AMTAtmospheric Measurement TechniquesAMTAtmos. Meas. Tech.1867-8548Copernicus PublicationsGöttingen, Germany10.5194/amt-9-1637-2016An automatic precipitation-phase distinction algorithm for optical disdrometer data over the global oceanBurdanowitzJörgjoerg.burdanowitz@mpimet.mpg.dehttps://orcid.org/0000-0002-0540-4034KleppChristianBakanStephanMax Planck Institute for Meteorology, Bundesstraße 53, Hamburg, GermanyUniversity of Hamburg, CliSAP/CEN, Bundesstraße 55, Hamburg, GermanyJörg Burdanowitz (joerg.burdanowitz@mpimet.mpg.de)13April2016941637165227November201521December201510March201628March2016This work is licensed under a Creative Commons Attribution 3.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by/3.0/This article is available from https://amt.copernicus.org/articles/9/1637/2016/amt-9-1637-2016.htmlThe full text article is available as a PDF file from https://amt.copernicus.org/articles/9/1637/2016/amt-9-1637-2016.pdf
The lack of high-quality in situ surface precipitation data over the global
ocean so far limits the capability to validate satellite precipitation
retrievals. The first systematic ship-based surface precipitation data set
OceanRAIN (Ocean Rainfall And Ice-phase precipitation measurement Network)
aims at providing a comprehensive statistical basis of in situ precipitation
reference data from optical disdrometers at 1 min resolution deployed on
various research vessels (RVs). Deriving the precipitation rate for rain and
snow requires a priori knowledge of the precipitation phase (PP). Therefore,
we present an automatic PP distinction algorithm using available data based
on more than 4 years of atmospheric measurements onboard RV
Polarstern that covers all climatic regions of the Atlantic Ocean. A
time-consuming manual PP distinction within the OceanRAIN post-processing
serves as reference, mainly based on 3-hourly present weather information
from a human observer. For automation, we find that the combination of air
temperature, relative humidity, and 99th percentile of the particle diameter
predicts best the PP with respect to the manually determined PP. Excluding
mixed phase, this variable combination reaches an accuracy of 91 % when
compared to the manually determined PP for 149 635 min of precipitation
from RV Polarstern. Including mixed phase (165 632 min), an
accuracy of 81.2 % is reached for two independent PP distributions with a
slight snow overprediction bias of 0.93. Using two independent PP
distributions represents a new method that outperforms the conventional
method of using only one PP distribution to statistically derive the PP. The
new statistical automatic PP distinction method considerably speeds up the
data post-processing within OceanRAIN while introducing an objective PP
probability for each PP at 1 min resolution.
Introduction
Rare and often low-quality gauge-based surface reference data sets challenge
the in situ validation of oceanic precipitation as observed by passive and
active microwave satellite sensors (;
). Over land, radar and gauge-based
precipitation monitoring networks cover a large fraction of the land surface
for more than 2 decades, which qualifies them to validate precipitation
satellite estimates . However, the ocean surface
lacks dense long-term in situ precipitation monitoring networks. Furthermore,
existing coastal and island-based precipitation measurements may not fully
represent oceanic precipitation because the measured particle size
distributions (PSDs), precipitation rates, and accumulations may differ from
those measured over the open ocean . However,
found no difference between PSDs over coastal
areas and ocean. Most existing in situ oceanic precipitation data sets sample
measurements from low-quality rain gauges on voluntary observing ships (VOSs;
) or buoy arrays .
Many of these in situ ocean data sets include present weather observations but
lack quantitative estimates of precipitation. The high-latitude ocean
completely lacks precipitation measurements that sample solid and mixed-phase
precipitation. However, recent and future precipitation satellite estimates
demand high-quality in situ quantitative precipitation estimates including
snowfall over the global ocean.
The large uncertainty in precipitation gauge measurements arises from the
rough open-ocean conditions that complicate precipitation monitoring. Under
high wind speed, standard rain gauges with horizontal catchment surfaces face
a large undercatch (;
). In the extratropics,
mixed-phase and solid precipitation cause further difficulties strongly
adding to the undercatch of
horizontally oriented measuring surfaces. In contrast, optical instruments
with a vertically oriented measuring surface such as disdrometers perform
better at capturing precipitation under high wind speeds, though varying wind
directions are challenging. Optical disdrometers are thus denoted as the
reference in situ instrument to measure precipitation
.
To provide systematic high-quality in situ precipitation data over the ocean,
the long-term Ocean Rainfall And Ice-phase precipitation measurement Network
(OceanRAIN; )
applies automatic optical disdrometers of type ODM470 that are deployed
onboard sea-going research vessels (RVs) for operation in all climatic
regions. The ODM470 was developed to measure under high wind speed and
frequently varying wind directions. Its cylindrical measuring volume ensures
being independent from the wind-driven incidence angle of the falling
hydrometeors while a wind vane keeps the measuring volume perpendicular to
the instantaneous wind direction. The ODM470 accuracy lies within a range of
3 % rain accumulation limited to rainfall at various wind conditions with
respect to an improved ship rain gauge including side collectors on RV
Alkor
on the Baltic Sea . Compared to an ANS410
WMO-reference rain gauge over land , the ODM470
deviates by 2 % under low wind speed . For
snow, a predecessor of the current ODM470 agreed with the observer's log
during the Lofoten Cyclones campaign (LOFZY;
)
in detecting snowfall events. More recent results for measuring snow and
mixed-phase precipitation are expected soon from the WMO Solid Precipitation
InterComparison Experiment (SPICE) at Marshall field site in Boulder (CO,
USA), where the ODM470 was compared against a multitude of in situ solid
precipitation instruments for more than 2 years. The ODM470 suits well to
measure various types of precipitation under open-ocean conditions onboard
sea-going RVs.
The deployment of the ODM470 on several RVs allows to sample OceanRAIN
precipitation data from all climate zones including the cold-season high
latitudes. This requires a precipitation-phase (PP) distinction between rain,
snow, and mixed phase in order to provide correct precipitation rates for
disdrometer-measured PSDs. The PP information usually originates from human
observers' reports saved in the WMO present weather code ww.
Efforts to automatize present weather observations impose high requirements
on instruments such as present weather sensors. Automated present weather
sensors encounter problems at temperatures around 0 ∘C as well as for
light precipitation and small particle sizes .
High wind speed also complicates the PP determination because the wind speed
strongly interferes with the particle fall speed that solely carries the PP
information. Thus, most studies to distinguish PPs limit the wind conditions
to low wind speed or calm conditions
(;
;
). Only few studies apply more
sophisticated instruments that use articulating particle size velocity
(PARSIVEL) disdrometers to account for wind effects and thus directly derive
the PP from the particle fall speed
. More simply constructed
instruments such as the ODM470 require ancillary data to determine the PP.
In OceanRAIN, we aim to replace the so far manual PP distinction method by an
automatic algorithm for three main reasons. First, the manual method consumes
a considerable amount of time and workforce because the 1 min
precipitation data requires visual inspection of air temperature, present
weather observations, and theoretical rain and snow rate. Second, the
human-based PP decision based on visual data inspection lacks objectivity
while the decision itself remains non-transparent to the user. Third,
temporal gaps exist between the 3-hourly present weather observational
timesteps, especially during nighttime adding to the uncertainty. Currently,
no measures of this PP uncertainty are provided in the manual method. For
these reasons, we present a new automatic PP distinction algorithm including
a PP probability for OceanRAIN precipitation data that is also applicable to
all other instruments sampling PSDs of precipitation.
The new PP distinction algorithm follows a statistical approach guided by
many other studies that relate atmospheric predictors to the PP
(; ; ; ). Most previous work focuses on PP
distinction over land only, while we introduce a new PP distinction algorithm
over the ocean. compares ocean and land areas
using a relatively coarse time step of few to several hours depending on
availability of observations. In contrast, OceanRAIN offers atmospheric
measurements at 1 min resolution while present weather observations are
limited to 3-hourly timesteps during daytime only. These high-resolution
ancillary data from the RV combined with PSD data from the optical
disdrometer enable a more accurate and reliable PP distinction.
Section introduces the optical disdrometer, the manual PP
distinction method, and the OceanRAIN data set in detail. Section
presents different atmospheric variable combinations and methods to predict
the PP. In Sect. , one PP distribution distinguishes between two
PPs, while in Sect. one PP distribution distinguishes between
three PPs. Section introduces a newly developed method to predict
three PPs using two PP distributions. Section discusses the
results by comparing with similar studies. Section completes our
investigations with a summary and concluding remarks.
Data and methods
Since 2010, OceanRAIN collects atmospheric data including precipitation rates
on several RVs. Current permanent deployments include the German ships RV
Polarstern (since June 2010), RV Meteor (since March 2014), RV Sonne
(since November 2014), and the Russian ship Akademik Ioffe (since September 2010). The
backbone of OceanRAIN is the optical disdrometer ODM470, which is explained
in detail in Sect. . Section introduces the
manual method that has been used so far to retrieve the PP in OceanRAIN.
These manually determined PPs function as a benchmarking reference for the
new automatic PP distinction algorithm. For the algorithm development, we
exclusively use RV Polarstern data (Sect. ) that
contains a high fraction of high-latitude solid and mixed-phase precipitation
being a prerequisite to develop a robust PP distinction algorithm. While
describes the OceanRAIN data post-processing and
quality checking before PP distinction we focus on presenting a new automatic
PP estimation method that provides uncertainty information.
The ODM470
The ODM470 is an optical disdrometer to measure precipitation, manufactured
by the German company Eigenbrodt GmbH & Co KG near Hamburg (Germany). The
instrument consists of an infrared (IR) light-emitting diode (LED) at
880 nm and a photo diode receiver .
The IR-LED of the ODM470 is only activated once at least 8 particles per
minute pass the active sensing area of the precipitation detector IRSS88
(Fig. , right) in order to increase IR-LED lifetime and exclude
measurement artifacts caused by birds or other non-precipitation objects. The
IRSS88 switches off the ODM470 after 1 min without any particle passing
the IRSS88 active sensing area. The entire ODM470 system was developed in a
way to minimize undesired influences by changing wind directions and high
wind speed. The ODM470 sensitive optical volume has a cylindrical shape of
120 mm length and 22 mm in diameter. The cylindrical shape guarantees an
independence from the incidence angle of the falling hydrometeors, which
becomes crucial under high wind speeds and superstructure-induced local
turbulence. Mounted on a pivotable axis, a wind vane ensures the optical
volume to adjust perpendicular to the instantaneous local wind direction. The
ODM470 mounting height typically ranges between 30 and 45 m height,
depending on the RV's specifications (Fig. ). This elevated
deployment reduces influences on the measured precipitation by splashing wave
water.
The image displays the automatic ODM470 measurement system
including a cup anemometer, the optical disdrometer ODM470, and the
precipitation detector IRSS88, deployed in the highest main mast at about
43 m height onboard RV Polarstern.
During precipitation events, the falling hydrometeors attenuate the emitted
IR radiation, which decreases the voltage signal measured. The duration of
the voltage drop determines the particle transit time, that is, the total time
it takes a particle to pass through the optical volume of the disdrometer.
From the amplitude of the detector voltage drop the cross-sectional area can
be deduced, which determines the particle diameter. The measured particle
diameters are split into 128 logarithmically distributed size bins, of
which
the smallest is less than 0.02 mm and the largest corresponds to the
optical volume diameter of 22 mm. However, wind- or wave-induced ship
vibrations passed on to the instrument might cause artificial signals that
are not distinguishable from real precipitation, which is why particles below
bin 14 (0.43 mm diameter) are not considered in OceanRAIN. This exclusion
of small particles also removes sea spray particles from the PSD spectra. The remaining particles are accumulated and
binned over 1 min. From the resulting PSD, the precipitation rate PR
(mm h-1) or liquid water equivalent (kg m-2 h-1) after
can be calculated using
PR=3600⋅∑bin=1128n(bin)⋅v(bin)⋅m(bin),
where v(bin) (m s-1) denotes the particle terminal fall speed and
m(bin) (kg) the particle mass; both are parameterized. n(bin)
(m-3)
denotes the PSD density per bin class that is calculated following
by considering the geometrical
features, diameter d (m) and length l (m), the sampling time t (s) of
the ODM470 as well as the sum of local wind speed Urel
(m s-1)
relative to the ship movement measured by a cup anemometer, and the empirical
terminal fall speed v(bin) (m s-1) as
n(bin)=N(bin)l⋅d⋅t⋅Urel2+[v(bin)]2.N(bin) is the number of measured particles per bin class, denoted as PSD.
As explained, in Eq. () empirical relationships utilize the particle
diameter D that strongly depends on the type of precipitation. Henceforth,
we refer to precipitation phase (PP), which means either liquid precipitation
(rain), solid precipitation (e.g., snow or graupel), or mixed-phase
precipitation. For rain the drop mass ml (kg), or liquid water content,
and the particle terminal fall speed vl (m s-1) are well known and
calculated using Eqs. () and () from
, respectively.
ml=1000⋅43π⋅(0.5D)3vl=9.65-10.3⋅e-600D
For snow, the measured cross-sectional area differs from the required maximum
dimension of the particle due to the non-spherical shape of snowflakes. This
difference requires applying a transfer function. However,
found that the product of particle
terminal fall speed and particle mass (liquid water equivalent) as a function
of cross-sectional area is in the same order of magnitude for various frozen
precipitation particle types. Hence, no transfer function between
cross-sectional area and maximum diameter is required when using a spherical
lump graupel assumption. The lump graupel assumption works well for frozen
precipitation particles between 0.4 and 9 mm in diameter, whereas particles
exceeding 9 mm in diameter rarely occur. Nevertheless, events with large
particles introduce larger errors to the estimate in the same way as the
retrieval quality may largely differ for individual snowfall events. Overall,
no unique snowfall retrieval can be derived using optical disdrometers
without recording the individual particle shape. Compared to a Geonor gauge,
the optical disdrometer agreed well in most cases and overestimated a few
light snowfall cases during the 1999/2000 winter period at Uppsala
. Following the lump graupel
approximation by , particle mass ms
(Eq. ) and particle terminal fall speed vs (Eq. ) are
calculated empirically as
ms=1.07⋅10-5⋅(100D)3.1,vs=7.33⋅(100D)0.78. observed lump graupel being the most frequently
occurring precipitation type over the cold-season Norwegian Sea during the
LOFZY campaign. discuss several
sources of error for a snow-measuring PARSIVEL whereof those for particle
shape and orientation, margin effects, and coinciding particles also apply to
the ODM470. However, the PARSIVEL is more sensitive to influences by wind
speed and wind direction on the falling precipitation particles because the
PARSIVEL has a fixed non-pivotable horizontal optical sensing area.
For mixed-phase precipitation, we generally use the snow retrieval
(Eqs. and ) to calculate the precipitation rate within
OceanRAIN because the absolute error of treating rain drops like snow
particles, and thus underestimating the precipitation rate, results in a
smaller error than vice versa. In more than 90 % of the precipitating cases
from RV Polarstern in OceanRAIN the precipitation rate calculated
with Eqs. () and () (theoretical rain rate) exceeds precipitation
rate calculated with Eqs. () and () (theoretical snow rate) by a
factor of 50 to 200. Accordingly, this large difference might cause large
biases in the precipitation rate for misclassified PP. Correctly classified
mixed-phase precipitation events might still strongly underestimate the
precipitation rate if the instantaneous rain fraction sharply exceeds 0.5.
The minute-aggregated fraction of liquid and solid particles cannot be
identified by the ODM470 and would require ancillary data such as a video
disdrometer. More details on the instrumentation can be found in
while algorithm features are explained
in .
Translation of WMO present weather codes ww
into the three PPs from ,
, and OceanRAIN. ww codes printed
in bold can be translated into multiple PPs in OceanRAIN depending on
ancillary data. “Indet./hail” denotes indeterminate precipitation or hail
used for classification in .
Though time-consuming, the manual PP distinction was so far employed to
determine the PP that is required to calculate the precipitation rate.
Because we apply this manual PP distinction data as reference to the new
automatic PP distinction algorithm, a detailed explanation follows. If
available, shipboard present weather observations stored in the WMO standard
meteorological present and past weather code ww are translated
into the three PPs: rain, snow, and mixed phase according to
, displayed in Table .
However, the translation of ww codes into a PP partly differs between
OceanRAIN and . While
assigns one single PP to each of the
ww codes, OceanRAIN allows multiple PPs for a single ambiguous ww code
(bold weather codes in Table ). Instead,
lists ambiguous ww codes in a
category called “indeterminable” (abbreviated “Indet.” in
Table ) that, however, includes no PP information anymore. For
that reason we deviate from this procedure to retain as much PP information
as possible. Another difference concerns ww codes for all kinds of freezing
rain (i.e., rain at freezing temperatures) without snow that
classifies as mixed phase. Classifying
freezing rain as mixed phase by applying the lump graupel assumption
(Eqs. and ) leads to an underestimated precipitation rate.
This underestimation arises because falling raindrops freeze only after
passing the disdrometer's optical volume when hitting any obstacle, which is
why we consider freezing rain cases in OceanRAIN as rain (ww=56,57,66,67). Likewise, we assign a snow flag to ice pellets (ww=79) as well as
mostly to hail (ww=89,90), graupel (ww=87,88), and combinations of both
(ww=93,94,96,99). The aim is mainly to separate frozen (solid) from
non-frozen (liquid) precipitation particles to account for differences in
density and cross-sectional area that affect Eqs. () to () and
hence the precipitation rate. In contrast, the study by
concentrates exclusively on clear rain,
snow, and mixed-phase observations (Table ) by neglecting
drizzle, freezing rain, and ice pellets, among others. In general, assigning
the correct PP for a given ambiguous ww code requires visual inspection of
PSDs and ancillary data collected onboard the RV.
The ww code from shipboard observations on RV Polarstern is
available 3 hourly during daytime only. Nighttime conditions and PP changes
between two consecutive 3-hourly observational time steps require
ancillary data from the RV to derive the PP. By ancillary data we refer to
atmospheric variables measured onboard the ship including the ODM470, such as
air temperature, humidity, and precipitation rate. These ancillary data are
available at a much higher resolution of 1 min compared to the 3-hourly
observations. Initially, we assign the PP derived from the ww code directly
to every single minute of precipitation that follows a 3-hourly observation
as a first-guess information. If available, air temperature as one of the
ancillary data serves to possibly correct this first-guess PP. For
near-freezing air temperatures, the manual procedure calculates the
precipitation rate after Eq. () for rain (Eqs. and ) and snow (Eqs. and ) assumption separately.
Large differences between theoretical rain and snow rate can help to identify
a plausible PP. However, if both theoretical rain and snow rate differ by
much less than 2 orders of magnitude, their influence on the PP decision
becomes negligible, which makes the PP more arbitrary. Accordingly, the
manual method might be biased by the worker who decides for a PP and the
observer on the RV. For these reasons, we aim at developing an automatic PP
distinction algorithm at 1 min resolution that statistically derives a PP
from atmospheric measurements.
OceanRAIN data from RV Polarstern
Map illustrates ship tracks from RV Polarstern ALL data
(11 June 2010–8 October 2014), whereby dots denote minutes of occurring
precipitation classified by the manual PP distinction (cyan: rain; orange:
mixed phase; purple: snow). Harbor times and minutes without precipitation
are not shown. Left side denotes the fraction of each PP averaged per
latitude.
The manual PP estimation has been applied to more than 4 years of
OceanRAIN data from RV Polarstern (11 June 2010–8 October 2014).
This period consists of several expeditions to Arctic and Antarctic regions.
In addition to the high latitudes, RV Polarstern collected
precipitation data from the tropics and subtropics when crossing the equator
in the Atlantic Ocean six times (Fig. ). The whole measuring period
amounts to more than 268 000 min of precipitation excluding periods of
maintenance in harbors and instrument outages. Snow or mixed-phase
precipitation occurred almost exclusively poleward of 45∘ S and
70∘ N, which largely depends on seasonality. During boreal warm
season, RV Polarstern sailed over the northern hemispheric Atlantic
Ocean and in the entire Arctic area, whereas during austral warm-season RV
Polarstern cruised on the southern hemispheric Atlantic Ocean and at
the Antarctic. As an exception, RV Polarstern spent the whole year
2013 including austral cold season in the Southern Hemisphere, which explains
the multitude of mixed-phase and snow precipitation cases between
45 and 75∘ S that are not sampled at corresponding
northern hemispheric latitudes. For the sake of polar research, RV
Polarstern spends most research time in the polar regions, which
results in a high time fraction of snow or mixed-phase precipitation of
0.57 (Table ). Accordingly, precipitation occurred most
frequently at temperatures around 0 ∘C and at high relative humidity
(Fig. ). The high time fraction of snow or mixed-phase
precipitation around 0 ∘C makes RV Polarstern an extremely
valuable data set for oceanic PP distinction analysis.
OceanRAIN data sets from RV Polarstern divided into sub-data sets that are used in the analysis. While RSM (rain, snow, mixed phase) and ALL (all data)
include the mixed phase, RS (rain, snow) sub-data exclude mixed-phase precipitation. RSM and RS
include only those minutes with at least 20 particles of precipitation
falling at mid- or high latitudes at air temperatures around the freezing
point (see Sect. ). The no-rain fraction (rain fraction
subtracted from 1) yields the fraction frozen precipitation meaning snow
cases for RS and snow and mixed phase for RSM and ALL.
The whole RV Polarstern data set, denoted ALL (for all data),
consists of about 268 000 min of precipitation. The ship's positions cover
large areas of distinctly high or low temperatures where the PP assignment is
trivial and does not help in developing the PP algorithm. Therefore, we
reduce the complete RV Polarstern data set ALL to minutes of highest
PP uncertainty (Table ). Air temperatures below -6 ∘C
and above 8 ∘C are excluded as well as ship locations between
45∘ S and 70∘ N latitude wherein virtually no solid or
mixed-phase precipitation was observed within the 4-year period
(Fig. ). We exclude minutes with a total particle number of less
than 20 particles because they cannot guarantee a meaningful PP
distinction. These limitations leave a subset of data denoted RSM (for rain,
snow, mixed phase) with 165 632 min of rain, snow, or mixed-phase
precipitation. By that, the no-rain time fraction including snow or
mixed-phase precipitation increases from 0.57 (ALL) to 0.61 (RSM). If we
further exclude mixed-phase precipitation the gained sub-sample, denoted RS
(for rain, snow), reduces to 149 635 min while the no-rain fraction
decreases to 0.57 (Table ).
Two-dimensional histogram shows relative occurrence (%) for each PP (top:
snow; middle: mixed phase; bottom: rain) after manual PP distinction from
OceanRAIN RSM data set of RV Polarstern. n denotes the number of
minutes used per PP (165 632 in total).
Atmospheric variables measured onboard RV Polarstern include
temperature-related (T, Td, T2h) and humidity-related variables
(rH, Td), air pressure (P), and, from the ODM470, wind speed (not used
for analysis) and particle diameter (D). Instead of D, we use the 99th
percentile of D, D99, which is a measure for the maximum particle
diameter measured within 1 min but excluding erroneously large particles
possibly caused by particle coincidences, drip-off drops, or other artifacts.
Table lists all relevant variables from RV Polarstern
and the ODM470. Note that all variables are measured distinctly higher than
2 m above the surface at about 43 m in order to reduce interfering sea spray
and splashing wave water.
List of available meteorological predictor variables in OceanRAIN used in the logistic regression model to predict the PP.
VariableDescriptionUnitSourceTAir temperature∘CPolarsternTdDew point temperature∘CPolarsternT2hAir temperature 2 h prior to observation∘CPolarsternrHRelative humidity%PolarsternPSea-level air pressurehPaPolarsternRRPrecipitation rate assuming rainmm h-1ODM470D9999th percentile of particle diametermmODM470The automatic PP distinctionOne PP distribution to predict two PPs (2P1D)
This study aims at predicting the PP automatically by using available in situ
atmospheric predictor variables (Table ). While we first focus
on predicting two PPs using one PP distribution (Sect. ; 2P1D), we
later apply one PP distribution to predict three PPs (incl. mixed phase;
Sect. ; 3P1D). Section presents a novel approach that
predicts three PPs applying two PP distributions (3P2D).
For the PP prediction we adopt a statistical model using logistic regression
to relate the available observed atmospheric variables (predictor variables)
to the PP as suggested by , henceforth KS98. The
predictor variables are fitted against binary dependent variables to
calculate the PP probability p(PP). Taken from the manual PP distinction
data (Sect. ), the binary dependent variables attain a rain
probability p(rain) [frac] of either 0 (snow) or 1 (rain). Once fitted,
p(rain) can attain any value between 0 and 1 depending on the predictor
variables. p(rain) is calculated by
p(rain)=11+eα+β⋅V1+γ⋅V2+…+ω⋅Vn,
whereby Vi represents the atmospheric predictor variables. α, β, γ,
…, ω denote the regression coefficients that are determined by
minimizing the sum of squared errors (nearest-neighbor method) with respect
to the PPs from the manual PP distinction. Generally, we use the term PP
probability, p(PP), representing both rain (p(rain)) and snow probability
(p(snow)) if not stated differently. The snow probability is calculated as
1-p(rain), excluding mixed phase for now in this simple model.
We calibrate various combinations of atmospheric predictor variables Vi
(Table ) for RS sub-data (Table ) to find the
combination that predicts best the PP. KS98 state that the combination of air
temperature T and relative humidity rH, called T_rH, is
suited best to predict the PP. For T_rH, Eq. () changes
to
p(rain)=11+e(α+β⋅T+γ⋅rH),
where the number of regression coefficients reduces to three. In lack of
alternative reference data, we evaluate the calculated regression
coefficients of RS sub-data using the same manually determined PPs as used
for the model calibration. Nevertheless, we investigated the robustness of
the regression coefficients using 100 realizations of only 50 % randomly
chosen minutes of precipitation from the RS data set. The standard deviation
of the 100 realizations rarely exceeds 10 % of the individual regression
coefficients from the whole RS data set, which confirms the robustness of the
calculated regression coefficients. If in the manual PP reference data set a
minute of precipitation is assigned rain, the statistical model by definition
“agrees” for p(rain)≥0.5 while it “disagrees” for p(rain)<0.5. For the
rain/snow distinction four possible combinations exist – rain agreement,
snow agreement, rain disagreement, and snow disagreement. For instance, rain
disagreement means that the statistical model predicts rain that disagrees
with the manual PP reference data indicating snow. Combined in a contingency
table we choose four scores to evaluate how well the atmospheric predictor
variable combinations serve to predict the PP in this statistical model.
First, the accuracy serves to evaluate the overall correctness of the
predictor variable combinations with respect to the manual PP reference
data set. The accuracy represents the sum of cases in which model and manually
determined PP reference data of RS sub-data agree divided by the total number
of minutes in RS sub-data. Ideally, the accuracy is close to 1.
Second, we consider the bias score defined as the ratio between the sum of
disagreeing rain predictions and agreeing rain predictions to the sum of
disagreeing snow predictions and agreeing rain predictions, all with respect
to the manually determined PP reference data. Accordingly, a bias score of
b<1 represents an overprediction of snow by the model, whereas b>1
represents an overprediction of rain by the model. However, the bias should
be interpreted with caution because the manual reference data set might be
biased itself. Thus, the bias rather carries the information in which
direction the predicted PP deviates from the manual reference data.
Third, we determine the percentage of cases misclassified (PM). Misclassified
means that predicted high-probability cases (p>0.95) disagree with the
manual PP reference data. For PM, the number of these misclassified cases is
divided by the number of all RS cases. Ideally, PM is close to 0.
Fourth, the percentage of uncertain cases (PU) estimates how well the PPs are
separated by the predictor variables used. PU represents the number of cases
with 0.05<p< 0.95 divided by all RS cases. Accordingly, PU measures the
fraction of cases that the model is unable to predict at a high level of
certainty. The definitions of PM and PU follow the evaluation method in
.
The four performance scores are calculated for both 100 realizations of
50 % randomly chosen minutes of precipitation (black boxes and whiskers in
Fig. ) and for all minutes of RS sub-data (red stars). The
percentiles (5th, 25th, 50th, 75th, and 95th) illustrate how strongly the RS
data set scatters and whether differences among predictor variable
combinations are significant (p=0.95, n=100).
The PPs calculated with the logistic regression model reach an accuracy of
more than 88 % for combinations of atmospheric predictor variables that all
include the air temperature T (Fig. ). T carries the most
straightforward PP information in most cases. Combining T with up to two
other relevant predictor variables (connected by underscores) aids to assess
their value in determining the PP. Table displays the most
important fitted regression coefficients for different combinations of
predictor variables using the OceanRAIN sub-sample RS (2P1D) and the
sub-sample RSM (3P1D and 3P2D).
Combining T with the air temperature 2 h prior to observation
(T2h) does not increase the accuracy of T (both 88.5 %). Other
time intervals led to similarly small performance changes being in agreement
with . Accordingly, weather fronts
associated with T drops do not seem to have an imprint on T_T2h or
they are currently underrepresented in the OceanRAIN data set. The air
pressure P may have an impact on the PP at higher elevations due to lower
air density . This, however, cannot
explain the better accuracy of 89.2 % for T_P compared to T. Over
the ocean, the additional skill in the predictor P might be caused by
certain weather situations that favor either rain or snow, and are
sufficiently sampled in the OceanRAIN data set. The relative humidity rH and
the dew point Td (not shown) reach about the same accuracy of 89.4 %.
The addition of P and rH to T leads to a statistically significant
(p=0.95, n=100) but only slight increase in accuracy compared to T
alone.
Box-and-whisker plot displays interquartile spread (black box: 25th,
50th, and 75th percentile) and lower (whisker: 5th percentile) as well as
upper (95th percentile) extremes, calculated from 100 realizations of each
50 % randomly chosen minutes of precipitation from RS sub-data. Red stars
denote the values for 100 % of RS sub-data. Accuracy (%), bias score
(frac), percentage misclassified (PM: fraction of disagreeing cases with high
certainty of p>0.95 in %), and percentage unclassified (PU: fraction of
uncertain cases of 0.05<p<0.95 in %) serve as performance scores using the
calculated coefficients in Table against the manually
determined PP reference data. Labels indicate variable combinations, whereby
all combinations include T.
List of regression coefficients calculated with Eq. () by
minimizing the sum of squared errors with respect to the manual PP reference
data for two PPs using one PP distribution (2P1D; Sect. ), three PPs
using one PP distribution (3P1D; Sect. ), and three PPs using two PP
distributions (3P2D; Sect. ). For 3P2D, the asterisk denotes the rain distribution that was derived setting the mixed phase to snow. KS98 denotes the coefficients
recommended by derived over Finland.
With the 99th percentile of the particle diameter D99 and the calculated
theoretical rain rate RR (Eqs. and ), physical
properties of precipitation particles directly enter the PP distinction. This
direct physical relation explains the notably higher accuracy of at least
90 % by T_RR, T_D99, and other combinations containing RR
and D99. The similarly high performance of these three predictor
combinations is driven by the particle diameter that mainly influences RR.
Combinations of T, a humidity-related variable such as rH, and a
diameter-related variable such as D99 reach the highest accuracy of more
than 91 %. Combinations of four or five of the available atmospheric
predictor variables such as T_rH_RR_D99 brought no noticeable
further increase in accuracy (not shown). From the considered predictors, a
combination of three out of the available predictor variables suits best to
accurately distinguish between rain and snow.
The bias provides the ratio of rain cases predicted by the statistical model
and observed rain cases from the manual PP reference data. All predictor
variable combinations range between 0.89 and 0.94, which implies an
underprediction of rain and hence an overprediction of snow. Combinations
that contain RR and D99 reach the smallest overprediction of snow,
whereas T holds the strongest snow bias. The lower snow bias combined with
the higher accuracy of predictor variables carrying particle diameter
information highlights the need to include physically related variables in a
statistical model to predict the PP.
Besides being accurate and unbiased, a small PP transition region of low PP
certainty (low PU) combined with a low fraction of highly certain but
misclassified PP cases (low PM) characterize a useful predictor variable
combination. The PU decreases with increasing accuracy. Consequently,
predictor variable combinations including rH and either D99 or RR
reach the lowest PU of about 36 %. This low PU and thus fairly narrow PP
distribution causes a slight increase in PM for T_rH_RR and
T_rH_D99 (1.5 %) compared to T_D99, T_RR, and
T_RR_D99 (1.3 %). However, the positive effect of using RR or
D99 outweighs the slightly negative influence of rH on PM.
Consequently, the physical related predictor variables confirm their good
performance in predicting the PP.
Rain probability using regression coefficients from Table
for OceanRAIN RS sub-data (2P1D) with the predictor variables
T_rH (black), T_rH_D99 (blue) both fitted
against OceanRAIN, compared to KS98-recommended coefficients for
T_rH (red). Dashed lines (black, red) indicate a PP distribution
where rH is set to 80 % while for solid lines it is set to 99 %. For
T_rH_D99 (blue lines), D99 is set to either 1 or
5 mm in addition to rH.
The T_rH coefficients that were calculated for Finland in KS98 and
confirmed in over Switzerland reach an
accuracy of 88.6 %, which is slightly lower than those coefficients
optimized for OceanRAIN (89.4 %). A two-tailed t test confirms the
difference to be statistically significant (p=0.99, n=100). The
OceanRAIN-adapted coefficients exhibit a shallower rain/snow transition that
results in a 0.8 ∘C lower temperature at p(rain)=0.1 while both
distributions equal at p(rain)=0.9 (Fig. ). Compared
to OceanRAIN, the steeper rain/snow transition against T fitted in KS98
holds a much lower PU of 24 % but to the expense of a much higher PM of
4 % and a snow bias of 0.8. Consequently, the coefficients from KS98
better predict most uncertain cases with T_rH but miss more extreme
cases such as freezing rain. For the OceanRAIN data set, the PP prediction
using the RS-fitted coefficients better reflects the OceanRAIN PP
distribution compared to the KS98-fitted coefficients as indicated by the
accuracy.
For T_rH_D99, the rain/snow transition shifts with T depending
on D99. While D99=1 mm shifts the rain/snow transition to even
lower temperatures by about 0.5 ∘C, D99=5 mm shifts it towards
higher temperatures by about 2 ∘C, both compared to T_rH derived
from OceanRAIN RS sub-data. The shallower rain/snow transition of the PP
distribution fitted for OceanRAIN compared to that over Finland is likely
caused by more freezing rain cases sampled in OceanRAIN, which the
KS98-fitted coefficients for T_rH cannot predict.
One PP distribution to predict three PPs (3P1D)
In a second step, we include mixed-phase precipitation into the algorithm
because mixed-phase precipitation marks the transition from frozen to liquid
particles and thus carries the highest uncertainty. We calculate the
regression coefficients using the RSM sub-data including 165 632 min of
precipitation measured onboard RV Polarstern. The three-phase
distinction 3P1D fits p(rain) against three PPs from the same manually
determined PP reference data set as before. However, the calculated transition
phase between snow with p(rain)=0 and rain with p(rain)=1 is interpreted
as mixed phase, defined in the range of 0.3≤p(rain)≤0.7 after KS98.
The approximated coefficients for predictor variable combinations Vi
differ considerably from those calculated for the two-phase method 2P1D (see
Table in Sect. ).
We evaluate the calculated PP probability against PPs from the manual PP
reference data using RSM sub-data. Again, accuracy, bias, and PM serve as a
measure of quality, while PU is no longer suitable for evaluation because the
transition region of highest uncertainty between snow and rain represents
mixed-phase precipitation. Overall, this three-phase method 3P1D yields an
accuracy between 74 and 78 %, which corresponds to an absolute decrease
of about 14 % compared to 2P1D (Fig. ). To that large
decrease in accuracy two reasons mainly contribute: (1) the manual PP
reference data, acting as reference data, holds large uncertainties in the
mixed phase, as well. The ww code represents snapshots of 3-hourly
observations. Therefore, they hardly satisfy the need for minute-based
observations because the mixed-phase rain/snow fraction can vary
dramatically, both temporally and spatially. (2) KS98 assume the mixed-phase
precipitation to occur in the transition region between rain and snow, which
is true in most cases. However, several cases exist in which mixed-phase
precipitation occurs at distinctly high or low air temperature
(Fig. ) and thus 3P1D misclassifies these cases.
Relative to each other, the individual variable combinations perform similar
compared to 2P1D. T, T_T2h, and T_P have the lowest accuracy
of below 75 % (Fig. ) and a bias below 0.92. The
addition of rH significantly increases the accuracy by about 1 %,
whereas T_rH_T2h, T_rH, and T_Td_T2h (not shown)
do not differ much from each other. The predictor variable combinations that
include the diameter-related predictors RR and D99 lead to the highest
accuracy of 76 up to 78 %. The highest accuracy of 78 % reached by
T_rH_D99 represents a statistically significant performance
increase to the remaining variable combinations in 3P1D, which contrasts to
2P1D where T_rH_RR does not perform significantly better than
T_rH_D99.
For the bias, predictor combinations including RR and/or D99 reach the
least pronounced snow bias of about 0.93, whereas the remaining predictor
combinations feature significantly lower biases, mostly below 0.92. In that
respect, the bias of 3P1D resembles that of 2P1D (see Fig. in
Sect. ) both in terms of magnitude and in the individual performance
of the predictor variable combinations.
Performance of fit is shown for different combinations of atmospheric
variables as in Fig. for RSM sub-data. All variable combinations
again include T.
While the ranking of predictor variable combinations with respect to accuracy
and bias looks very similar compared to 2P1D, PM tends to form three
clusters. The first cluster comprises predictor variables without particle
diameter information, holding the lowest PM of 2.2 to 2.4 %. The second
cluster includes RR but not D99, holding the highest PM (3.4 %). In
the third cluster each predictor variable combination includes D99 but
performs better than the second cluster with PM of about 2.8 %.
T_rH_D99 in the third cluster offers the best compromise in
maximizing the accuracy while minimizing the fraction of misclassified cases.
In contrast to 2P1D, for 3P1D PM tends to scale with accuracy for many
predictor variable combinations. While T_rH_D99 exhibits an about
0.5 % larger PM than T, the PM of T_rH_RR and T_RR are
even 1.1 % larger. A high PM indicates a clear disagreement between
calculated PP and manually estimated PP. Note, however, that not in all of
these clearly disagreeing cases the manual PP reference data necessarily
contains the correct PP. Physically related predictor variables such as
D99 can assist to unveil cases falsely classified by the manual PP
estimation. For example, D99 is able to identify snow or mixed-phase
cases, falsely classified as rain in the manual reference data. Except for
the tropics, rain drops rarely exceed drop diameters of 5 mm
(; ). Larger
drops mainly break up or collide with neighboring drops. D99 excludes
coincidences of drops as well as artificial drops dripping off the instrument
housing by discarding the uppermost percentile of measured drop diameters per
minute. Therefore, particles classified as rain drops with D99>5 mm
very likely represent frozen particles, which means that they were falsely
classified as rain (Fig. ). Below 4 ∘C, 163
rain cases in RSM sub-data (about 0.25 %) are likely falsely classified.
This could explain about half of the 0.5 % PM difference of
T_rH_D99 to T in Fig. ).
The T_rH coefficients calculated in KS98 reach an accuracy of
78.6 %, but PM amounts to 7.2 % misclassified cases (not shown),
which is more than a factor of 2.5 higher than the PM of
T_rH_D99. The shift towards higher air temperatures and the
steeper rain/snow transition in the PP distribution using the coefficients
recommended in KS98 (see Fig. ) explain the large amount
of misclassified cases. However, as stated before, the coefficients in KS98
derived over Finland cannot represent the temperature distribution of PPs in
the OceanRAIN data set.
Two PP distributions to predict three PPs (3P2D)
The relatively low accuracy reached with the three-phase method after KS98
using one PP distribution (3P1D) motivates a novel investigation of how to
further improve the PP prediction for three PPs. Instead of applying one PP
distribution to determine rain, mixed-phase, and snow precipitation, we
suggest to approximate two separate PP distributions for rain and snow
(3P2D). These two individual PP distributions are derived analogous to the
method for one PP distribution by assigning the mixed-phase PP differently –
first set it to rain to calculate the snow PP distribution, then set it to
snow to calculate the rain PP distribution. Subtracting the sum of both
individually calculated PP distributions from 1 yields the PP distribution
for mixed phase. In contrast to 3P1D, the separately calculated coefficients
for rain and snow (Table ) lead to individual distributions
only connected via the mixed phase.
Analogous to 2P1D (Sect. ), the accuracy represents the percentage of
cases with p(PP)>0.5 that agree in their respective PP with the manual PP
reference data. The bias represents the ratio between the sum of predicted
rain cases and the sum of rain cases in the manual PP reference data. Please
note that the bias definition remains unchanged for 3P2D that includes
mixed phase compared to 2P1D. However, the additional PP distribution
slightly modifies the calculation of PM and PU, illustrated in
Fig. . PM represents the percentage of all certain cases
(p(PP)>0.95; hatched area in Fig. ) in which either
one of the PPs disagrees with the manual PP reference data. PU as the
percentage of uncertain cases (0.05<p(PP)<0.95; shaded area) represents
only those cases where all PPs are uncertain after definition. We introduce
this limitation because if p(PP)<0.05 holds for at least one PP then we
would not consider this PP uncertain anymore. Note that for mathematical
reasons we cannot display PMmix>0 and PU>0 in the same figure, which is
why we set PMmix>0.
Two-dimensional histogram of temperature and the 99th percentile of the
particle diameter for cases classified as rain by the manual PP estimation in
RSM.
This 3P2D method using two individual PP distributions reaches on average a
higher accuracy compared to 3P1D (Fig. ). Whereas T,
T_T2h, and T_P hold less than 78 % accuracy,
T_rH_D99 reaches the highest accuracy of 81.2 %. As for 3P1D,
the improvement is mainly caused by adding the predictor D99 that
performs significantly better than when adding the predictor RR. Also the
overprediction of snow by all predictor variable combinations with respect to
the manually determined PP reference data stays the same in 3P2D. The
physically related variables are least biased (about 0.93), which
consistently highlights the improvement of including them in the predictor
variable combination. However, for PM stronger differences among these
physically related predictor variables arise. While T_RR holds the
highest PM (about 2.3 %), T_rH_D99 reaches 1.9 % PM, which
is on the order of the predictor variable combinations without RR and
D99 (1.8 %). However, the physically related predictors reach again
lowest PUs of about 38 % while T holds a PU of 51 %. In combination
with the other scores we recommend T_rH_D99 followed by
T_RR_D99 and T_D99 to most accurately predict the PP using
two independent PP distributions.
Compared to 3P1D after KS98, the PM decreases for 3P2D. This decrease in PM
ranges between 0.5 and 1 % and thus highlights the improved performance
of using two PP distributions instead of one to predict the PP. The lower PM
and higher accuracy approve that the novel method applying two independent PP
distributions better represents the PP distribution in OceanRAIN RSM.
Graph illustrates the calculation of PU (framed) and PM (hatched)
including snow (dashed/purple), mixed phase (dotted/orange), and rain
(solid/cyan), analogous to Fig. 3 in . PU
divides the sum of cases with 0.05<p(PP)<0.95 for all PPs by the sum
of all RSM cases. PM divides the sum of cases with p(PP)>0.95 for
one of the PPs that disagrees with the manual PP estimation by the sum of all
RSM cases. We set PMmix>0 because otherwise we could not display it
in the same PP distribution (rH kept constant) with PU>0.
As Fig. but for RSM including mixed phase, using two
independent PP distributions (3P2D). The calculation of PM and PU differs
from Fig. as displayed and explained in
Fig. .
Lines show PP fraction for rain (solid, cyan), mixed phase (dotted,
orange), and snow (purple, dashed) from OceanRAIN RSM (165 632 min)
determined with the manual PP estimation against temperature. Gray bars
represent the temperature frequency of occurrence (in 103).
Air temperature versus predicted PP by the different methods: two
PPs (2P1D; solid blue), three one-PP distributions (3P1D; dashed red), and
three two-PP distributions (3P2D; dotted black). 3P2D consists of two curves
(left: snow distribution as 1-p(snow); right: rain distribution as
p(rain)) for the calculated coefficients of
T_rH_D99 (left panel; rH=85 %, D99=5 mm) and
T_rH (right panel; rH=85 %).
To understand the better performance of 3P2D compared to 3P1D after KS98, we
consider how the PP fraction is distributed with respect to T around the
freezing point (rain/snow transition) in the manual PP reference data
(Fig. ). While the rain occurrence shows a relatively
low skewness, the mixed-phase/snow distribution is slightly left-skewed. This
higher skewness with a secondary maximum in the mixed-phase distribution at
-3 ∘C (minimum in snow distribution) cannot be well represented by
one PP distribution. One PP distribution is limited to match all three PP
distributions at the same time that can only represent an average skewness.
In that respect, deriving two independent PP distributions driven by
mixed-phase precipitation better reflects the PP distribution of each PP
individually with respect to the manual PP reference data in OceanRAIN RSM.
The question arises whether the left-skewed distribution of snow and
mixed-phase precipitation in OceanRAIN sub-data RSM represents a feature of
the oceanic PP distribution or if it simply reflects a currently insufficient
length of the OceanRAIN time series. Though the latter seems more likely,
addressing this question comprehensively, however, remains beyond the scope
of this study due to the limited available OceanRAIN data sample. Future
studies could clarify this aspect by reanalyzing the constantly growing
OceanRAIN database.
Nevertheless, differences remain due to the chosen PP distinction method. By
discriminating three PPs, 3P1D and 3P2D enable a smoother rain/snow
transition compared to 2P1D due to included mixed-phase precipitation
(Fig. ). At lower temperatures, 2P1D approaches the snow
distribution of 3P2D, while at higher temperatures it approaches the rain
distribution of 3P2D. In other words, the steeper rain probability
distribution for 2P1D clarifies the slightly smaller unclassified range
(0.3<p(PP)<0.7) compared to 3P2D as seen in the percentage unclassified
(compare Fig. and Fig. ).
D99 as additional variable in T_rH_D99 tends to shift the
snow and rain distributions to higher temperatures and apart of each other,
which also resolves more extreme cases. This distribution shift with
temperature follows a physical reason: large snow particles better withstand
melting at high air temperatures than small snow particles. This physical
information lacks in T_rH, which notably decreases its accuracy (cf.
Fig. ).
Discussion
After finding suitable methods for both the rain/snow distinction
(Sect. ) as well as for the rain/snow/mixed-phase distinction
(Sect. ) we compare the results to those of similar studies. For
the rain/snow distinction over Switzerland using T_rH derived over
Finland by KS98, find a higher accuracy of
92.4 % compared to our calculated accuracy of 88.6 % when using the
same KS98 regression coefficients α=22, β=2.7, γ=0.2.
find an overprediction of snow cases (bias
0.82), very similar to the OceanRAIN RS snow overprediction (bias 0.8)
using the same coefficients derived by KS98. However, for fitting the
regression coefficients to our data set (Table ) we still
obtain a slightly lower accuracy of 89.4 % calculated for T_rH and
91 % for T_rH_D99 while the low-bias decreases to 0.92 and
0.93, respectively. These performance improvements indicate, first,
different conditions for PP transition over the ocean compared to Finland of
KS98 while, second, the OceanRAIN data set is in relatively close agreement
with the Swiss data.
With respect to two PPs, including the mixed phase decreases the accuracy to
below 78 % while PM almost doubles. To elaborate on reasons for that
accuracy decrease we consider a study of , who
applied the same translation of ww codes into PPs for ww codes between
50 and 86. Using 3P1D, they find an accuracy of 86 % compared to
Norwegian synoptic stations (6 months winter period) and 85 % compared
to independent climatological stations over Norway (1 month). They obtain
an overprediction of snow (bias of 0.92) and problems in predicting the PP
of supercooled rain during prevailing temperature inversions. In OceanRAIN we
find a similar overprediction of snow (bias T_rH: 0.91;
T_rH_D99: 0.93) with respect to the manual PP reference data in
OceanRAIN. This overprediction of snow occurs predominantly around
0 ∘C that is the temperature range sampled most frequently (cf.
Fig. ). Hence, OceanRAIN is likely to face the same
problems underpredicting rain when supercooled raindrops fall under
prevailing temperature inversions. Further work is required in order to
clarify whether we need additional ancillary data to reduce the bias or
whether the logistic regression model is unable to provide a less biased PP
prediction.
Assuming that mixed-phase precipitation causes most of the accuracy decrease
between 2P1D and 3P1D as well as 3P2D, we consider the individual probability
of detection (POD) for rain, snow, and mixed phase. For rain, the POD is
calculated by dividing the number of agreeing rain cases by the number of all
observed rain cases. For the POD of 3P1D using the KS98-fitted coefficients
for T_rH for rain, snow, and mixed phase we find 0.92, 0.78, and
0.21 (T_rH_D99: 0.92, 0.86, and 0.25). The respective PODs
from for the same settings reveal slightly
different PODs of 0.81, 0.97, and 0.25. Whereas they obtain a notably
higher POD for snow, the rain POD is lower compared to OceanRAIN.
Nevertheless, mixed-phase precipitation confirms to carry the largest
prediction uncertainty of all three PPs.
PP probability shown using the new 3P2D method with two individual PP
distributions (T_rH_D99) as frequency of occurrence (%) in
gray shades against air temperature according to PP reference data that
separates rain, snow, and mixed phase in OceanRAIN ALL for more than 4 years
of RV Polarstern data. Solid red lines represent the mean PP
fraction from observations in the Swiss Alps (1991–2010) from
; dashed blue lines show mean PP fraction
for oceanic ship data (DS464.0; 1977–2007) from
.
The variable combination T_rH_D99 distinguishes best rain, snow,
and mixed-phase precipitation in OceanRAIN data. In comparison with PP
fractions allocated into temperature bins from 30 years of Swiss Alps data
, in most cases the PP transition
in OceanRAIN occurs at lower temperatures (Fig. ).
However, the analysis by , among other
conditions, neglects all kinds of freezing rain (ww=56,57,66,67) that we
assign to rain. Without these “cold rain” cases, the rain/snow transition
shifts towards higher temperature that may in parts explain the temperature
difference in Fig. . Additionally, the PP probability
distribution in the OceanRAIN RV Polarstern data sample is biased by
the high number of temperatures around 0 ∘C that occur by a factor
of 3 to 4 more often than temperatures between -4 and 4 ∘C
(cf. Fig. ), and relative humidity close to 100 %.
These frequently sampled conditions put their mark on the average rain/snow
transition by reducing the rain/snow transition temperature compared to the
Swiss Alps where T and rH were sampled more homogeneously (Fig. 9 in
). Despite the high number of available
minutes with precipitation in OceanRAIN, the rather short time series on
climatological timescales and the spatial distribution of along-track data
limit the representativeness. However, a different land–ocean rain/snow
transition might be observable. found a
systematic land–ocean difference in the rain/snow transition between land
and ocean in 30 years of NCEP ADP Operational Global Surface Observations
(DS464.0; 1977–2007). Whereas over land, rain transitions into snow
relatively quickly (-2<T<4∘C), over ocean the transition zone is
wider (-3<T<6∘C). Although the rain/snow transition zone within
OceanRAIN appears wider compared to regression coefficients recommended by
as seen in Fig. , the rain/snow transition
in OceanRAIN compares better to the Swiss Alps data
than to the NCEP DS464.0 ocean data
that reveal a wider transition zone. In
specific, OceanRAIN relatively closely agrees with the NCEP DS464.0 ocean
data for T<0∘C, whereas larger differences of >1∘C occur in
the range of 2<T<5∘C. Two main reasons can explain the different
rain/snow transitions between OceanRAIN and NCEP DS464.0 ocean data by
. (1) ww codes used in the NCEP ocean data are
subject to larger uncertainty compared to OceanRAIN. In contrast to the RV
Polarstern onboard weather observatory by the German Meteorological
Service, many VOSs such as cargo ships in NCEP DS464.0 ocean data have
inadequately trained observers that might use certain ww codes
preferentially, ships possibly avoid bad weather, or measurement quality may
suffer from instrument biases . For these
reasons, the wider rain/snow transition zone likely reflects a higher
uncertainty of the NCEP DS464.0 ocean data compared to the OceanRAIN data
from RV Polarstern or the Swiss Alps data. (2) RV
Polarstern mainly sampled warm-season precipitation in the Atlantic
Arctic and Antarctic regions with the exception of the austral cold season in
2013. In addition to that, the heterogeneous regional and seasonal sampling
by RV Polarstern might have favored conditions under which
inversions prevail that allow rain at fairly low temperatures but inhibit
snow under relatively high temperatures. While the sampling imbalance of RV
Polarstern may indicate a restricted representativeness of PPs in
OceanRAIN, the T_rH_D99 predictor variable combination recommended
as the new automatic PP distinction method for OceanRAIN well represents the
observed PPs within OceanRAIN. The continuously growing time series of RV
Polarstern among other RVs in OceanRAIN allows to recalibrate or
refine the algorithm geographically for a longer time series with
comprehensive statistical sampling in the future.
Summary and concluding remarks
We developed a novel automatic algorithm to distinguish the PP within OceanRAIN in situ precipitation data to introduce a
statistical PP probability and to increase the data post-processing
efficiency. The analysis focused on identifying the most suitable combination
of available atmospheric predictor variables to determine the PP. For that
purpose, we applied a simple logistic regression model suggested by
that was shown to perform well over land. Previous studies
mainly rely on air temperature T, relative humidity rH, air pressure P,
and others to predict the PP. We investigated several of these atmospheric
predictor variable combinations to obtain a PP probability. In particular, we
tested the performance of the logistic regression model after
for OceanRAIN using two (excl. mixed phase) and three PPs (incl. mixed phase)
against the manually estimated observation-based PP in OceanRAIN. Besides
increasing the efficiency in predicting the PP with an automatic method, we
developed a novel three-phase method that uses two individual and independent
PP distributions to predict the PP more accurately.
The study led to the following main results.
In OceanRAIN, the combination of air temperature
T, relative humidity rH, and the 99th percentile of the particle diameter
D99 (called T_rH_D99) predicts best the PP for all
investigated methods.
Applying more than three of the chosen atmospheric predictor variables negligibly
increases the accuracy in predicting the PP.
The two-phase method (2P1D) using the predictor variable combination T_rH_D99
and regression coefficients fitted to OceanRAIN reaches an accuracy above
91 % with a slight overestimation of snow cases for the mid- and
high latitudes between -6 and 8 ∘C in the OceanRAIN data set with
respect to the manual PP reference data including shipboard present weather
observations.
A novel three-phase method using two individual PP distributions (3P2D) for rain and
snow performs better than a three-phase method that relies exclusively on one
PP distribution (3P1D after ). As a
reason, two individual PP distributions are capable of better representing
unequally distributed or skewed PP distributions of atmospheric predictor
variables as well as certain weather situations that might currently be over-
or undersampled. Accordingly, this performance difference might decrease once
the investigated 4-year OceanRAIN time series grows further while sampling
biases vanish.
The OceanRAIN data using 3P2D reveal a wider rain/snow transition zone compared to
data derived over Finland . The rain/snow transition in
OceanRAIN occurs at slightly lower temperatures compared to the data from
Finland as well as NCEP DS464.0 global ocean ship data
. The difference in the rain/snow transition zone
likely originates from heterogeneous spatial and seasonal sampling in
OceanRAIN that is likely to decrease with an increasing OceanRAIN time
series. In contrast, a higher quality of the derived ww codes in OceanRAIN
compared to the average VOS suggests a higher certainty of the derived PPs.
The Swiss Alps data shows a similar
rain/snow transition at slightly higher temperatures, likely caused by
neglected cases of freezing rain, among others. Due to these differences we
obtain the highest accuracy and lowest bias when applying regression
coefficients fitted to the OceanRAIN data set instead of using recommended
coefficients from the literature such as those from .
The new PP distinction algorithm 3P2D including D99 as essential physical
information identified several cases that were erroneously classified as rain
within the manual PP estimation. Large particle diameters indicate that the
PP should be classified as snow or at least mixed-phase precipitation instead
of rain.
Mixed-phase precipitation carries the largest uncertainty of the three
PPs and is most challenging to detect for the new algorithm with a
probability of detection of up to 0.25 using the predictor variable
combination T_rH_D99 and 3P2D.
Even though the newly developed automatic PP distinction algorithm
strongly depends on the currently still limited OceanRAIN data set, remarkable
improvements are made. First, a PP probability is provided on a minute basis
that limits the number of highly uncertain cases requiring visual inspection
of atmospheric variables. The PP probability further allows error
characterizing other precipitation data sets such as satellite data using
OceanRAIN precipitation rates to unveil systematic errors with respect to PP.
Second, the PPs of a few critical cases could be corrected that were falsely
classified by the manual method. Third, we give evidence that the particle
diameter of the falling precipitation particles contributes valuable
information to the PP separation and by that in a physical way significantly
improves the algorithm accuracy. Fourth, the new PP distinction algorithm
considerably speeds up the data processing within OceanRAIN, which is an
important step towards a fast-growing global surface precipitation data set
for validating and evaluating other oceanic precipitation data sets.
Data availability
The OceanRAIN data set is publicly available upon request
free of charge. A registration with a digital object identifier is planned.
Further information are available on
http://oceanrain.org.
Acknowledgements
We would like to thank the crew of RV Polarstern including the
German Meteorological Service and the Alfred Wegener Institute for their
support and assistance in supervising the instrument and caring for a smooth
operation and ship-board access. We further thank Nicole Albern for carefully
processing the data. Helpful comments by Bjorn Stevens are well appreciated
to finalize the manuscript. We also like to thank two anonymous reviewers for
their valuable comments. We acknowledge the financial support of the German
Research Foundation in research unit FOR1740.
The article processing charges for this open-access publication were covered by the Max Planck Society.
Edited by: M. Kulie
ReferencesAdler, R. F., Gu, G., and Huffman, G. J.: Estimating Climatological Bias Errors
for the Global Precipitation Climatology Project (GPCP), J. Appl. Meteorol.
Clim., 51, 84–99, 10.1175/JAMC-D-11-052.1, 2012.
Atlas, D. and Ulbrich, C.: The physical basis for attenuation-rainfall
relationships and the measurement of rainfall parameters by combined
attenuation and radar methods, J. Rech. Atmos., 8, 275–298, 1974.Battaglia, A., Rustemeier, E., Tokay, A., Blahak, U., and Simmer, C.:
PARSIVEL Snow Observations: A Critical Assessment, J. Atmos.
Ocean. Tech., 27, 333–344, 10.1175/2009JTECHA1332.1,
2010.Bentley, W.: Studies of Raindrops and Raindrop Phenomena, Mon. Weather Rev.,
32, 450–456, 10.1175/1520-0493(1904)32<450:SORARP>2.0.CO;2, 1904.Bumke, K. and Seltmann, J.: Analysis of Measured Drop Size Spectra over Land
and Sea, ISRN Meteorology, 2012, 1–10, 10.5402/2012/296575, 2012.Clemens, M.: Machbarkeitsstudie zur räumlichen Niederschlagsanalyse aus
Schiffsmessungen über der Ostsee, Berichte aus dem Institut für
Meereskunde an der Christian-Albrechts-Universität Kiel, 1–50,
10.3289/ifm_ber_327,
2002.Dai, A.: Temperature and pressure dependence of the rain-snow phase transition
over land and ocean, Geophys. Res. Lett., 35, L12802,
10.1029/2008GL033295, 2008.Friedrich, K., Higgins, S., Masters, F. J., and Lopez, C. R.: Articulating and
Stationary PARSIVEL Disdrometer Measurements in Conditions with Strong
Winds and Heavy Rainfall, J. Atmos. Ocean. Tech., 30, 2063–2080,
10.1175/JTECH-D-12-00254.1, 2013.Froidurot, S., Zin, I., Hingray, B., and Gautheron, A.: Sensitivity of
Precipitation Phase over the Swiss Alps to Different Meteorological
Variables, J. Hydrometeorol., 15, 685–696, 10.1175/JHM-D-13-073.1,
2014.Fuchs, T., Rapp, J., Rubel, F., and Rudolf, B.: Correction of synoptic
precipitation observations due to systematic measuring errors with special
regard to precipitation phases, Phys. Chem. Earth Pt. B, 26, 689–693,
10.1016/S1464-1909(01)00070-3, 2001.Gjertsen, U. and Ødegaard, V.: The water phase of precipitation – a comparison
between observed, estimated and predicted values, Atmos. Res., 77, 218–231,
10.1016/j.atmosres.2004.10.030, 2005.Goodison, B. E.: Accuracy of Canadian Snow Gage Measurements, J. Appl.
Meteorol., 17, 1542–1548,
10.1175/1520-0450(1978)017<1542:AOCSGM>2.0.CO;2, 1978.
Großklaus, M.: Niederschlagsmessung auf dem Ozean von fahrenden
Schiffen, PhD thesis, Institut für Meereskunde,
Christian-Albrechts-Universität Kiel, Kiel, Germany, 1996.Hogan, A. W.: Objective Estimates of Airborne Snow Properties, J.
Atmos. Ocean. Tech., 11, 432–444,
10.1175/1520-0426(1994)011<0432:OEOASP>2.0.CO;2, 1994.Ishizaka, M., Motoyoshi, H., Nakai, S., Shiina, T., Kumakura, T., and Muramoto,
K.-i.: A New Method for Identifying the Main Type of Solid Hydrometeors
Contributing to Snowfall from Measured Size-Fall Speed Relationship, J.
Meteorol. Soc. Jpn., 91, 747–762, 10.2151/jmsj.2013-602, 2013.Kent, E. C., Ball, G., Berry, D. I., Fletcher, J., Hall, A., North, S., and
Woodruff, S.: The Voluntary Observing Ship (VOS) Scheme, European Space Agency, 518–528 10.5270/OceanObs09.cwp.48, 2010.Kidd, C. and Levizzani, V.: Status of satellite precipitation retrievals,
Hydrol. Earth Syst. Sci., 15, 1109–1116, 10.5194/hess-15-1109-2011, 2011.Klepp, C.: The oceanic shipboard precipitation measurement network for surface
validation – OceanRAIN, Atmos. Res., 163, 74–90,
10.1016/j.atmosres.2014.12.014, 2015.Klepp, C., Bumke, K., Bakan, S., and Bauer, P.: Ground validation of oceanic
snowfall detection in satellite climatologies during LOFZY, Tellus A, 62,
469–480, 10.1111/j.1600-0870.2010.00459.x, 2010.
Koistinen, J. and Saltikoff, E.: Eperience of customer products of accumulated
snow, sleet and rain, COST75 Advanced Weather Radar Systems, 1998.Lanza, L. G. and Vuerich, E.: The WMO Field Intercomparison of Rain
Intensity Gauges, Atmos. Res., 94, 534–543,
10.1016/j.atmosres.2009.06.012, 2009.Lempio, G. E., Bumke, K., and Macke, A.: Measurement of solid precipitation
with an optical disdrometer, Adv. Geosci., 10, 91–97,
10.5194/adgeo-10-91-2007, 2007.Löffler-Mang, M. and Joss, J.: An Optical Disdrometer for Measuring Size
and Velocity of Hydrometeors, J. Atmos. Ocean. Tech., 17, 130–139,
10.1175/1520-0426(2000)017<0130:AODFMS>2.0.CO;2, 2000.Merenti-Välimäki, H.-L., Lönnqvist, J., and Laininen, P.: Present
weather: comparing human observations and one type of automated sensor,
Meteorol. Appl., 8, 491–496, 10.1017/S1350482701004108, 2001.Michelson, D. B.: Systematic correction of precipitation gauge observations
using analyzed meteorological variables, J. Hydrol., 290, 161–177,
10.1016/j.jhydrol.2003.10.005, 2004.Ocean Rainfall And Ice phase precipitation measurement Network (OceanRAIN): precipitation
particle spectra, precipitation rates and ancillary data from RV Polarstern
over the Atlantic Ocean including polar regions, University Hamburg,
CliSAP/CEN, Hamburg, Germany, available at:
http://oceanrain.org, last access: 7 April 2016.Petty, G. W.: Frequencies and Characteristics of Global Oceanic Precipitation
from Shipboard Present-Weather Reports, B. Am. Meteorol. Soc., 76,
1593–1616, 10.1175/1520-0477(1995)076<1593:FACOGO>2.0.CO;2, 1995.
Schmid, W. and Mathis, A.: Validation of methods to detect winter precipitation
and retrieve precipitation type, 12th SIRWEC Conference, 16–18 June 2004, Bingen, Germany, 2004.Schneider, U., Becker, A., Finger, P., Meyer-Christoffer, A., Ziese, M., and
Rudolf, B.: GPCC's new land surface precipitation climatology based on
quality-controlled in situ data and its role in quantifying the global water
cycle, Theor. Appl. Climatol., 115, 15–40, 10.1007/s00704-013-0860-x, 2014.
Taylor, P.: Intercomparison and Validation of Ocean-Atmosphere Energy Flux
Fields – Final report of the Joint WCRP/SCOR Working Group on Air-Sea
Fluxes, 325 pp., WCRP-112, WMO/TD-1036, 2000.Villermaux, E. and Bossa, B.: Single-drop fragmentation determines size
distribution of raindrops, Nat. Phys., 5, 697–702, 10.1038/nphys1340, 2009.Weller, R. A., Bradley, E. F., Edson, J. B., Fairall, C. W., Brooks, I.,
Yelland, M. J., and Pascal, R. W.: Sensors for physical fluxes at the sea
surface: energy, heat, water, salt, Ocean Sci., 4, 247–263,
10.5194/os-4-247-2008, 2008.WMO: Code Tables and Flag Tables Associated with BUFR/CREX Table B, available at: http://www.wmo.int/pages/prog/www/WMOCodes/WMO306_vI2/LatestVERSION/WMO306_vI2_BUFRCREX_CodeFlag_en.pdf (last access: 7 April 2016), 2015.Yuter, S. E. and Parker, W. S.: Rainfall Measurement on Ship Revisited: The
1997 PACS TEPPS Cruise, J. Appl. Meteorol., 40, 1003–1018,
10.1175/1520-0450(2001)040<1003:RMOSRT>2.0.CO;2, 2001.Yuter, S. E., Kingsmill, D. E., Nance, L. B., and Löffler-Mang, M.:
Observations of Precipitation Size and Fall Speed Characteristics within
Coexisting Rain and Wet Snow, J. Appl. Meteorol. Clim., 45, 1450–1464,
10.1175/JAM2406.1, 2006.