearby
grid boxes (⫾1 grid box horizontally and ⫾0.3 km ver-
tically). As described earlier, CIP incorporates PIREPs
from the hour prior to the valid time. Thus, the verifi-
cation analyses only used PIREPs from the hour fol-
lowing the forecast valid time.
The ROC curve indicates that CIP is skillful in dis-
criminating between yes and no PIREPs. In fact, the
area under this curve is ⬃0.78, which is significantly
larger than the no-skill value of 0.50. At a low threshold
(0.05), CIP was able to diagnose correctly 83.4% of the
(moderate or greater) yes and 62.4% of the explicit no
PIREPs, respectively. Statistics were very similar when
all positive icing severities were examined.
6. Summary
CIP combines satellite, surface, radar, lightning, and
pilot observations with model output to provide a de-
tailed three-dimensional hourly diagnosis of the poten-
tial for icing and SLD. It uses a physically based situ-
ational approach derived from basic and applied cloud
physics principles, combined with forecaster and on-
board flight experience from field programs. CIP uses a
conservative approach to the determination of the lo-
cations of clouds and precipitation and in its depiction
of the potential for icing and SLD, showing the worst
possible conditions that are likely to exist within a given
portion of airspace (3D grid volume). The technique
was demonstrated using examples of icing encountered
by a research aircraft, and its quality was validated us-
ing pilot reports for a 3-month period.
CIP provides users with accurate, high-resolution de-
pictions of icing and SLD potential, allowing them to
make route-specific decisions that can help aircraft to
avoid icing, including that associated with SLD. The use
of the operational CIP in combination with high-
resolution diagnoses and forecasts of convection, tur-
bulence, ceiling and visibility, flight-level winds, and so
on in an easy-to-use graphical form may soon allow
pilots and dispatchers to choose their routes and alti-
tudes appropriately to allow for more efficient and,
most important, safer flights.
Acknowledgments. This research is in response to re-
quirements and funding by the Federal Aviation Ad-
FIG. 6. Relationship between PODy [moderate-or-greater-
severity (mog) PIREPs] and 1 – PODn for CIP, using data from
the winter of 2000, with all valid times combined. The thresholds
used (starting in the upper-right corner) are 0.0, 0.02, 0.05, 0.15,
0.25, ...,0.95.
J
ULY 2005 B E R N STEIN ET AL. 983

ministration. The views expressed are those of the au-
thors and do not necessarily represent the official policy
or position of the FAA.
The authors appreciate the thoughtful comments
supplied by reviewers, including Alexi Korolev and
Roy Rasmussen. Engineers, pilots, and others at the
NASA Glenn Research Center provided invaluable in-
flight research data and the opportunity to gain expe-
rience by flying on board their aircraft. Thanks are
given to regional airlines Air Wisconsin, SkyWest,
COMAIR, Peninsula Airways, and Atlantic Coast Air-
lines for serving as test users for CIP, providing valu-
able feedback and participating in algorithm assess-
ments. Greg Thompson and Mike Dixon provided sig-
nificant help with real-time ingest of satellite and radar
data. Gary Cunning was instrumental in the develop-
ment of the high-resolution and operational versions of
CIP. The verification group at NCAR and NOAA
Forecast Systems Laboratory and the weather assess-
ment group at the FAA Technical Center made high-
quality algorithm verifications and evaluations. Last,
we thank the FAA’s Aviation Weather Research Pro-
gram managers for their strong support of the in-flight
icing program at NCAR, the development of the CIP,
and its subsequent passage into the operational envi-
ronment.
APPENDIX
Abbreviations
ATR Avions de Transport Régional
CIP Current icing potential algorithm
CTT Cloud-top temperature
CTT
map
Cloud-top temperature interest map
DZ Drizzle
EMB Embraer (Empresa Brasileira de Aero-
náutica S.A.)
FAA Federal Aviation Administration
FZDZ Freezing drizzle
FZRA Freezing rain
GOES Geostationary Operational Environmental
Satellites
METAR Aviation Routine Weather Report
NASA National Aeronautics and Space Adminis-
tration
NCAR National Center for Atmospheric Re-
search
NEXRAD Next-Generation Weather Radar
NTSB National Transportation Safety Board
NWS National Weather Service
PE Ice pellets
PIREP Pilot report
PIREP
map
Pilot-report interest map
POD Probability of detection (suffixes “y” and
“n” indicate “yes” and “no,” respectively)
RA Rain
RH Relative humidity with respect to water
RH
map
Relative humidity interest map
ROC Relative operating characteristic
RUC Rapid Update Cycle numerical weather
prediction model
SLD Supercooled large droplets
SLW Supercooled liquid water
SLW
map
Supercooled liquid water interest map
SN Snow
T Temperature
T
map
Temperature interest map
UTC Universal coordinated time
VV Vertical velocity
VV
map
Vertical velocity interest map
REFERENCES
Air Weather Service, 1980: Forecasters’ guide on aircraft icing.
Air Weather Service Rep. AWS/TR-80/001, 58 pp. [Available
from U.S. Air Force, Scott AFB, IL 62225.]
Benjamin, S. G., and Coauthors, 2004: An hourly assimilation–
forecast cycle: The RUC. Mon. Wea. Rev., 132, 495–518.
Bernstein, B. C., 1996: A new technique for identifying locations
where supercooled large droplets are likely to exist: The
stovepipe algorithm. Preprints, 15th Conf. on Weather Analy-
sis and Forecasting, Norfolk, VA, Amer. Meteor. Soc., 5–8.
——, T. A. Omeron, F. McDonough, and M. K. Politovich, 1997:
The relationship between aircraft icing and synoptic scale
weather conditions. Wea. Forecasting, 12, 742–762.
Brown, B. G., G. Thompson, R. T. Bruintjes, R. Bullock, and T.
Kane, 1997: Intercomparison of in-flight icing algorithms.
Part II: Statistical verification results. Wea. Forecasting, 12,
890–914.
——, T. L. Kane, R. Bullock, and M. K. Politovich, 1999: Evi-
dence of improvements in the quality of in-flight icing algo-
rithms. Preprints, Eighth Conf. on Aviation, Range, and
Aerospace Meteorology, Dallas, TX, Amer. Meteor. Soc., 48–
52.
——, J. L. Mahoney, R. Bullock, T. L. Fowler, J. Henderson, and
A. Loughe, 2001: Quality assessment report: Integrated Icing
Diagnostic Algorithm (IIDA). AWIT Technology Review
Panel Internal Report, 35 pp. [Available from B. Brown, Re-
search Applications Program, National Center for Atmo-
spheric Research, P.O. Box 3000, Boulder, CO 80307-3000.]
Carriere, J., S. Alquier, C. LeBot, and E. Moulin, 1997: Some
results of a statistical evaluation of forecast icing risk algo-
rithms. Meteor. Appl., 4, 115–130.
Cober, S. G., G. A. Isaac, and J. W. Strapp, 1995: Aircraft icing
measurements in East Coast winter storms. J. Appl. Meteor.,
34, 88–100.
——, J. W. Strapp, and G. A. Isaac, 1996: An example of super-
cooled drizzle drops formed through a collision–coalescence
process. J. Appl. Meteor., 35, 2250–2260.
——, G. A. Isaac, and J. W. Strapp, 2001: Characterizations of
aircraft icing environments that include supercooled large
drops. J. Appl. Meteor., 40, 1984–2002.
984 JOURNAL OF APPLIED METEOROLOGY VOLUME 44

Ellrod, G. P., 1996: The use of GOES-8 multispectral imagery for
the detection of aircraft icing regions. Preprints, Eighth Conf.
on Satellite Meteorology and Oceanography, Atlanta, GA,
Amer. Meteor. Soc., 168–171.
Forbes, G. S., Y. Hu, B. G. Brown, B. C. Bernstein, and M. K.
Politovich, 1993: Examination of soundings in the proximity
of pilot reports of aircraft icing during STORM-FEST. Pre-
prints, Fifth Conf. on Aviation Weather Systems, Vienna, VA,
Amer. Meteor. Soc., 282–286.
Geresdi, I., R. Rasmussen, W. Grabowski, and B. Bernstein, 2005:
Sensitivity of freezing drizzle formation in stably stratified
clouds to ice processes. Meteor. Atmos. Phys., 88, 91–105.
Guan, H., S. G. Cober, and G. A. Isaac, 2001: Verification of su-
percooled cloud water forecasts with in situ aircraft measure-
ments. Wea. Forecasting, 16, 145–155.
Hanesiak, J. M., and R. E. Stewart, 1995: The mesoscale and mi-
croscale structure of a severe ice pellet storm. Mon. Wea.
Rev., 123, 3144–3162.
Hill, G. E., 1980: Seeding-opportunity recognition in winter oro-
graphic clouds. J. Appl. Meteor., 19, 1371–1381.
Isaac, G. A., S. G. Cober, J. W. Strapp, A. V. Korolev, A. Trem-
blay, and D. L. Marcotte, 2001: Recent Canadian research on
in-flight aircraft icing. Can. Aeronaut. Space J., 47, 1–9.
Jensen, C. E., 1963: Prediction of large-scale cloudiness and air-
frame icing conditions by machine methods. J. Appl. Meteor.,
2, 337–344.
Kelsch, M., and L. Wharton, 1996: Comparing PIREPs with
NAWAU turbulence and icing forecasts: Issues and results.
Wea. Forecasting, 11, 385–390.
Knight, C. A., and P. Squires, Eds., 1982: Case Studies of the Na-
tional Hail Research Experiment. Vol. 2, Hailstorms of the
Central High Plains, Colorado Associated University Press,
245 pp.
Korolev, A. V., G. A. Isaac, S. G. Cober, J. W. Strapp, and J.
Hallett, 2003: Microphysical characterization of mixed-phase
clouds. Quart. J. Roy. Meteor. Soc., 129, 39–65.
Le Bot, C., 2004: SIGMA: System of Icing Geographic Identifi-
cation in Meteorology for Aviation. Preprints, 11th Conf. on
Aviation, Range and Aerospace Meteorology, Hyannis, MA,
Amer. Meteor. Soc., CD-ROM.
Lee, T. F., F. J. Turk, and K. Richardson, 1997: Stratus and fog
products using GOES-8–9 3.9-m data. Wea. Forecasting, 12,
664–677.
Marwitz, J. D., M. K. Politovich, B. C. Bernstein, F. M. Ralph,
P. J. Neiman, R. Ashenden, and J. Bresch, 1997: Meteoro-
logical conditions associated with the ATR-72 aircraft acci-
dent near Roselawn, Indiana, on 31 October 1994. Bull.
Amer. Meteor. Soc., 78, 41–52.
Mason, I. B., 1982: A model for assessment of weather forecasts.
Aust. Meteor. Mag., 30, 291–303.
Miller, D., T. Ratvasky, B. Bernstein, F. McDonough, and J. W.
Strapp, 1998: NASA/FAA/NCAR supercooled large droplet
icing flight research: Summary of winter 96-97 flight opera-
tions. Extended Abstracts, 36th Aerospace Science Meeting
and Exhibit, Reno, NV, American Institute of Aeronautics
and Astronautics, AIAA 98-0557, 20 pp.
National Transportation Safety Board, 1996: Aircraft accident re-
port. In-flight icing encounter and loss of control, Simmons
Airlines, d.b.a. American Eagle flight 4184, Avions de Trans-
port Regional (ATR) Model 72-212, N401MA, Roselawn,
IN, October 31 1994, Vol. 1, Safety Board Rep. NTSB/AAR-
96/01, PB96-910401, DCA95MA001, 322 pp. [Available from
National Transportation Safety Board, Public Inquiries
Section RE-51, 490 L’Enfant Plaza S.W., Washington, DC
20594.]
——, 1998: Aircraft accident report. In-flight icing encounter and
uncontrolled collision with terrain, COMAIR flight 3272,
Embraer EMB-120RT, N265 CA, Monroe, MI, January 9,
1997, Safety Board Rep. NTSB/AAR-98/04, PB08-910404,
DCA97MA017, 348 pp. [Available from National Transpor-
tation Safety Board, Public Inquiries Section RE-51, 490
L’Enfant Plaza S.W., Washington, DC 20594.]
Pobanz, B. M., J. D. Marwitz, and M. K. Politovich, 1994: Condi-
tions associated with large-drop regions. J. Appl. Meteor., 33,
1366–1372.
Politovich, M. K., and B. C. Bernstein, 1995: Production and
depletion of supercooled liquid water in a Colorado winter
storm. J. Appl. Meteor., 34, 2631–2648.
Rasmussen, R., and Coauthors, 1992: Winter Icing and Storms
Project (WISP). Bull. Amer. Meteor. Soc., 73, 951–974.
——, B. C. Bernstein, M. Murakami, G. Stossmeister, J. Reisner,
and B. Stankov, 1995: The 1990 Valentine’s Day Arctic out-
break. Part I: Mesoscale and microscale structure and evolu-
tion of a Colorado Front Range shallow upslope cloud.
J. Appl. Meteor., 34, 1481–1511.
Rauber, R. M., L. S. Olthoff, M. K. Ramamurthy, and K. E.
Kunkel, 2000: The relative importance of warm rain and
melting processes in freezing precipitation events. J. Appl.
Meteor., 39, 1185–1195.
Reinking, R. F., S. Y. Matrosov, R. T. Bruintjes, and B. E. Mart-
ner, 1997: Identification of hydrometeors with elliptical and lin-
ear polarization Ka-band radar. J. Appl. Meteor., 36, 322–339.
Reisner, J., R. M. Rasmussen, and R. T. Bruintjes, 1998: Explicit
forecasting of supercooled liquid water in winter storms using
the MM5 mesoscale model. Quart. J. Roy. Meteor. Soc., 124,
1071–1107.
Rogers, R. R., and M. K. Yau, 1989: A Short Course in Cloud
Physics. 3d ed. Pergamon Press, 151 pp.
Rosenfeld, D., and W. L. Woodley, 2000: Deep convective clouds
with sustained supercooled liquid water down to ⫺37.5°C.
Nature, 405, 440–442.
Ryerson, C., M. K. Politovich, K. L. Rancourt, G. G. Koenig,
R. F. Reinking, and D. R. Miller, 2000: Overview of Mt.
Washington icing sensors project. Extended Abstracts, 38th
AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV,
American Institute of Aeronautics and Astronautics, Paper
AIAA-2000-0488.
Sand, W. R., W. A. Cooper, M. K. Politovich, and D. L. Veal,
1984: Icing conditions encountered by a research aircraft. J.
Climate Appl. Meteor., 23, 1427–1440.
Schultz, P., and M. K. Politovich, 1992: Toward the improvement
of aircraft-icing forecasts for the continental United States.
Wea. Forecasting, 7, 491–500.
Smith, W. L., Jr., P. Minnis, B. Bernstein, and P. W. Heck, 2002:
Supercooled liquid water cloud properties derived from
GOES: Comparisons with in-situ aircraft measurements. Pre-
prints, 10th Conf. on Aviation, Range and Aerospace Meteo-
rology, Portland, OR, Amer. Meteor. Soc., 89–92.
Tafferner, A., T. Hauf, C. Leifeld, T. Hafner, H. Leykauf, and U.
Voigt, 2003: ADWICE: Advanced Diagnosis and Warning
System for Aircraft Icing Environments. Wea. Forecasting,
18, 184–203.
Thompson, G., R. T. Bruintjes, B. G. Brown, and F. Hage, 1997a:
Intercomparison of in-flight icing algorithms. Part I: WISP94
real-time icing prediction and evaluation program. Wea.
Forecasting, 12, 878–889.
J
ULY 2005 B E R N STEIN ET AL. 985

——, R. Bullock, and T. F. Lee, 1997b: Using satellite data to
reduce spatial extent of diagnosed icing. Wea. Forecasting, 12,
185–190.
——, R. M. Rasmussen, and K. Manning, 2004: Explicit forecasts
of winter precipitation using an improved bulk microphysics
scheme. Part I: Description and sensitivity analysis. Mon.
Wea. Rev., 132, 519–542.
Tremblay, A., and A. Glazer, 2000: An improved modeling
scheme for freezing precipitation forecasts. Mon. Wea. Rev.,
128, 1289–1308.
Turk, J., J. Vivekanandan, T. Lee, P. Durkee, and K. Nielsen,
1998: Derivation and applications of near-infrared cloud re-
flectances from GOES-8 and GOES-9. J. Appl. Meteor., 37,
819–831.
Vivekanandan, J., S. M. Ellis, R. Oye, D. S. Zrnic, A. V. Ryzhkov,
and J. Straka, 1999: Cloud microphysics retrieval using S-
band dual-polarization radar measurements. Bull. Amer. Me-
teor. Soc., 80, 381–388.
Wade, C. G., 2003: A multi-sensor approach to detecting drizzle
on ASOS. J. Atmos. Oceanic Technol., 20, 820–832.
986 JOURNAL OF APPLIED METEOROLOGY VOLUME 44