The Atmospheric River Tracking Method Intercomparison Project (ARTMIP): Quantifying Uncertainties in Atmospheric River Climatology

Abstract

Atmospheric rivers (ARs) are now widely known for their association with high‐impact weather events and long‐term water supply in many regions. Researchers within the scientific community have developed numerous methods to identify and track of ARs—a necessary step for analyses on gridded data sets, and objective attribution of impacts to ARs. These different methods have been developed to answer specific research questions and hence use different criteria (e.g., geometry, threshold values of key variables, and time dependence). Furthermore, these methods are often employed using different reanalysis data sets, time periods, and regions of interest. The goal of the Atmospheric River Tracking Method Intercomparison Project (ARTMIP) is to understand and quantify uncertainties in AR science that arise due to differences in these methods. This paper presents results for key AR‐related metrics based on 20+ different AR identification and tracking methods applied to Modern‐Era Retrospective Analysis for Research and Applications Version 2 reanalysis data from January 1980 through June 2017. We show that AR frequency, duration, and seasonality exhibit a wide range of results, while the meridional distribution of these metrics along selected coastal (but not interior) transects are quite similar across methods. Furthermore, methods are grouped into criteria‐based clusters, within which the range of results is reduced. AR case studies and an evaluation of individual method deviation from an all‐method mean highlight advantages/disadvantages of certain approaches. For example, methods with less (more) restrictive criteria identify more (less) ARs and AR‐related impacts. Finally, this paper concludes with a discussion and recommendations for those conducting AR‐related research to consider.

Full text

The Atmospheric River Tracking Method Intercomparison
Project (ARTMIP): Quantifying Uncertainties
in Atmospheric River Climatology
Jonathan J. Rutz
1
, Christine A. Shields
2
, Juan M. Lora
3
, Ashley E. Payne
4
, Bin Guan
5
,
Paul Ullrich
6
, Travis O’Brien
7
, L. Ruby Leung
8
, F. Martin Ralph
9
, Michael Wehner
10
,
Swen Brands
11
, Allison Collow
12
, Naomi Goldenson
13
, Irina Gorodetskaya
14
,
Helen Griffith
15
, Karthik Kashinath
16
, Brian Kawzenuk
9
, Harinarayan Krishnan
10
,
Vitaliy Kurlin
17
, David Lavers
18
, Gudrun Magnusdottir
19
, Kelly Mahoney
20
,
Elizabeth McClenny
6
, Grzegorz Muszynski
16,17
, Phu Dinh Nguyen
21
, Mr. Prabhat
16
,
Yun Qian
8
, Alexandre M. Ramos
22
, Chandan Sarangi
8
, Scott Sellars
23
, T. Shulgina
9
,
Ricardo Tome
22
, Duane Waliser
5
, Daniel Walton
24
, Gary Wick
19
, Anna M. Wilson
9
,
and Maximiliano Viale
25
1
Science and Technology Infusion Division, National Weather Service Western Region Headquarters, National Oceanic
and Atmospheric Administration, Salt Lake City, UT, USA,
2
Climate and Global Dynamics Division, National Center for
Atmospheric Research, Boulder, CO, USA,
3
Department of Geology and Geophysics, Yale University, New Haven, CT,
USA,
4
Department of Climate and Space Sciences and Engineering, University of Michigan, Ann Arbor, MI, USA,
5
Jet
Propulsion Laboratory, California Institute of Technology, Pasadena, CA, USA,
6
Department of Land, Air and Water
Resources, University of California, Davis, CA, USA,
7
Climate and Ecosystem Sciences Division, Lawrence Berkeley
National Laboratory, Berkeley, CA, USA,
8
Earth Systems Analysis and Modeling, Pacific Northwest National Laboratory,
Richland, WA, USA,
9
Center for Western Weather and Water Extremes, Scripps Institution of Oceanography, La Jolla,
CA, USA,
10
Computational Chemistry, Materials, and Climate Group, Lawrence Berkeley National Laboratory, Berkeley,
CA, USA,
11
MeteoGalicia–Xunta de Galicia, Santiago de Compostela, Spain,
12
Universities Space Research Association,
Columbia, MD, USA,
13
Center for Climate Science, University of California, Los Angeles, CA, USA,
14
Centre for
Environmental and Marine Studies, University of Aveiro, Aveiro, Portugal,
15
Department of Geography and
Environmental Science, University of Reading, Reading, UK,
16
Data and Analytics Services, National Energy Research
Scientific Computing Center (NERSC), Lawrence Bekeley National Laboratory, Berkeley, CA, USA,
17
Department of
Computer Science, University of Liverpool, Liverpool, UK,
18
European Centre for Medium‐Range Weather Forecasts,
Reading, UK,
19
Department of Earth System Science, University of California, Irvine, CA, USA,
20
Physical Sciences
Division, Earth System Research Laboratory, National Oceanic and Atmospheric Administration, Boulder, CO, USA,
21
Department of Civil and Environmental Engineering, University of California, Irvine, CA, USA,
22
Instituto Dom Luiz,
Faculdade de Ciências, Universidade de Lisboa, Lisbon, Portugal,
23
University of California, San Diego Qualcomm
Institute, La Jolla, California, USA,
24
Institute of the Environment and Sustainability, University of California, Los
Angeles, CA, USA,
25
Instituto Argentino de Nivologia, Glaciologia y Ciencias Ambientales, CCT‐CONICET, Mendoza,
Argentina
Abstract Atmospheric rivers (ARs) are now widely known for their association with high‐impact
weather events and long‐term water supply in many regions. Researchers within the scientific community
have developed numerous methods to identify and track of ARs—a necessary step for analyses on gridded
data sets, and objective attribution of impacts to ARs. These different methods have been developed to
answer specific research questions and hence use different criteria (e.g., geometry, threshold values of key
variables, and time dependence). Furthermore, these methods are often employed using different reanalysis
data sets, time periods, and regions of interest. The goal of the Atmospheric River Tracking Method
Intercomparison Project (ARTMIP) is to understand and quantify uncertainties in AR science that arise due
to differences in these methods. This paper presents results for key AR‐related metrics based on 20+
different AR identification and tracking methods applied to Modern‐Era Retrospective Analysis for Research
and Applications Version 2 reanalysis data from January 1980 through June 2017. We show that AR
frequency, duration, and seasonality exhibit a wide range of results, while the meridional distribution of
these metrics along selected coastal (but not interior) transects are quite similar across methods.
Furthermore, methods are grouped into criteria‐based clusters, within which the range of results is reduced.
AR case studies and an evaluation of individual method deviation from an all‐method mean highlight
©2019. American Geophysical Union.
All Rights Reserved. This article has
been contributed to by US Government
employees and their work is in the
public domain in the USA.
RESEARCH ARTICLE
10.1029/2019JD030936
Special Section:
Atmospheric Rivers:
Intersection of Weather and
Climate
Key Points:
•The large number of atmospheric
river identification/tracking
methods produces large uncertainty
related to AR climatology and
impacts
•Uncertainty is quantified using the
same data (MERRA v2), time period
(1980–2017), region (global where
possible), and common metrics
•This study presents
recommendations regarding the
advantages/disadvantages of certain
approaches based on science
application
Supporting Information:
•Supporting Information S1
Correspondence to:
J. J. Rutz,
jonathan.rutz@noaa.gov
Citation:
Rutz, J. J., Shields, C. A., Lora, J. M.,
Payne, A. E., Guan, B., Ullrich, P., et al.
(2019). The atmospheric river tracking
method intercomparison project
(ARTMIP): quantifying uncertainties in
atmospheric river climatology. Journal
of Geophysical Research: Atmospheres,
10.1029/2019JD030936
Received 2 MAY 2019
Accepted 21 NOV 2019
Accepted article online 24 NOV 2019
RUTZ ET AL.
Published online 24 DEC 2019
124, 13,777–13,802. https://doi.org/
13,777

advantages/disadvantages of certain approaches. For example, methods with less (more) restrictive criteria
identify more (less) ARs and AR‐related impacts. Finally, this paper concludes with a discussion and
recommendations for those conducting AR‐related research to consider.
1. Introduction
Over the past several years, interest in atmospheric river (AR) science and applications has increased rapidly.
Beyond the now well‐known impacts of heavy rain and flooding (e.g., Lamjiri et al., 2017; Neiman et al.,
2008; Ralph et al., 2013), ARs have been shown to have applications in areas as diverse as avalanche hazard
(Hatchett et al., 2017), dust transport (Ault et al., 2011), and postfire debris flows (Oakley et al., 2017).
Furthermore, the study of ARs has become global in scope, and international in terms of participation, as
evidenced by the well‐attended 2018 International Atmospheric Rivers Conference (Ramos et al., 2019).
The American Meteorological Society (AMS) Glossary of Meteorology defines an AR as “a long, narrow,
and transient corridor of strong horizontal water vapor transport that is typically associated with a low‐level
jet stream ahead of the cold front of an extratropical cyclone.”The development of this definition, a process
described by Ralph et al. (2018a), was marked by open engagement with the atmospheric and geosciences
community throughout the process and should be considered a major success in the field. However, the ele-
gance of this definition depends on its qualitative description of ARs, whereas, in practice, the peer‐reviewed
literature contains dozens of quantitative definitions of ARs, as needed in analysis and modeling. These
quantitative definitions are manifested as different AR identification and tracking methods that researchers
have developed to answer a wide variety of questions. Note also that the large majority of these methods
were developed prior to the development of the AR definition within the AMS Glossary of Meteorology.
Each individual method identifies and/or tracks ARs on the basis of selected criteria being met, as summar-
ized in Figure 1. A first step in development of these methods is often the choice of a thresholding variable
and magnitude, which serves as the minimum requirement for identifying ARs. The thresholding variable
can be integrated water vapor (IWV; e.g., Wick et al., 2013) but is most commonly IWV transport (IVT),
and the magnitude can be either absolute (e.g., IVT ≥250 kg m
−1
s
−1
; e.g., Rutz et al., 2014) or relative
(e.g., IVT ≥85th percentile of local climatological IVT; e.g., Lavers et al., 2012). Research has shown that
using IVT extends medium‐range predictability for high‐impact hydrological events (Lavers et al., 2017),
and recent field campaigns have used probabilistic IVT forecasts to determine AR location and intensity
(Cordeira et al., 2017). Once the thresholding process is applied to the data, features meeting or exceeding
the threshold are examined with respect to geometric parameters such as length, width, shape, axis, and
orientation. Throughout this study, methods with lower‐magnitude thresholds and less geometric require-
ments will generally be referred to as “less restrictive methods,”whereas methods with higher‐magnitude
thresholds and more geometric requirements will generally be referred to as “more restrictive methods.”
Note also that some methods, particularly those based on machine learning techniques, do not directly
use any thresholds as requirements. Temporal requirements may also be chosen (i.e., either AR identifica-
tion is independent of time [time slicing], or it is dependent on criteria being met for a certain duration [time
stitching]). The choices described above lead to many possible permutations, and while some methods fea-
ture similar criteria, others vary widely. Of course, in addition to using different identification and tracking
methods, many researchers examine different regions, using different data sets, and different periods of
records, to do so. More recently, machine learning techniques have been developed to identify and track ARs
(e.g., Mudigonda et al., 2017; Muszynski et al., 2019; Radićet al., 2015).
These different methods produce differences in AR climatologies and, consequently, differences in the
impacts attributable to ARs. These differences produce uncertainty in operational weather research and
forecasting, water management, and climate projections, which require a current baseline of AR climatology
and impacts to assess future changes. The differences in identified ARs that can be observed during a single
event are highlighted using a case from 0000 UTC 15 February 2014, shown in Figure 2. Notice that some
methods identify an AR only over the greatest values of IVT offshore, others extend near just inland of
the coast, and some extend well into the continental interior. These differences have major consequences.
For example, one question the water management community might ask is, “what fraction of precipitation
is attributable to ARs, and how might that change under future climate change scenarios?”Before even
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,778

exploring climate change scenarios, one needs to answer the first question, and the answer depends on
which method is chosen. Figure 3 shows the fraction of cool‐season or annual precipitation attributable
to ARs based on three studies (Dettinger et al., 2011; Guan & Waliser, 2015; Rutz et al., 2014). These
studies found broadly similar spatial patterns, but quite widely varying values from southern California
(~15–35%) to coastal Washington (~25–60%). It is worth noting that in addition to different AR identification
methods, these studies also used different data sets, different periods of record, and different methods of
attributing precipitation to ARs, all of which contribute to this range.
It is critical to remember that each AR identification and tracking method was developed to answer a specific
question or set of questions and that these questions vary widely from one study to the next. Having a sense
of these original questions better informs the reader as to the original intent or goal of each method, as
described in the supplemental material provided by method developers. For example, Ramos et al. (2015)
examined the relationship between persistent ARs and extreme precipitation over the Iberian Peninsula;
Rutz et al. (2014) identified ARs and their impacts over the complex topography of the western United
States; and Guan and Waliser (2015) produced a global climatology of ARs and their characteristics.
Furthermore, Shields and Kiehl (2016a, 2016b) and Gershunov et al. (2017) explored the climate scale varia-
bility of ARs along the North American West Coast. Still other methods are using machine learning techni-
ques to determine whether ARs can be identified without the use of defined thresholds (e.g., Muszynski
et al., 2019). With such a variety of different questions asked, and such different goals pursued, it should
not be surprising that many different results have been found. Nevertheless, a growing awareness of the
uncertainties that these differences produce has led to the development of a community‐based project to bet-
ter understand and quantify them.
The goal of the Atmospheric River Tracking Method Intercomparison Project (ARTMIP; Shields et al., 2018)
is to quantify and understand the uncertainties in AR climatology (e.g., frequency, duration, and intensity),
precipitation, and related impacts that arise from different AR identification and tracking methods, and how
uncertainties in these AR‐related metrics may change in the future. Furthermore, ARTMIP aims to under-
stand the implications of those uncertainties in terms of our recent, current, and future climate. A few recent
studies have focused on this topic. Huning et al. (2017) examined the sensitivity of AR‐attributable snowfall
in California's Sierra Nevada to AR detection methods based on two different AR catalogs. Guan and Waliser
(2015) examined the sensitivity of AR detection to intensity/geometry thresholds and input data sets, but
only based on a single AR detection algorithm. Ralph et al. (2018b), in an initial pre‐ARTMIP study, quan-
tified uncertainties in AR‐related metrics using ~10 AR detection algorithms but focused on only one loca-
tion along the California coast. This paper provides a systematic and global intercomparison between
Figure 1. Schematic diagram illustrating the diversity of AR identification and tracking methods found in current litera-
ture by categorizing the variety of parameters used as criteria and then listing different types of choices available per
category.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,779

different methods by quantifying the uncertainty in current (1980–2017) AR climatology on a global scale,
using over 20 AR identification and tracking methods. To do so, it leverages a variety of metrics, which
are described in more detail in the following sections. An assessment of method‐related uncertainty
affecting AR climatology under climate change scenarios will be the subject of another paper, discussed at
the end of section 4.
2. Data and Methods
The progression of ARTMIP is organized into “tiers,”and this study is a summary of results from the Tier 1
phase of the project. The data used in Tier 1 of ARTMIP are described at length in Shields et al. (2018), and a
brief overview is also given here.
A key aspect of ARTMIP is that analyses are performed using the same atmospheric data set, over the same
period of record, and over the entire globe. This enables a clean comparison of AR‐related metrics across all
methods, whereas previous studies used different atmospheric data sets, different periods of records and
examined only certain regions. Note, however, that some methods' criteria explicitly limit their results to cer-
tain regions, and a mask is used to indicate these regions. Basic quantities such as IWV and IVT, which is
often a derived variable, were precomputed for ARTMIP to ensure that all algorithms use exactly the same
data. The atmospheric data for these calculations comes from the Modern‐Era Retrospective Analysis for
Research and Applications Version 2 (MERRA‐2) reanalysis (Gelaro et al., 2017) for the period of January
1980 through June 2017, at a horizontal resolution of 0.625 × 0.5° and a 3‐hr temporal resolution. The
ARTMIP catalogs are then produced by developers applying their identification and tracking methods to
these data. For each 3‐hr time slice, each grid point is flagged with a 0 for “AR conditions do not exist”or
a 1 for “AR conditions exist.”Catalogues produced for Tier 1 as well as the source MERRA‐2 data used by
all ARTMIP participants are available on the Climate Data Gateway. MERRA‐2 source data can be found
at https://doi.org/10.5065/D62R3QFS (NCAR/UCAR Climate Data Gateway), and ARTMIP Tier 1 output
data catalogues, also housed on the Climate Data Gateway, online (doi:10.5065/D6R78D1M). Table 1 sum-
marizes all the methods participating in ARTMIP with notation specifying Tier 1 algorithms only.
Key results are presented along selected, roughly meridional transects along the North American West
Coast, through interior western North America, and along the European West Coast (Figure 4). These trans-
ects are selected because most regional methods have been developed, and produce data, for one of these two
regions. The coastal transect points are determined by selecting all MERRA‐2 reanalysis grid points that
Figure 2. Example of how AR identification and tracking methods differ over the northeastern Pacific, based on MERRA
Version 2 data from 0000 UTC 15 February 2014. Gray shading represents IVT (kg m
−1
s
−1
), and colored contours
represent the spatial regions designated as ARs by the various methods. Note that only algorithms available in this region
are shown.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,780

have fractional land/sea cover between 32°N and 55°N (and 130–115°W) for North America, and between
35°N and 62°N (and 15°W to 10°E) for Europe. The interior transect points are determined by subjectively
selecting grid points that represent a significant topographic “crest”between 32°N and 55°N. The interior
transect facilitates comparison between results for AR‐related metrics along a coastline, which lies down-
stream of an ocean, and results for AR‐related metrics over an interior region, which lies downstream of,
and embedded within, complex topography.
This paper will present a number of results based on grouping methods into “clusters”that have similar
approaches to identifying ARs. Throughout this section, refer to Figure 5 for a summary of which clusters
each method is grouped into, and Table 1 for more in‐depth information regarding each method. Note also
that many groups have joined ARTMIP and contributed data since the beginning of this analysis and are not
listed here but can be found online. Their areas of focus include South America (Viale et al., 2018) and Polar
regions (Gorodetskaya et al., 2014), among others. The first key cluster pair is that differentiating between
methods using absolute thresholds (e.g., IVT ≥250 kg m
−1
s
−1
) and methods using relative thresholds
(e.g., IVT ≥85th percentile of climatological IVT). This is done because these are fundamentally different
ways of identifying and tracking ARs, and the visualization of results benefits from the distinction.
Throughout this paper, the terms absolute and relative will be italicized when used in this context. There
is also a subtle, but important difference among the relative methods themselves: those whose thresholds
vary as a function of latitude (latitude‐dependent relative methods) and those that do not (latitude‐indepen-
dent relative methods). Latitude‐dependent relative methods use thresholds based on the climatology of each
grid point and can be expected to produce smaller meridional gradients in AR statistics. Latitude‐indepen-
dent relative methods use one threshold based on the climatology of a given region and can be expected to
produce larger gradients in AR statistics, which will likely be more similar to results produced by absolute
methods. Furthermore, this paper includes one method based on machine learning (TDA_ML; Muszynski
et al., 2019), which defies many of the threshold‐based groupings outlined above. It is currently employed
over the western Unites States but could be readily applied to other regions.
Another distinction made in this study, and key cluster pair, is that between global and regional meth-
ods, which simply describes the area over which the method was originally developed and applied.
Masks for each regional method are found in the supplemental material of the experimental design
paper, Shields et al. (2018).
Finally, a subjective distinction will at times be made between methods that are either less restrictive or
more restrictive. Here, “less restrictive”generally denotes a method or methods with less restrictive
Figure 3. Fraction of total cool‐season precipitation attributable to ARs from (a) Dettinger et al. (2011) and (b) Rutz et al.
(2014). (c) As in panels (a) and (b) but for annual precipitation from Guan and Waliser (2015). These studies use different
AR identification methods, as well as different atmospheric reanalyses, observed precipitation data sets, and methods of
attributable precipitation to ARs.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,781

Table 1
Listed Algorithms (A1, A2, etc.) Have Submitted Catalogues for Tier 1
Developer (algorithm name) Type Geometry requirements Threshold requirements
Temporal
requirements Region DOI/reference
Brands et al. (Brands_v1,
Brands_v2, Brands_v3)
Condition >1,500 km Both latitude‐dependent relative and
absolute IVT: 95th (A16)/90th (A23)/
90th (A24)/percentile at point of
detection, 90th/85th/85th percentile
along the AR structure, consider all
months for threshold calculation
with a min threshold of 240 kg m
−1
s
−1
/500 kg m
−1
s
−1
/250 kg m
−1
s
−
1.
Spatial tracking guided by vector
IVT.
Time slice 150°W to 30°E,
30°N to 62°N
10.1007/s00382‐016‐
3095‐6
Gershunov et al.
b
(Gershunov)
Condition and track ≥1,500 km long Absolute: 250 kg m
−1
s
−1
IVT1.5 cm
IWV
Time
stitching‐18
hr (three
time steps for
6‐hourly
data)
Western United
States
10.1002/2017GL074175
Goldenson
b
(Goldenson) Condition >2000 km long and < 1000
km wide, Object
recognition
Absolute:2 cm IWV Time slice Western United
States
Goldenson et. al.,
submitted
Guan and Waliser
b
,
c
(Guan_Waliser)
Condition Length > 2000 km and
length‐width ratio > 2;
Coherent IVT direction
within 45
0
of AR shape
orientation and with a
poleward component
Latitude‐dependent Relative: 85th
percentile IVT; Absolute min
requirement designed for polar
locations: 100kgm
−1
s
−1
IVT
Time slice Global 10.1002/2015JD024257;
10.1175/JHM‐D‐17‐
0114.1
Hagos et al.
b
(PNNL1_Hagos)
Condition Dependent on threshold
requirements to determine
footprint;> 2000 km long
and < 1000 km wide
Absolute: 2 cm IWV 10 ms
−1
wind
speed
Time slice Western United
States
10.1175/JCLI‐D‐14‐
00567.1
Lavers et al. (Lavers) Condition 4.5
o
latitude movement
allowed
Latitude‐independent Relative:
~85th percentile determined by
evaluation of reanalysis products
Time slice United Kingdom, 10.1029/2012JD018027
Leung and
Qian
b
(PNNL2_LQ)
Track Moisture flux has an
eastward or northward
component at landfall;
tracks originating north of
25 N and east of 140 W are
rejected
Absolute: mean IVT along track
>500 kgm
−1
s
−1
and IVT at landfall
>200 kg m
−1
s
−1
; grid points up to
500 km to the north and south along
the AR tracks are included as part of
the AR if their mean IVT > 300 kg m
−1
s
−1
Time slice Western United
States
10.1029/2008GL036445
Lora et al.
b
(Lora_global,
Lora_NPac)
Condition Length > = 2000 km Latitude‐independent Relative: IVT
100 kg m
−1
s
−1
above climatological
area means for N. Pacific
Time slice Global (A6), North
Pacific (A7)
10.1002/2016GL071541
Mundhenk et al.
(Mundhenk)
Condition >1400 km length, aspect
ratio 1:4, lat limit >16 N/S,
Latitude‐dependent Relative IVT
percentiles
Time slice Global 10.1175/JCLI‐D‐15‐
0655.1
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,782

Table 1
(continued)
Developer (algorithm name) Type Geometry requirements Threshold requirements
Temporal
requirements Region DOI/reference
axis orientation based on
IVT
Muszynski et al.
(TDA_ML)
Condition Topological analysis and
machine learned
Threshold‐free N/A Western United
States, adaptable
to other regions
Experimental
Payne and Magnusdottir
b
,
c
et al. (Payne)
Condition Length > 1200 km,
landfalling only
Latitude‐independentRelative:85th
Percentile of maximum IVT
(1,000–500 mb)
Absolute: IWV >2 cm, 850 mb wind
speed >10 m/s
Time
stitching (12‐
hr
minimum)
Western United
States 10.1002/2015JD023586;
10.1002/2016JD025549
Ramos et al.
b
,
c
(Ramos) Condition Detected for reference
meridians, length ≥1500
km, latitudinal movement
<4.5
0
N
Latitude‐dependent relative: IVT
85th percentile (1,000–300 mb)
Time slice,
but 18‐hr
minimum
for persistent
ARs
Western Europe,
South Africa,
adaptable to other
regions
10.5194/esd‐7‐371‐2016
Rutz et al.
b
(Rutz) Condition Length ≥2,000 km Absolute: IVT (surface to 100 mb) =
250 kg m
−1
s
−1
Time slice Global, low value
on tropics
10.1175/MWR‐D‐13‐
00168.1
Sellars et al.
b
(CONNECT500, CONNECT700)
Track Object identification Absolute: IVT, thresholds tested =
300 (A11), 500 (A12), 700 (A13) kg m
−1
s
−1
Time
stitching,
minimum
24‐hr period
Global 10.1002/2013EO320001;
10.1175/JHM‐D‐14‐
0101.1
Shields and Kiehl
b
(Shields)
Condition Ratio 2:1, length to width
grid points min 200 km
length; 850mb wind
direction from specified
regional quadrants,
landfalling only
Latitude‐dependent Relative:
a
ZN
moisture threshold using IWV; Wind
threshold defined by regional 85th
percentile 850 mb wind magnitudes
Time slice Western United
States Iberian
Peninsula, UK,
adaptable but
regional specific
10.1002/
2016G-
L069476; 10.1002/
2016GL070470
TEMPEST
b
(TEMPEST) Track Laplacian IVT thresholds
most effective for widths
>1,000 km; cluster size
minimum = 120000km
2
IVT > =250 kg m
−1
s
−1
Time
stitching
(time slice
only for this
study)
Global, but
latitude ≥15°
Experimental
Walton et al. (Walton) Condition and track Length ≥2,000 km Latitude‐dependent Relative: IVT >
250 kg/m/s + daily IVT climatology
Time
stitching
Western United
States
Experimental
Note. Algorithm numbering is determined by overall alphabetical order with the first set of numbers assigned to developers who participated in the 1‐month proof of concept (A1–A15), (see GMD
paper), followed by Tier 1‐only participation (A16–A24). ARTMIP algorithm identifier will remain consistent across tiers and scientific papers. Shorthand identifiers also are listed in the devel-
oper column (italicized in parenthesis) to help navigate figure interpretation.
a
ZN relative threshold formula: Q>=Q
zonal_mean
+AR
coeff
(Q
zonalmax
−Q
zonamean
), where Q= moisture variable, either IVT (kg m
−1
s
−1
) or IWV (cm). AR
coeff
= 0.3 except where noted.
(Zhu & Newell, 1998). The Gorodetskaya method uses Q
sat
, where Q
sat
represents maximum moisture holding capacity calculated based on temperature (Clausius‐Clapeyron), an important
distinction for polar ARs. Additional analysis on the ZN method can be found in Newman et al. (2012).
b
Methods used in a 1‐month proof‐of‐concept test (section 5). These methods are
assigned an algorithm id, that is, A1, A2, etc.
c
These 1‐month proof‐of‐concept methods apply a percentile approach to determining ARs. A3 and A8 applied the full MERRA2 climatology
to compute percentiles. A9, applied the February 2017 climatology for this test only. For the full catalogues, A9 will apply extended winter and extended summer climatologies to compute per-
centiles. Please refer to individual publications (DOI reference column in this table) for climatologies used in earlier published studies by each developer. The climatology used to compute per-
centile is often dependent on the data set (reanalysis or model data) being used.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,783

criteria required for AR identification, leading to a greater number of ARs being identified. Similarly, “more
restrictive”generally denotes a method or methods with more restrictive criteria required for AR
identification, leading to a smaller number of ARs being identified. Future work will quantify the
“restrictiveness”of such methods, but these generalizations will be used throughout this paper.
3. Results
This section describes climatological characteristics of ARs based on the ARTMIP methods used in Tier 1.
These characteristics are highlighted via a few key metrics, including AR frequency and duration. Here,
results are presented either along selected transects or in a zonal‐mean framework to facilitate a more
focused analysis.
3.1. AR Frequency
This section discusses AR frequency, which is defined as the percentage of time that a given location is
experiencing AR conditions (i.e., is located within the spatial footprint of an AR). For example, if a given
method produces an AR frequency of 10% at some location, it means that this location is within the spatial
footprint of ARs, as identified by that method, 10% of the time from January 1980 through June 2017, inclu-
sive of all months. AR frequency along the North American and European West Coasts, as well as through
interior western North America, varies greatly as a function of method used (Figure 6). Focusing on the
North American West Coast, nearly all methods exhibit a rapid increase in AR frequency from a minimum
near 32°N toward a maximum near 45°N, followed by a more gradual decrease northward toward 56°N
(Figure 6, top). This distribution closely resembles that of North Pacific storm track density shown by
Lukens et al. (2018; their Figure 4b), among others. In general, both less restrictive criteria and absolute
thresholds lead to more dramatic changes in AR frequency as a function of latitude, with the “Rutz”method
exhibiting the greatest maximum (~14%) and range (~12%) along this transect. The AR frequency of the
“Guan_Waliser”method is an exception to the generalized statements above—it exhibits a gradual increase
as a function of latitude and a small range (~4%) relative to the number of events it identifies. A similar beha-
vior is also seen with Brands_v1. These characteristics arise from the fact that they are both percentile‐based
Figure 4. Selected transects along the North American West Coast (left panel, black dots), through interior western North
America (left panel, red dots), and along the European West Coast (right panel, black dots). The coastal transect points are
determined by selecting all MERRA‐2 reanalysis grid points that have fractional land/sea cover between 32°N and
55°N (and 130–115°W) for North America, and between 35°N and 62°N (and 15°W –10°E) for Europe. The interior
transect points are determined by subjectively selecting grid points that represent a significant topographic “crest”
between 32°N and 55°N.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,784

relative methods that use a latitude‐dependent IVT threshold, where the direct influence on AR frequency
from the climatological meridional gradient in IVT tends to be smoothed out. Such a smoothing effect is
less obvious with “Brands_v2”and “Brands_v3”, likely because the fixed lower limit of IVT becomes
dominant compared to the less restrictive percentile thresholds in these two methods, making them
inclined toward absolute methods. In a similar sense, relative methods that use latitude‐independent IVT
thresholds (“Payne”,“Lora_NPac”,“Lora_global”) agree better with absolute methods, because for a given
region the IVT threshold in these latitude‐independent relative methods is nothing but a fixed value.
Relative methods can vary substantially in their methodology, from percentile‐and climatology‐based
thresholds (“Brands,”“Guan and Waliser,”“Lavers,”“Lora,”“Mundhenk,”“Payne and Magnusdottir,”
“Ramos,”“Viale,”and “Walton”) to thresholds based on spatial anomalies (“Gorodetskaya”and “Shields
and Kiehl”). Interestingly, the AR frequency from the machine learning method, “TDA_ML,”is
characterized by a maximum of just 2% near 39°N and declines to 0% north of 45°N, where many
methods produce their larger frequency values. Muszynski et al. (2019) note that this method frequently
produces false negatives (i.e., it fails to detect ARs) when an AR merges with another AR or “some other
event with high concentration of water vapor and similar topological structure, such as an extratropical
cyclone.”This would happen with higher frequency in the more active storm track at latitudes north of
45°N, which may explain the rapid drop‐off in AR detection associated with this method.
Focusing on the selected transect through interior western North America, the AR frequency is greatly
reduced for nearly all methods (Figure 6, center). However, the “Guan_Waliser”method is a remarkable
outlier here, as it exhibits an AR frequency (~8–10%) only slightly lower than at the same latitudes along
Figure 5. Tables showing the names of ARTMIP Tier 1 methods grouped into (top) absolute/relative/machine learning
clusters and (bottom) global/regional clusters. For the bottom table, the region(s) over which data are used from each
method are given in parenthesis following the method name. Note that this is not a comprehensive list of all AR identi-
fication and tracking methods found in scientifically relevant literature; only those methods used in this study are shown.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,785

the North American West Coast. This clearly results from being a less
restrictive and relative method. The AR frequency of other methods is
much lower (~1–4%), with the “Rutz”and “Brands_v1”methods, owing
to their less restrictive thresholds, being the largest of these at most lati-
tudes. Most other methods exhibit a coastal maximum near 45°N (albeit
of varying magnitude) that shifts northward to an interior maximum near
48°N. This shift arises because ARs making landfall near 45°N preferen-
tially extend inland toward the east‐northeast along the relatively low‐
elevation corridor of the Columbia River Basin, as shown by Rutz et al.
(2015, their Figure 3). Additional, secondary, corridors of inland penetra-
tion are located south of 32°N and north of 52°N, and all of these corridors
play an important role in heavy precipitation events (Alexander et al.,
2015) and growth of vegetation (Albano et al., 2017) over interior regions.
In contrast, areas downstream of major topographic barriers feature a lar-
ger decrease from coastal to interior frequency due to AR decay, as moist-
ure is more effectively removed by orographic precipitation. This is
particularly true for ARs making landfall between ~32°N and 38°N, which
are severely disrupted by the southern Sierra Nevada Mountains (eleva-
tion 2–4 km), drastically lowering the inland frequency between ~38–
44°N (following a typical trajectory of inland penetration; see Rutz et al.,
2015, their Figure 3).
Along the European West Coast, “Rutz”identifies the greatest AR fre-
quency nearly everywhere, and diverges markedly from
“Guan_Waliser”between 44°N and 60°N. This divergence is likely due
to a higher climatological value of IVT at these latitudes, which generally
causes absolute methods to identify a greater number of ARs than relative
methods. Other methods such as “Lora,”“Mundhenk,”and “Tempest”
follow a distribution very similar to “Rutz,”but with smaller amplitudes.
Once again, the fairly good agreement between the absolute method of
Rutz and latitude‐independent method of Lora is not surprising. The dra-
matic jump in AR frequency near 45°N may be due to some combination
of climatology (i.e., placement of the storm track; e.g., Lukens et al., 2018)
and the greater number of coastal transect points at latitudes north of
45°N. It is worth noting that during the ARTMIP 1‐month experiment
described in Shields et al. (2018; their Figure 3), the human‐control analy-
sis yielded a greater AR frequency than any automated method along both
coastal transects. (The human‐control analysis consisted of two graduate students counting “by eye”all ARs
making landfall for the North American and European coastlines for the month of February 2017.)
Clearly, different AR identification and tracking methods produce widely varying results for AR frequency
along coastal transects. It is important to remember that the criteria of each individual method used in
ARTMIP have been developed to answer specific scientific questions, often driven by regional and/or
impacts‐specific considerations. Since different questions were asked, it should be no surprise that different
methods are used and different results produced. Nevertheless, while the ARTMIP methods do not agree on
absolute values of AR frequency, they do exhibit remarkable agreement in their latitudinal distribution,
except for an outlier (“Guan_Waliser”), which, among all the methods examined in this study, is the only
relative method that both has a global coverage and uses percentile‐based, latitude‐and longitude‐dependent
IVT thresholds. Also, despite generally large intermethod differences, Brands_v2 and CONNECT500 yield
practically identical results on both continents, which is quite surprising since the two methods have been
developed independently.
The ARTMIP methods' good agreement regarding the latitudinal distribution of AR frequency is more
clearly seen by normalizing each method as follows: for each method, the largest and smallest value along
a given transect are given values of 1 and 0, respectively, and all values are then normalized to this scale
(Figure 7). For example, if the largest and smallest AR frequency along a given transect are 24% and 8%,
Figure 6. AR frequency of ARTMIP methods for selected transects (a) along
the North American West Coast, (b) through interior western North
America, and (c) along the European West Coast. Note that some methods
are only available over certain regions.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,786

respectively, a value of 12% will be normalized to 0.25. Exceptions to good
agreement (e.g., “TDA_ML,”“Shields,”and “PNNL_lq”) are more
prominent at lower latitudes, where they identify a relatively larger
number of ARs than most methods (the machine learning method,
“TDA_ML,”is an outlier here). The “Guan_Waliser”method is another
exception owing to its steady rise with latitude throughout the North
American coast, which appears more dramatic when normalized
precisely because it is so gradual in absolute terms. These normalized
results can be further examined by clustering methods according to key
differentiating criteria such as whether absolute or relative thresholds
are used (Figure 8). Focusing on the median of the absolute and relative
clusters (the thick black and blue lines, respectively) reveals excellent
agreement in the distribution of AR frequency.
These results suggest that the ARTMIP methods are not identifying funda-
mentally different features (as could be inferred from the nonnormalized
results), but rather that their numbers are simply scaled as a function of
how restrictive their criteria are. To investigate this further, Figure 9 pre-
sents, for each method, composites of IVT magnitude and identified ARs
anytime that method identifies an AR at a point along the northern
California coast (39°N, 123.75°W). There are notable differences: Less
restrictive methods are characterized by a smaller composited AR because
they identify both weak and strong events (e.g., “Guan_Waliser,”“Rutz,”
and “Tempest”), whereas more restrictive methods are characterized by a
larger composited AR because they identify only strong events (e.g.,
“CONNECT700”and “PNNL_LQ”). In addition, while most methods'
composite ARs exhibit a west/southwest to east/northeast orientation,
those based entirely or partly on IWV have a more zonal orientation
(e.g., “Goldenson”and “Shields”). However, the methods' composited
AR footprints generally cover the same region. The bottom right panel in
Figure 9 highlights this aspect—anytime the “PNNL_LQ”method, one of
the most restrictive, identifies an AR at the coastal point, the number of
other methods identifying ARs within the domain shown are counted,
and the average over all “PNNL_LQ”ARs is shown in this panel. The
results indicate that when one of the most restrictive methods identifies
an AR at the coastal point, most other methods (~15), which are less restrictive, also identify an AR near this
point, and this number decreases with distance as a function of decreasing IVT. Hence, more restrictive
methods' AR composites are shown to be an approximate subset of their less restrictive counterparts.
There are advantages and disadvantages to this normalization approach. One key advantage is that baselin-
ing the distribution of AR frequency along these transects is necessary to assess changes predicted by climate
models; these results increase confidence in the general shape of the latitudinal distribution. One key disad-
vantage is that AR‐related impacts cannot simply be normalized—emergency management is much more
interested in how often these impacts will be encountered than in the general shape of AR frequency along
the coast. Thus, more work is needed to constrain the range of AR frequency depicted above, and this is dis-
cussed in more detail in section 5.
3.2. AR Duration
This section discusses AR duration, which is defined as the continuous length of time that a given location is
experiencing AR conditions (i.e., is located within the spatial footprint of an AR). For example, if a given
method produces an AR duration of 10 hr at some location, it means that when this location is within the
spatial footprint of ARs, as identified by that method, the average duration of such conditions is 10 hr from
January 1980 through June 2017, inclusive of all months. AR duration along the North American and
European West Coasts varies as a function of method used, but not as greatly as AR frequency
(Figure 10). Focusing on the North American West Coast, most methods exhibit a gradual increase in AR
Figure 7. Normalized AR frequency of ARTMIP methods for selected trans-
ects (a) along the North American West Coast, (b) through interior western
North America, and (c) along the European West Coast. Note that some
methods are only available over certain regions.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,787

duration from 32°N toward a maximum near 42–44°N (matching the
maximum in AR frequency), followed by a steadier decrease northward
toward 50°N, plateauing north of there (Figure 10, top). In their pre‐
ARTMIP study, focused on Bodega Bay (~38°N along the northern
California coast), Ralph et al. (2018b) showed that a group of methods fea-
turing lower IVT thresholds (“Rutz,”“Guan_Waliser,”and “Gershunov”)
clusters strongly in both frequency (~23 events per year) and mean event
duration (~24 hr). Figure 4 of Ralph et al. (2018b) shows that this agree-
ment, in terms of the number of events (which can be related to AR fre-
quency, given the similar duration), is primarily the product of the
fortuitous latitude at Bodega Bay, where results from “Guan_Waliser”
cross over with those of “Rutz”and “Gershunov”(Figure 6). Figure 10
also shows that these two methods agree on duration at this latitude.
More restrictive methods, which typically detect fewer ARs, may also
translate into ARs having shorter average durations (e.g.,
“CONNECT700,”“TDA_ML,”,“Shields and Kiehl”). However, this is
not necessarily the case for specific events, as can be seen by examining
the AR identification time series shown later in Figure 12 (top panels).
In general, both less restrictive method criteria and absolute thresholds
lead to greater AR duration, with the “Rutz”and “Gershunov”methods
described above being among the largest along this transect (the
latitude‐independent “Lora_global”following closely behind). The
“Guan_Waliser”method also produces large AR durations, including
the largest south of ~37°N. In addition, the “Guan_Waliser”method also
produces the largest mean AR duration (as it does mean AR frequency)
through interior western North America (Figure 10, center). These char-
acteristics arise from the fact that it is a relative method that uses
percentile‐based, latitude‐and longitude‐dependent IVT thresholds, as
explained earlier. In fact, north of ~46°N, this method produces slightly
larger mean durations along the interior transect than it does at similar
latitudes along the coastal transect, perhaps because its criteria preferen-
tially select for more powerful (and hence, longer‐lived) events over
regions where IVT is climatologically weaker. In Europe, the “Lavers”
and “Ramos”methods cluster closely together, which is interesting since
these two methods were developed within this region (Figure 10, bottom). Over this region, for many meth-
ods, there is little change or a slight decrease in mean event duration as a function of increasing latitude,
although the less restrictive “Rutz”method is an exception, peaking at 33 hr near 50°N. As with AR fre-
quency, machine learning methods tend to cluster toward lower values of AR duration.
The mean AR duration, after each method is normalized from 0 to 1 and then clustered, shows that the rela-
tive distributions of most methods are in good agreement along the North American and European West
Coast (Figure 11). This agreement, however, is not as robust as that observed for AR frequency. For the
North American West Coast, relative methods produce relatively greater durations between 35°N and
43°N and absolute methods produce relatively greater durations between 47°N and 55°—a similar pattern
is observed along the European West Coast. These relative differences as a function of latitude are not as
apparent for AR frequency (Figure 7), suggesting that while relative and absolute methods share similar dis-
tributions in overall AR activity, relative methods tend to observe longer duration events further south.
3.3. AR Concurrence
This section analyzes the extent to which the ARTMIP methods agree or disagree on the identification of AR
conditions along the North American West Coast during events of varying intensity, and the relationship
between the methods' identification of ARs conditions and observed precipitation. To do so, the methods'
identification of AR conditions along a selected coastal transect during two events, one strong and one weak,
are explored. It is important to note that this analysis and the results shown in Figure 12 are based on the
Figure 8. Normalized and clustered (based on absolute or relative thresh-
olds) AR frequency of ARTMIP methods for selected transects (a) along
the North American West Coast, (b) through interior western North
America, and (c) along the European West Coast. Note that some methods
are only available over certain regions.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,788

peak IVT (blue time series in top panel) and the presence (or absence) of AR conditions (black dots in top
panel) along the entire coastal transect, and not at an individual point.
The first event (12–16 February 2014) is characterized by a broad area of IVT ≥250 kg m
−1
s
−1
making land-
fall along the U.S. West Coast and extending inland, with areas exceeding the 85th percentile of climatolo-
gical IVT embedded within the core (Figure 12a). This event produced heavy precipitation along the coastal
and the interior northwestern United States, triggering a series of avalanches that resulted in 10 fatalities
(Hatchett et al., 2017). Most of the ARTMIP methods identify AR conditions along the coastal transect either
throughout, or nearly throughout, the entire period. Some methods, such as “Brands_v2,”“Connect500,”
and “Payne,”are very sensitive to periodic surges and lulls in IVT magnitude, identifying ARs during the for-
mer. Methods that are more restrictive, such as “Connect700,”“PNNL1_hagos,”“PNNL_lq,”and
“TDA_ML”do not identify AR conditions as frequently, particularly at times when peak IVT along the coast
drops below their more restrictive thresholds. Furthermore, it must be noted that some methods, such as the
“PNNL”methods above, only identify ARs if and when they intersect the coast, but not before or after.
The second event (23–24 October 2006) is characterized by a broad area of IVT ≥250 kg m
−1
s
−1
terminating
along the coast of British Columbia, with the 250 kg m
−1
s
−1
contour overlapping the northern edge of the
selected transect (Figure 12b). Precipitation is very light, and the authors are not aware of significant impacts
associated with this event. The ARTMIP methods generally disagree as to whether or not AR conditions
occur along this transect, and the disagreement extends beyond differentiation into absolute and relative
methods. The “Rutz”and “Tempest”methods most frequently identify an AR along the transect during this
time period—both are based on an absolute threshold of IVT ≥250 kg m
−1
s
−1
, but the “Rutz”method uti-
lizes an Eulerian framework for identifying ARs, whereas the “Tempest”method utilizes a Lagrangian fra-
mework. Certain relative methods such as “Guan_Waliser,”“Lora,”and “Shields”also identify an AR along
Figure 9. (all except bottom right) For each method, composites of IVT at all times when that method identifies an AR at
the coastal location of 39°N, 123.75°W (black circle). Blue color shading represents IVT magnitude and purple contour
indicates IVT of 350 kg m
−1
s
−1
. (bottom right) Shading indicates the number of methods identifying an AR anytime the
“PNNL_LQ”method identifies an AR at 39°N, 123.75°W.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,789

this transect nearly 50% of the time (or during adjacent time steps). Other
methods, such as “PNNL1_hagos”and “PNNL2_lq,”only identify
instances when AR conditions are met at the coast (i.e., landfalling ARs)
so very few instances were denoted as AR conditions for this AR that
barely made landfall along the coast of British Columbia.
One key point is that there is a difference between identifying AR condi-
tions at one time step and identifying an AR event, which is often defined
as having some minimum duration (such as the 12‐hr minimum in
Figure 10). AR duration plays a key role in storm‐total precipitation and
streamflow (Ralph et al., 2013), and along with peak IVT intensity, forms
the basis of a forthcoming AR scale, which categorizes the strength and
impacts of ARs (Ralph et al., 2019). A further consideration is that while
prolonged AR duration often drives impacts over land, the peaks and
troughs in IVT intensity along the coast are of great interest to those focus-
ing on the physical processes involved in strengthening, maintaining, or
weakening ARs. Finally, based on this limited analysis, ARTMIP methods
exhibit greater agreement regarding those storms that are more meteoro-
logically impressive and associated with heavier precipitation. Of course,
significant meteorological events do not always produce significant
impacts, and many factors need to be considered, but it is encouraging
that all else held constant, a large majority of methods agree to classify
“the big ones”as ARs.
3.4. AR Seasonality
In this section, we assess AR seasonality by calculating, at each latitude,
the number of methods that yield a maximum AR frequency during a
given month (Figure 13; i.e., for any latitude, the sum across all columns
will be equal to the number of methods for that transect). The month of
maximum AR frequency along the North American West Coast is charac-
terized by a gradual shift from north to south during the course of the bor-
eal cool season (Figure 13, top). More specifically, it occurs between 48°N
and 54°N (near British Columbia and Vancouver Island) during October,
42–48°N (Washington and Oregon) during November, 36–42°N
(Northern California) during December, and south of 36°N (Southern
California) during January. This result agrees well with the results for
peak IVT intensity shown by Dettinger et al. (2018), among others. The month of maximum AR frequency
should not be confused with AR frequency—in other words, the blue shading indicating a December max-
imum near 37°N means that a large majority of methods agree that AR frequency, at this latitude, features a
maximum in December. It does not mean that the December AR frequency is greater here than some other
location, though it might be.
Along the interior western North America transect, the month of maximum AR frequency is quite varied as
a function of latitude (Figure 13, center). North of ~45°N, the pattern is similar to that along the coast, with
the month of maximum AR frequency gradually shifting southward September through November. South of
~41°N, this southward shift continues, to some extent, into December and January, but is less clearly seen
because of the influence of the monsoon circulation, which produces a maxima during September and
October at these latitudes. Perhaps the most unexpected phenomenon shown here is the June maximum
between 41°N and 45°N. At these latitudes, during June, seasonally increasing moisture interacts with a
jet stream that remains sufficiently strong to produce a relatively large number of ARs. However, given a
warmer atmosphere in June than during winter months, it is likely that many of these ARs fail to fully satu-
rate the atmospheric column and produce less precipitation than ARs of similar IVT magnitude during the
winter, particularly over lower elevations.
The month of maximum AR frequency along the European West Coast is characterized by a rapid shift from
north to south concentrated during September–October, at the onset of the boreal cool season (Figure 13,
Figure 10. AR duration of ARTMIP methods for selected transects (a) along
the North American West Coast, (b) through interior western North
America, and (c) along the European West Coast. Note that some methods
are only available over certain regions. Only AR events lasting ≥12 hr
qualify.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,790

bottom). More specifically, it is maximized north of 55°N during
September and south of 55°N during October (similar to the season-
ality of these same latitudes along the North American West Coast;
Gershunov et al., 2017). At some latitudes, such as those near
Scotland or southern Spain, some methods identify December as
the month of maximum AR frequency. One possible explanation is
that these latitudes can observe landfalling ARs from a greater variety
of directions (see the inset maps of grid points used for coastal trans-
ects), and the tendency to do so may vary by month. It is also possible
that some of this is due to a more fragmented European West Coast
versus that of North America. For both the North American and
European West Coasts, it is notable that at least 1 or 2 methods iden-
tify January as the month of maximum AR frequency at nearly every
latitude. This results from a wide variety of methods identifying
January along the North American West Coast, and the “Lavers”
(further south) and “Shields”(further north) methods consistently
identifying January along the European West Coast.
3.5. AR Zonal‐Mean Area, Poleward IVT, and “Efficiency”
One oft‐quoted result, developed from an early series of seminal
papers on ARs (Newell et al., 1992; Zhu & Newell, 1998), is that
ARs are responsible for ~90% of poleward water vapor transport in
the midlatitudes, despite encompassing only ~10% of global circum-
ference at any given latitude and time. This section derives motiva-
tion from this early work and explores related metrics across
various ARTMIP methods. Results shown are limited to global meth-
ods, which can be compared to each other because they consider all
latitudes and longitudes. In contrast, regional methods cannot be
compared because they consider only certain regions.
3.5.1. Zonal‐Mean Area
The first metric examined here is the zonal‐mean AR area (i.e., the
time‐mean spatial footprint, along a given latitude band, of identified
or tracked ARs), expressed as a fraction of global circumference, for
each global method (Figure 14, top). Most of the global methods are
characterized by a maximum zonal‐mean AR area in the midlati-
tudes (~10% of global circumference), a rapid decrease toward higher latitudes, and a gradual decrease
toward lower latitudes. The rapid decrease toward higher latitudes is due to rapidly decreasing mean water
vapor (and hence, IVT) at these latitudes, whereas the gradual decrease toward lower latitudes is dominated
by decreasing mean wind (and hence, IVT) further from the mean storm track. One notable exception is the
“Rutz”method, which identifies a large fraction of the intertropical convergence zone as an AR, since it does
not account for climatology and has no width requirement. In general, absolute methods exhibit greater var-
iance in zonal‐mean AR areas as a function of latitude than relative methods. In addition, less restrictive
methods predictably identify greater zonal‐mean AR areas overall than more restrictive methods.
Here, the “Guan_waliser”method is interesting in two respects. First, it identifies larger zonal‐mean AR
areas at higher latitudes of both hemispheres than any other method. This arises because this method uses
a latitude‐dependent 85th percentile IVT threshold, where the direct influence on AR activity from the cli-
matological meridional gradient in IVT tends to be smoothed out, resulting in a more gradual decrease in AR
occurrence toward high latitudes, as also explained earlier. Second, for the same reason above, the fractional
zonal‐mean AR area identified is remarkably stable throughout the midlatitudes of both hemispheres at
nearly 10%, which is the value from Zhu and Newell (1998) for the fraction of global circumference encom-
passed by ARs. A number of other methods (i.e., “Rutz,”“Lora_global,”“Lora_npac,”“Tempest,”
“Mundhenk,”and “Connect500”) also approach this value at low or middle latitudes, but their distributions
are more variable as a function of latitude.
Figure 11. Normalized and clustered (based on absolute or relative thresholds)
AR duration of ARTMIP methods for selected transects (a) along the North
American West Coast, (b) through interior western North America, and (c) along
the European West Coast. Note that some methods are only available over certain
regions. Only AR events lasting ≥12 hr qualify.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,791

The midlatitude (~30–60°N/S) global circumference occupied by ARs is in the range of ~2–15% when con-
sidering all global methods, excluding the “CONNECT”methods (for which the range is lower, due to their
fairly restrictive thresholds). This is important for three reasons. First, being significantly below 50% means
that even the least restrictive methods examined here (such as “Rutz”) are identifying discrete features that
are making large contributions to IVT relative to their size. Second, the average global circumference occu-
pied in the core of the midlatitudes (~45°N/S) being ~5–15% means that these features occupy more space
than the cold‐frontal zones associated with extratropical cyclones, and therefore the concept of an AR is dis-
tinct and useful. Finally, it is encouraging that these results, based on 5 methods and ~38 years of global data,
align so well with those discussed by Zhu and Newell 20 years ago.
3.5.2. Zonal‐Mean Poleward IVT
The second metric examined here is the AR‐related zonal‐mean poleward IVT (i.e., poleward IVT occurring
within the spatial area of ARs; Figure 14, middle). For all methods, the AR‐related zonal‐mean poleward IVT
is maximized in the midlatitudes. The variation in the magnitude of its maximum and meridional range
clearly exhibits a dependence on the threshold magnitudes chosen for identifying ARs, which is most clearly
seen using the absolute methods. For example, “connect500”attributes a greater amount of zonal‐mean pole-
ward IVT to ARs than “connect700,”simply because the former uses an IVT threshold of 500 (versus 700) kg
Figure 12. ARTMIP methods' identification of AR conditions (dots) along a selected transect (hatched in the spatial
panels), peak IVT (light blue line, with values below 250 kg m
−1
s
−1
dashed) and mean precipitation (blue line) along
the transect. Composite of IVT (IVT ≥250 kg m
−1
s
−1
and IVT ≥85th percentile contoured as black dashed and black
solid lines, respectively), composite of IWV (IWV ≥20 mm contoured as a solid black line), and cumulative precipitation
for events centered on (a) 12–16 February 2014 and (b) 23–24 October 2006. The time steps composited for each event are
lightly shaded in the top panel. Listed methods use relative thresholds if italicized, no thresholds if bolded, and absolute
thresholds otherwise.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,792

m
−1
s
−1
, and identifies ARs as spatially larger features. Similarly, the
“tempest”and “rutz”methods, based on 250 kg m
−1
s
−1
, attribute an even
greater amount of zonal‐mean poleward IVT to ARs, with “rutz”attribut-
ing more than “tempest”because it is less restrictive with other criteria.
The relative “lora”method, with a less restrictive threshold requirement
of IVT ≥100 kg m
−1
s
−1
above climatology, attributes nearly as much
zonal‐mean poleward IVT to ARs as “rutz.”The other relative methods
“guan_waliser”and “mundhenk”have more restrictive criteria, and attri-
bute smaller fractions of zonal‐mean poleward IVT to ARs. The “guan_-
waliser”method is notable because of its relative smoothness at high
latitudes, being the most generous in attributing zonal‐mean poleward
IVT to ARs north of ~55°N. It is interesting that this is roughly the same
latitude at which the AR frequency of the “guan_waliser”method
becomes greater than that of “rutz”and “lora”along the North
American and European West Coasts (Figure 6).
3.5.3. Zonal‐Mean AR Efficiency
The final metric examined in this section is the zonal‐mean AR “effi-
ciency,”defined, in terms of the two metrics examined previously, as
the ratio of zonal‐mean poleward IVT to the fractional (i.e., unitless)
zonal‐mean spatial area of ARs. It is referred to here as the zonal‐mean
AR efficiency because it describes the quantity of poleward water vapor
transport per unit area of AR. One inherent problem with this metric is
that efficiency, as defined above, is naturally higher for more restrictive
AR identification and tracking methods. However, it is still interesting
to explore this metric, and particularly how it changes as a function
of latitude.
The methods with the largest zonal‐mean AR efficiency across most lati-
tudes are “CONNECT500”and “CONNECT700”(Figure 14, bottom).
This is not surprising given that the fairly restrictive criteria of IVT ≥
500 and 700 kg m
−1
s
−1
limits the number of ARs identified by these
methods to only the strongest of those identified by other methods. In fact,
the lack of events due to these restrictive criteria is clearly seen to affect
the results over high latitudes. The other methods cluster more closely
together, particularly over mid and high latitudes, although the “guan_-
waliser”method identifies more ARs over Antarctica, owing to its less
restrictive criteria, and hence, the efficiency is lower. In the tropics, the
“rutz”method, which was designed for midlatitude applications, is the
least efficient, as it often identifies regions of broad tropical moisture
transport as ARs. In contrast, the “lora”method becomes more efficient
at these latitudes, since it requires that IVT exceed the climatological
mean by 100 kg m
−1
s
−1
. Finally, AR efficiency using the “mundhenk”
and “tempest”methods, which are based on IVT ≥94th percentile of
the anomalies above the climatology and IVT ≥250 kg m
−1
s
−1
, respectively, are relatively steady across
all latitudes.
In summary, this section shows that most global methods used within ARTMIP broadly reproduce the clas-
sic results of Zhu and Newell (1998) in terms of AR size and significance for global water vapor transport.
3.6. Spread Among Methods
This section explores the relative difference between results for each individual method and the all‐
method median for AR frequency, the month of maximum AR frequency, and the seasonal range of
AR frequency (Figure 15). Results are presented along seven selected transects: the Pacific Northwest
(PNW; 41–52.5°N), Northern California (NorCal; 35–41°N), Southern California (SoCal; 32–35°N), the
interior western United States (WUS_In; 32–54°N), South America (SAmer; 18–56°S), the United
Figure 13. AR seasonality (month of maximum frequency) of ARTMIP
methods for selected transects (a) along the North American West Coast,
(b) through interior western North America, and (c) along the European
West Coast. Note that some methods are only available over certain regions.
Color shading indicates the number of methods for which a given month is
the month of maximum AR frequency at each latitude.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,793

Kingdom (UK; 49–60°N), and Iberia (Ib; 35–48°N). This analysis is fairly exhaustive, and a full descrip-
tion of every aspect would be very lengthy, so this section focuses on the highlights. This analysis offers
some insight as to which methods produce results closer to the median, and which methods produce
results further from the median, perhaps even being characterized as outliers. Here, relative difference
for a given region and given method is calculated as the difference between a given method and the
all‐method median normalized by the all‐method median.
For AR frequency, the methods closest (lighter color shading) to the all‐method median (<±36%) are
“Goldenson,”“Mundhenk,”“Payne,”and “Shields,”whereas the methods furthest (darker color shading)
from the all‐method median (>±60%) are “CONNECT700,”“Gershunov,”“Guan_Waliser,”“PNNL2_lq,”
and “Rutz”(Figure 15, top). Generally, the methods closest to the all‐method median are relative methods,
with the exception of “Goldenson,”whereas the methods furthest from the all‐method median are absolute
methods, with the exception of “Guan_Waliser.”This breakdown by absolute/relative is not too surprising
since absolute methods tend to accentuate climatological differences while relative methods tend to dimin-
ish them. Of those methods furthest from the median, the two methods identifying much lower AR fre-
quency are “CONNECT700,”which uses a high IVT threshold of 700 kg m
−1
s
−1
, and “PNNL2_lq,”
which has a number of restrictive criteria (Table 1). The three methods identifying much higher AR fre-
quency are “Gershunov”and “Rutz,”which have similar and less restrictive criteria, and
“Guan_Waliser,”which also has fairly low criteria (85% climatological IVT) for a relative method. It is
Figure 14. AR mean area (top), poleward IVT (center), and “efficiency”(bottom) of the ARs identified and tracked by the
various ARTMIP methods.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,794

worth noting that most methods are relatively far from the median for AR frequency along the transect
through the interior western United States, which results partly from differences in how methods assess
ARs over complex terrain, and partly from calculating percentage differences between small numbers.
Another point of interest is the rather notable differences in the three “Brands”methods, which shows
that relatively minor changes to the AR identification method (e.g., the threshold percentiles) can
significantly alter results. Finally, the regional methods “Lavers,”“Ramos,”and “Viale,”which are available
over the United Kingdom, Iberia, and Chile, respectively, all produce a lower AR frequency than the median
(i.e., global methods) over these regions. One possible explanation for this interesting result is that regional
methods are more finely tuned to their respective areas, and that this is manifested as more restrictive
criteria. For example, both “Lavers”and “Ramos”use time‐dependent percentiles based on climatological
IVT, and “Viale”imposes the restriction that an AR must be associated with a frontal system. The machine
learning technique, “TDA_ML,”produces AR frequencies below the median, possibly because the algorithm
Figure 15. Diagrams showing the relative difference of results from each ARTMIP method to the all‐method median for
the metrics of annual (top) AR frequency, (center) month of peak AR frequency, and (bottom) seasonality (or range) of AR
frequency. Results are shown for coastal transects of the Pacific Northwest (PNW), Northern California (NorCal),
Southern California (SoCal), the interior western United States (WUS_In), South America (SAmer), the United Kingdom
(UK), and Iberia (Ib).
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,795

employed by Muszynski et al. (2019) exhibits a fairly strong resolution‐dependent decrease in the “sensitivity
score”(the proportion of identified ARs that are correctly identified) as resolution increases, with over 25%
of features being misclassified (relative to their training data set) as non‐ARs at high resolution. Muszynski
et al. (2019) hypothesize that this is due to an interaction between the decrease in smoothness of
the IWV field as resolution increases and the underlying topology‐based method that they use to identify
potential ARs.
The month of maximum AR frequency is also examined (Figure 15, middle), and the following examples
assist with interpretation: the median month of maximum AR frequency over the United Kingdom is
September (i.e., peak month of 9), whereas for “CONNECT700,”it is October (i.e., peak month difference
of 1). Similarly, the median month of maximum AR frequency over SoCal is February (i.e., peak month of
2), whereas for “Goldenson,”it is December (i.e., peak month difference of −2). In many cases, there is
no difference between the all‐method median and the month of maximum AR frequency identified by most
methods, though there are some notable exceptions. For SAmer, both the “Guan_Waliser”and “Viale”
methods, the latter of which is focused on this region, feature a month of maximum frequency 4+ months
later than the median of February. This effectively means that whereas most methods are identifying the
maximum in Austral Summer, these methods identify the maximum in Austral Winter, which is possible
because the storm track in this region is less seasonally variable than over the Northern Hemisphere (e.g.,
Trenberth, 1991). The “Shields”method has a tendency to identify frequency maxima a few months later
than the median over nearly every region with the exception of SoCal, and this may be due to the tendency
of this method to detect only the stronger storms, typically in midwinter, when the eddy‐driven jet is further
south. Only the “Lora”and “Tempest”methods show no deviations from the median over all regions,
though a few other methods come close.
The third metric assessed here is the seasonality (i.e., range) of AR frequency (Figure 15, top), and many
results are similar to those for AR frequency itself. The methods closest (lighter color shading) to the all‐
method median (<±30%) are “Brands_v1,”“Goldenson,”“Mundhenk,”and “Payne,”the latter three of
which are all close to the median for AR frequency as well (Figure 15, top). The methods furthest (darker
color shading) from the all‐method median (>±40%) are “CONNECT700,”“PNNL1_hagos,”“PNNL2_lq,”
“Rutz,”“
Shields,”and “TDA_ML.”As with AR frequency, the methods closest to the all‐method median
are relative methods, with the exception of “Goldenson,”whereas the methods furthest from the all‐method
median are more mixed. Of those methods furthest from the median, the four methods identifying much
weaker seasonality are “CONNECT700”and the “PNNL”methods, which are both quite restrictive, and
the machine learning technique, “TDA_ML.”The two methods identifying much stronger seasonality are
“Rutz”and “Shields,”the latter of which features a larger range in AR frequency across most regions, despite
having a smaller AR frequency in a few of them. As with AR frequency, most methods are relatively far from
the median for the range in AR frequency along the transect through the interior western United States. Of
the global methods, “Lora_global”and “Rutz”exhibit a stronger seasonality over all regions, whereas
“CONNECT700”exhibits a weaker seasonality over all regions except SAmer. As with AR frequency, the
regional methods “Lavers,”“Ramos,”and “Viale”all produce a weaker seasonality than the median (i.e.,
global methods) over their respective regions.
For both the AR frequency and the seasonality (i.e., range) of AR frequency, it is notable that each method
generally exhibits either a positive or a negative relative difference from the median across all transects.
Exceptions to this generalization are most commonly noted along the interior western U.S. transect, which
is the only one located amidst complex topography far from a coast. Hence, the ARTMIP methods used to
identify ARs do not seem particularly sensitive to the region in which they are employed.
4. Discussion
Results based on the ARTMIP methods have been described in terms of “clusters,”which are groupings of
methods that approach AR identification and tracking similarly in a few critical ways. These clusters differ-
entiate between methods with very different approaches, and often very different results, without knowing
the nuances of each individual method within the cluster. The three cluster pairs this text has focused on, as
discussed in section 2, are absolute/relative,global/regional, and the more subjectively defined less/more
restrictive. The subdivision of relative methods into latitude‐dependent and latitude‐independent relative
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,796

methods has also been noted. There are many other cluster pairs (e.g., length/no length,ortime‐slicing/time‐
stitching) that have been omitted from this study in the interest of brevity. In each of these cluster pairs, both
clusters feature advantages and disadvantages, some of which are discussed below.
The absolute and relative clusters are perhaps the most fundamentally different in their approach to identify-
ing ARs. One key advantage of absolute methods is that they ensure a minimum physical threshold is met
before features are identified as ARs, which can be useful when considering only stronger events. Some
methods, such as “CONNECT700”(threshold IVT ≥700 kg m
−1
s
−1
), are designed to consider only the
strongest events. Furthermore, ARs identified following an absolute method are sometimes comparable
across regions, provided the appropriate threshold is carefully chosen to pursue specific applications. In that
regard, a relative method might be particularly useful when a single absolute threshold does not work well
across all the regions of interest—an example being over polar regions, where a temperature‐adjusted (i.e.,
climatology‐dependent) AR threshold has proven useful in detecting AR landfalls (Gorodetskaya et al.,
2014). Another good example is the lengthy southwestern coast of South America, which stretches from
18–56°S and encompasses a wide range of climatological IVT values. Some methods combine relative and
absolute thresholds to leverage the advantages of each. For example, the latitude‐dependent relative
“Guan_Waliser”method combines a relative threshold (IVT ≥85th percentile of climatological IVT) with
an absolute threshold (IVT ≥100 kg m
−1
s
−1
), the latter of which eliminates extremely weak features, par-
ticularly closer to the poles.
Another key advantage of relative methods is that they facilitate the pursuit of AR science in regions where it
is more difficult to do so using absolute methods. For example, imagine one wants to investigate the impacts
associated with the inland‐penetrating AR depicted in Figure 2. The “Guan_Waliser”method may be a good
choice since it identifies a broad inland region as being located within the AR (other relative methods have
more restrictive criteria and identify less area within the AR), and many impacts within this region could be
AR related. Of course, it may be that in some cases this region is too broad, and choosing an absolute method
that still highlights the inland penetration of the AR, but focuses more closely along its axis or region of core
intensity, is appropriate. These are very difficult decisions that need to be made based on the specifics of the
question being asked.
Climate change poses yet another point for consideration. As atmospheric temperature and moisture
increase following the Clausius‐Clapeyron equation, IWV will increase, and IVT will increase (unless
increases in water vapor are offset by decreases in wind). As the background moisture field increases, abso-
lute methods using thresholds based on the current climatology may struggle to distinguish between this
increased background moisture and coherent ARs resulting from dynamical processes. Hence, relative meth-
ods using thresholds based on climatology may be better suited to assess relative changes in ARs due to
dynamic and thermodynamic factors between our current climate and that of the future. For example,
one can compare two relative methods—one in which the percentile threshold is applied to the respective
climatology of the present and future (to isolate the dynamic factor) and one in which the percentile thresh-
old is applied to the present climatology and then the corresponding absolute threshold is used in the future
(hence including both thermodynamic and dynamic factors)—to separate the dynamic and thermodynamic
effects. On the other hand, impacts are generally not considered in relative terms, and one must be careful in
this regard. Forthcoming work by the ARTMIP community will address this issue in depth by examining
ARTMIP methods under future climate scenarios, and data processing is already underway.
Another important set of clusters examined in this study is that of global and regional methods. One key
advantage of global methods is simply the global coverage of results, unlike regional methods, which are lim-
ited. Another, more speculative, advantage of global methods is that their development may benefit from
using a global perspective rather than a focus on one region, where ARs may frequently take on character-
istics not observed in most locales. In contrast, one key advantage of regional methods is that they are
specifically tuned to ARs and AR‐related impacts over a specific region and hence may be the most useful
for answering key science questions particular to those regions.
Afinal distinction made, qualitatively, throughout this study is that between less restrictive and more restric-
tive methods—a very subjective distinction based on their criteria, and usually only useful when comparing
one method to another. One key advantage of less restrictive methods is that they facilitate AR science and
impacts in regions where ARs are very rare using more restrictive methods (e.g., the usefulness of
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,797

“Gorodetskaya”over polar regions, and “Guan_Waliser”or “Rutz”over continental interiors). Of course,
the disadvantage is the reverse—less restrictive methods may result in the attribution of impacts to ARs,
when in fact the associated dynamics and vapor fluxes are very weak, or merely remnants of a once‐robust
AR. Researchers need to carefully weigh the advantages and disadvantages of their approach to answering a
given scientific question. It is important to remember that in this paper, distinctions between less restrictive
and more restrictive methods can really only be made within a cluster (work is underway to more objectively
quantify restrictiveness across clusters). For example, Figure 6 (middle) makes every method seem restric-
tive in comparison to the less restrictive relative “Guan_Waliser”method, when in fact there are also more
restrictive relative methods (e.g., “Payne”), less restrictive absolute methods (e.g., “Rutz”) and more restric-
tive absolute methods (e.g., “Tempest”). In the case of “Tempest”, greater restrictiveness arises because
objects are required to remain long and thin across time, whereas ARs tend to spread out as they encounter
land. Finally, one key advantage of using more restrictive methods is that they highlight only the strongest
events, which will likely (though not always) produce the most severe impacts.
A number of ARTMIP methods are based on the identification of features meeting certain geometric criteria,
throughout which either IVT ≥250 kg m
−1
s
−1
(e.g., “Brands,”“Gershunov,”“Rutz,”and “Tempest”)or
IWV ≥20 mm (e.g., “Goldenson,”“Hagos,”“Ralph,”and “Wick”). In addition, the recently developed AR
scale, described by Ralph et al. (2019), establishes IVT ≥250 kg m
−1
s
−1
as the minimum threshold required
to categorize an event as an AR. The AR science community increasingly recognizes the importance of water
vapor transport within ARs and now strongly favors IVT over IWV for diagnosing such features. Therefore,
IVT ≥250 kg m
−1
s
−1
seems to be a reasonable starting point. However, there are many cases in which IVT
of this magnitude will be primarily beneficial, and it becomes worthwhile to identify only stronger, more
hazardous ARs. This is one rationale for higher minimum thresholds such as 500 or 700 kg m
−1
s
−1
, as used
in “CONNECT500”and “CONNECT700,”respectively. The rationale for a higher minimum threshold can
also be climatologically and/or regionally based, as is the case for the very high background moisture field
over the southeastern United States, a region in which 500 kg m
−1
s
−1
was used by Mahoney et al. (2016;
results not available for this study). It should also be noted that IVT thresholds below 250 kg m
−1
s
−1
can
be useful both in regions with climatologically lower IVT, and in cases where long‐duration, low‐intensity
IVT events may produce significant impacts.
A number of relative methods use thresholds based on IVT ≥85th percentile of climatological IVT, along
with an absolute IVT threshold that serves as a floor, or minimum IVT requirement, to identify features
as ARs (e.g., “Guan_Waliser,”“Lavers,”“Payne,”“Ramos,”and “Viale”). Still other relative methods thresh-
old based on IVT exceeding daily climatology by some raw value such as 100 or 250 kg m
−1
s
−1
(e.g., “Lora”
and “Walton”), or use some other method (e.g. “Gorodetskaya,”“Mundhenk,”and “Shields”). Among these,
“Guan_Waliser”is the least restrictive due to a minimum IVT requirement of only 100 kg m
−1
s
−1
, and this
causes its AR frequency (and results directly associated with AR frequency) to be clear outliers in polar
regions and continental interiors where IVT is climatologically low. That said, it is an extremely useful out-
lier, because it often identifies regions downstream of mountain barriers as within an AR, whereas most
other methods do not. It can be argued, based on the AMS Glossary definition, that these regions are not
necessarily located within the spatial footprint of an AR. However, the usefulness is found in identifying
and attributing impacts to the ARs likely responsible for them, even if the spatial footprints of these ARs,
and their impacts, do not directly overlap. The “Gorodetskaya”method, from which results were not avail-
able for this study, also identifies ARs (and AR‐related impacts) in regions of very low IWV/IVT, having been
designed specifically to identify intrusions of anomalously moist air into polar regions. To be more consistent
with the AMS Glossary definition, such features could potentially be described as “decaying”ARs, or by
some other term, which indicates that they are no longer associated with the extratropical cyclones
and/or dynamic processes critical to their genesis.
Quantifying the uncertainty in AR‐related impacts (and how they may change in the future), most of which
are in some way related to precipitation, is a major motivation behind ARTMIP. Some sense of impacts can
likely be inferred from the results for AR climatology highlighted in this study. However, a more complete
assessment of the advantages and disadvantages associated with individual methods and with certain clus-
ters will be possible only after some quantification of the uncertainty in AR‐related precipitation takes place.
Future ARTMIP work plans to address this subject.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,798

Finally, ARTMIP expects to produce a number of new results and publications over the coming years. The
most salient of these is a pair of Tier 2 summary papers, which will present results from all ARTMIP methods
applied to output from a high resolution version of the Community Atmospheric Model (Wehner et al. 2014)
and available CMIP5 models under historical and RCP8.5 forcing scenarios. Numerous studies (e.g.,
Espinoza et al., 2018; Gao et al., 2016; Shields & Kiehl, 2016a, 2016b; Warner et al., 2015) have already exam-
ined changes in AR climate and impacts under climate change scenarios, but, as with studies of current AR
climate and impacts, these suffer from uncertainty that arises due to the usage of different AR identification
and tracking methods. The Tier 2 summary papers will quantify these uncertainties. In addition, ARTMIP
participants have already planned a number of studies on topics ranging from quantifying differences in
ARs based on the reanalysis product used to trends in ARs over time, and a variety of other topics.
5. Recommendations
The results presented in this study indicate a large degree of uncertainty in the climatological characteristics
of ARs resulting from differences in the methods used to identify and track them. This uncertainty is reduced
within “clusters”of methods that share similar approaches to AR identification and tracking, but even then,
uncertainty arises due to differences in thresholding variable and magnitude, geometric considerations, and
other criteria. As stated in the introduction, this should not be surprising—each method was developed to
answer a different question, and different answers naturally arise. This diversity benefits the community
in that it offers a wide variety of approaches to answering new questions that may arise. Nevertheless, the
AR science community will be interested in recommendations regarding which of these methods or clusters
best answers their questions.
Here, the authors provide generalized recommendations regarding the types of AR identification and
tracking methods that are most advantageous for certain applications, and ideas regarding future
method development.
The authors generally recommend absolute methods for studies focused on the relationship between ARs
and large‐scale atmospheric patterns, dynamic processes, and physical mechanisms for our current climate
in the midlatitudes (~30–60°N/S). The AR science community increasingly recognizes these features, parti-
cularly as they relate to extratropical cyclones, as playing a key role in defining the AR, as well as intense
water vapor transport (see the AMS Glossary definition). These methods have the advantage of being based
on fixed, observable thresholds (e.g., 20 mm of IWV or 250 kg m
−1
s
−1
of IVT), which are well suited to
answer questions related to AR dynamics such as their growth, maintenance, and decay. Hence, these meth-
ods are preferred at midlatitudes, where the dynamics associated with extratropical cyclones typically drive
the life cycle of ARs. Absolute methods are also recommended for most weather forecasting applications
because they are straightforward and intuitive, though forecasts based on anomalies, percentiles, or return
intervals can be very effective in communicating to more knowledgeable audiences. Finally, it must be
remembered that AR‐related impacts can occur well outside the spatial footprint outlined by absolute meth-
ods, particularly when the threshold used is more restrictive.
The authors generally recommend relative methods for studies focused on attributing a wide variety of
hazards (e.g., heavy rainfall, flooding, and wind) to ARs. These methods have the advantage, as was
described at length above in reference to the “Guan_Waliser”method, of placing regions far from the core
of an AR within the spatial footprint of the AR, which facilitates the attribution of impacts within that foot-
print to the AR itself. These methods are preferred at tropical and subtropical latitudes, because they often
more effectively filter the broad regions of IVT that occur throughout the tropics. They are also preferred at
polar latitudes, because they often identify features that are climatologically anomalous, despite having low
IWV or IVT values relative to midlatitude features. Hence, these methods are well suited to answer questions
related to the occurrence and impacts of climatologically anomalous moisture surges around the globe, but it
must be remembered that some results may not translate well from one region to another.
The authors, at this time, defer recommendations regarding studies involving climate change until the Tier 2
analyses have been completed. It is likely that there will be advantages and disadvantages to both absolute
and relative methods, just as there will be to both sides of any number of other cluster pairs.
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL. 13,799

The authors recommend more restrictive methods for studies focusing on dynamic processes related to the
core of the AR and for studies focused on a subset of generally stronger ARs, which more restrictive methods
will select. For studies related to the attribution of precipitation and other impacts to ARs, the authors
recommend carefully considering the goals and objectives of each study. One advantage of less restrictive
methods is that identified ARs are associated with larger spatial footprints, which aid in evaluating all
impacts potentially related to ARs. However, these larger footprints can cause the average impacts associated
with ARs to be quite low, which could prove misleading to the public. Hence, more restrictive methods may
be more suitable for highlighting the extreme impacts that occur typically along the axis of greatest IVT
within an AR.
The authors also recommend that studies focusing on a particular region should consider basing their ana-
lyses on regional methods developed to assess that region. These methods may take into account important
regional characteristics of ARs and their impacts, whereas those focused on other regions, or global methods,
may not. Of course, previously developed methods will not be sufficient to answer certain questions, and yet
other studies might benefit from using global methods, which facilitate comparisons between different parts
of the world.
The authors recommend a few directions for future work in the area of AR identification and tracking
method development that would benefit the AR science community. The AMS Glossary defines an AR as
“a long, narrow, and transient corridor of strong horizontal water vapor transport that is typically associated
with a low‐level jet stream ahead of the cold front of an extratropical cyclone.”The AR science community is
increasingly recognizing the key role of dynamic processes in terms of defining the AR (personal correspon-
dence with participants and observation of presentations at 2018 International Atmospheric Rivers
Conference). Hence, the first recommendation is to emphasize methods based on IVT, which incorporates
wind as well as moisture, over those based on IWV. In this regard, wind serves as a proxy variable for a
number of dynamic processes, and is fundamental to the moisture transport associated with ARs. The
second recommendation is development of an interactive online tool that allows researchers to compare
multiple methods, along with other relevant layers (e.g., IWV, IVT, geopotential height, temperature, and
precipitation), in real time. This tool would assist researchers in determining which methods are most useful
for their specific applications. The third recommendation is that future work should consult the ARTMIP
archive and literature provided online. This provides future studies with the contextual background of what
methods already exist, which studies have been performed, and how new results best fit into this emerging
field of study.
ARTMIP has produced, and will continue to produce, an astonishing quantity of data that can be mined to
improve our understanding of ARs and their impacts. The authors anticipate, and indeed already plan, a
number of studies that will address various topics, and others are encouraged to do the same. The general-
ized recommendations above are the authors' best guidance at this time, but they are by no means a panacea.
While we think that these recommendations are useful, the history of science suggests that the most inter-
esting results will arise only when they are ignored. We welcome those developments.
Data Availability Statement
All ARTMIP data (including the MERRA‐2 source data) are available from the Climate Data Gateway,
DOI:10.5065/D6R78D1M (ARTMIP Tier 1 Catalogues), DOI: 10.5065/D62R3QFS (MERRA‐2 source data).
References
Albano, C. M., Dettinger, M. D., & Soulard, C. E. (2017). Influence of atmospheric rivers on vegetation productivity and fire patterns in the
southwestern U.S. Journal of Geophysical Research: Biogeosciences,122(2), 308–323. https://doi.org/10.1002/2016JG003608
Alexander, M. A., Scott, J. D., Swales, D., Hughes, M., Mahoney, K., & Smith, C. A. (2015). Moisture pathways into the U.S. Intermountain
West Associated with heavy winter precipitation events. Journal of Hydrometeorology,16(3), 1184–1206. https://doi.org/10.1175/JHM‐
D‐14‐0139.1
Ault, A. P., Williams, C. R., White, A. B., Neiman, P. J., Creamean, J. M., Gaston, C. J., et al. (2011). Detection of Asian dust in California
orographic precipitation. Journal of Geophysical Research,116(D16), D16205. https://doi.org/10.1029/2010JD015351
Brands, S., Gutiérrez, J. M., & San‐Martín, D. (2017). Twentieth‐century atmospheric river activity along the west coasts of Europe and
North America: Algorithm formulation, reanalysis uncertainty and links to atmospheric circulation patterns. Climate Dynamics,
48(9‐10), 2771–2795. https://doi.org/10.1007/s00382‐016‐3095‐6
10.1029/2019JD030936
Journal of Geophysical Research: Atmospheres
RUTZ ET AL.
Acknowledgments
ARTMIP is a grassroots community
effort and includes a collection of
international researchers from
universities, laboratories, and agencies.
Cochairs and committee members
include Jonathan Rutz, Christine
Shields, L. Ruby Leung, F. Martin
Ralph, Michael Wehner, Ashley Payne,
and Travis O'Brien. Details on
catalogues developers can be found on
the ARTMIP website. ARTMIP has
received support from the U.S.
Department of Energy Office of Science
Biological and Environmental Research
(BER) as part of the Regional and
Global Climate Modeling program, and
the Center for Western Weather and
Water Extremes (CW3E) at Scripps
Institute for Oceanography at the
University of California, San Diego.
13,800

Cordeira, J. M., Ralph, F. M., Martin, A., Gaggini, N., Spackman, J. R., Neiman, P. J., et al. (2017). Forecasting atmospheric rivers during
CalWater 2015. Bulletin of the American Meteorological Society,98(3), 449–459. https://doi.org/10.1175/BAMS‐D‐15‐00245.1
Dettinger, M. D., Ralph, F. M., Das, T., Neiman, P. J., & Cayan, D. R. (2011). Atmospheric rivers, floods, and the water resources of
California. Water,2011(3), 445–478.
Dettinger, M. D., Ralph, F. M., & Rutz, J. J. (2018). Empirical return periods of the most intense vapor transports during historical
atmospheric river landfalls on the U.S. West Coast. Journal of Hydrometeorology,19(8), 1363–1377. https://doi.org/10.1175/JHM‐D‐17‐
0247.1
Espinoza, V., Waliser, D. E., Guan, B., Lavers, D. A., & Ralph, F. M. (2018). Global analysis of climate change projection effects on atmo-
spheric rivers. Geophysical Research Letters,45(9), 4299–4308. https://doi.org/10.1029/2017GL076968
Gao, Y., Lu, J., & Leung, L. R. (2016). Uncertainties in projecting future changes in atmospheric rivers and their impacts on heavy preci-
pitation over Europe. Journal of Climate,29(18), 6711–6726. https://doi.org/10.1175/JCLI‐D‐16‐0088.1
Gelaro, R., McCarty, W., Suárez, M. J., Todling, R., Molod, A., Takacs, L., et al. (2017). The Modern‐Era Retrospective Analysis for Research
and Applications, Version 2 (MERRA‐2). Journal of Climate,30(14), 5419–5454. https://doi.org/10.1175/jcli‐d‐16‐0758.1
Gershunov, A., Shulgina, T., Ralph, F. M., Laver, D. A., & Rutz, J. J. (2017). Assessing the climate‐scale variability of atmospheric rivers
affecting western North America. Geophysical Research Letters,44(15), 7900–7908. https://doi.org/10.1002/2017GL074175
Goldenson, N., Leung, L. R., Bitz, C. M., & Blanchard‐Wrigglesworth, E. (2018). Influence of atm ospheric rivers on mountain snowpack in
the Western United States. Journal of Climate,31(24), 9921–9940. https://doi.org/10.1175/JCLI‐D‐18‐0268.1
Gorodetskaya, I. V., Tsukernik, M., Claes, K., Ralph, M. F., Neff, W. D., & Van Lipzig, N. P. M. (2014). The role of atmospheric rivers in
anomalous snow accumulation in East Antarctica. Geophysical Research Letters,41(17), 6199–6206. https://doi.org/10.1002/
2014GL060881
Guan, B., & Waliser, D. E. (2015). Detection of atmospheric rivers: Evaluation and application of an algorithm for global studies. Journal of
Geophysical Research: Atmospheres,120(24), 12514–12535. https://doi.org/10.1002/2015JD024257
Hagos, S. M., Leung, L. R., Yoon, J.‐H., Lu, J., & Gao, Y. (2016). A projection of changes in landfalling atmospheric river frequency and
extreme precipitation over western North America from the large ensemble CESM simulations. Geophysical Research Letters,43(3),
1357–1363. https://doi.org/10.1002/2015GL067392
Hatchett, B. J., Burak, S., Rutz, J. J., Oakley, N. S., Bair, E. H., & Kaplan, M. L. (2017). Avalanche fatalities during atmospheric river events
in the Western United States. Journal of Hydrometeorology,18(5), 1359–1374. https://doi.org/10.1175/JHM‐D‐16‐0219.1
Huning, L. S., Margulis, S. A., Guan, B., Waliser, D. E., & Neiman, P. J. (2017). Implications of detection methods on characterizing
atmospheric river contribution to seasonal snowfall across Sierra Nevada, USA. Geophysical Research Letters,44(20), 10,445–10,453.
https://doi.org/10.1002/2017GL075201
Lamjiri, M. A., Dettinger, M. D., Ralph, F. M., & Guan, B. (2017). Hourly storm characteristics along the U.S. West Coast: Role of atmo-
spheric rivers in extreme precipitation. Geophysical Research Letters,44(13), 7020–7028. https://doi.org/10.1002/2017GL074193
Lavers, D. A., & Villarini, G. (2013). Atmospheric Rivers and flooding over the central United States. Journal of Climate,26(20), 7829–7836.
https://doi.org/10.1175/JCLI‐D‐13‐00212.1
Lavers, D. A., Villarini, G., Allan, R. P., Wood, E. F., & Wade, A. J. (2012). The detection of </