Page 1
ORIGINAL PAPER
Validating a rapid-update satellite precipitation analysis acrosstelescoping space and time scales
Francis Joseph Turk Æ Byung-Ju Sohn ÆHyun-Jong Oh Æ Elizabeth E. Ebert ÆVincenzo Levizzani Æ Eric A. Smith
Received: 13 December 2006 / Accepted: 6 August 2009 / Published online: 22 August 2009
� Springer-Verlag 2009
Abstract In order to properly utilize remotely sensed
precipitation estimates in hydrometeorological applica-
tions, knowledge of the accuracy of the estimates are
needed. However, relatively few ground validation net-
works operate with the necessary spatial density and time-
resolution required for validation of high-resolution pre-
cipitation products (HRPP) generated at fine space and
time scales (e.g., hourly accumulations produced on a 0.25�spatial scale). In this article, we examine over-land vali-
dation statistics for an operationally designed, meteoro-
logical satellite-based global rainfall analysis that blends
intermittent passive microwave-derived rainfall estimates
aboard a variety of low Earth-orbiting satellite platforms
with sub-hourly time sampling capabilities of visible and
infrared imagers aboard operational geostationary plat-
forms. The validation dataset is comprised of raingauge
data collected from the dense, nearly homogeneous, 1-min
reporting Automated Weather Station (network of the
Korean Meteorological Administration during the June to
August 2000 summer monsoon season. The space-time
RMS error, mean bias, and correlation matrices were
computed using various time windows for the gauge
averaging, centered about the satellite observation time.
For ±10 min time window, a correlation of 0.6 was
achieved at 0.1� spatial scale by averaging more than 3
days; coarsening the spatial scale to 1.8� produced the
same correlation by averaging over 1 h. Finer than
approximately 24-h and 1� time and space scales, respec-
tively, a rapid decay of the error statistics was obtained by
trading-off either spatial or time resolution. Beyond a daily
time scale, the blended estimates were nearly unbiased and
with an RMS error of no worse than 1 mm day-1.
1 Introduction
The past two decades have witnessed the rapid evolution of
the low Earth-orbiting (LEO) passive microwave (PMW)
imaging sensor from a research setting into routine oper-
ational and climate applications. This is substantiated by
the mid-2005 decision to extend the lifetime of the joint
United States/Japan Tropical Rainfall Measuring Mission
F. J. Turk
Marine Meteorology Division, Naval Research Laboratory,
7 Grace Hopper Avenue, Monterey, CA 93940, USA
B.-J. Sohn � H.-J. Oh
School of Earth and Environmental Sciences,
Seoul National University, Seoul 151-747, Korea
e-mail: [email protected]
H.-J. Oh
e-mail: [email protected]
E. E. Ebert
Centre for Australian Weather and Climate Research,
Bureau of Meteorology Research Centre, GPO Box 1289K,
Melbourne, VIC 3001, Australia
e-mail: [email protected]
V. Levizzani
National Council of Research, Institute of Atmospheric Sciences
and Climate, ISAC-CNR, via Gobetti 101, 40129 Bologna, Italy
e-mail: [email protected]
E. A. Smith
National Aeronautics and Space Administration, Goddard Space
Flight Center, Code 613.6, Greenbelt, MD 20771, USA
e-mail: [email protected]
Present Address:F. J. Turk (&)
Jet Propulsion Laboratory, California Institute of Technology,
Pasadena, CA 91109, USA
e-mail: [email protected]
123
Meteorol Atmos Phys (2009) 105:99–108
DOI 10.1007/s00703-009-0037-4
Page 2
(TRMM), with its companion Microwave Imager (TMI)
and Precipitation Radar (PR) (Kummerow et al. 2000), and
preparations for the upcoming Global Precipitation Mission
(GPM 2008). Together with the Advanced Microwave
Scanning Radiometer (AMSR-E) onboard the Earth
Observing System (EOS) Aqua satellite, the Defense
Meteorological Satellite Program (DMSP) Special Sensor
Microwave Imager (SSMI) and Imager-Sounder (SSMIS),
the Advanced Microwave Sounding Unit (AMSU-B) and
Microwave Humidity Sounder (MHS) onboard the
National Oceanic and Atmospheric Administration
(NOAA) and EUMETSAT MetOp satellites, and the
WindSat sensor onboard Coriolis, these PMW sensors all
possess imaging and/or sounding channels sensitive to
precipitation-sized rain and frozen hydrometeors over land
and water surfaces. The verification of satellite-derived
precipitation from these sensors is problematic owing to
many factors, such as the evolving time and spatial scales
of the precipitation process, the intermittent and unequally
spaced satellite overpass times, the instantaneous nature of
a moving-platform satellite observation, and the method by
which these instantaneous satellite precipitation data are
integrated to into accumulation products. Moreover, the
verification data itself should be properly aligned with the
spatial and temporal scales of the satellite-based precipi-
tation products.
Early SSMI-era validation efforts such as the Precipi-
tation and Algorithm Intercomparison Programs (PIP and
AIP, respectively) evaluated PMW precipitation algorithms
on a global scale or by meteorological event (AIP-1: Arkin
and Xie 1994; AIP-3: Ebert et al. 1996; PIP-1: Barrett et al.
1994; PIP-2: Smith et al. 1998; PIP-3: Adler et al. 2001).
The ongoing Program for the Evaluation of High Resolu-
tion Precipitation Products (PEHRPP; Turk et al. 2008), an
initiative of the International Precipitation Working Group
(IPWG; Turk and Bauer 2006), is geared toward daily time
scale validation of many multi-satellite techniques (as well
as several numerical weather prediction models) using
national radar and raingauge analyses. High Resolution
Precipitation Products (HRPP) combine a multitude of
space borne remotely estimated and ground-based datasets
in order to generate a precipitation product that is of a finer
spatial and/or temporal resolution than any of the individ-
ual input datasets. The idea behind all HRPPs is to augment
the infrequent, but physically based precipitation estimates
from the PMW sensors with indirect, but fast-update geo-
stationary Earth orbiting (GEO) environmental satellite
visible/infrared imagery (VIS/IR) data, augmented in some
cases by surface radar and raingauge information and
analyses from numerical weather prediction (NWP) mod-
els. Examples of commonly used HRPPs are the Tropical
Rainfall Measuring Mission (TRMM) Multisatellite Pre-
cipitation Analysis (TMPA; Huffman et al. 2007), the
Precipitation Estimation from Remotely Sensed Informa-
tion using Artificial Neural Networks (PERSIANN) data-
sets (Sorooshian et al. 2000), the Climate Prediction Center
morphing technique (CMORPH; Joyce et al. 2004), the
Global Satellite Mapping of Precipitation (GSMaP; Ushio
et al. 2004), and the NRL-Blend (Turk and Miller 2005),
among others. Typically, these HRPPs combine multiple
satellite datasets (and some add in additional raingauge and
other non-satellite data) and produce estimates of 3-h
accumulated precipitation between ±60� latitude, updated
every 3 h, at a gridded spatial resolution of 0.25�, although
some HRPPs have even finer space and time scales. Some
of these HRPPs are designed to be strictly operated in near
realtime (e.g, NRL-Blend), while others create near real-
time as well as a higher quality, post-processed non-real-
time datasets.
However, to properly evaluate the performance of a
HRPP dataset at sub-daily time and sub-one-degree spatial
scales requires a ground-based validation system with a
dense, homogeneous spatial coverage and a time sampling
rate fast enough (and extended over a long enough period
of time) to coordinate matched comparisons with the
HRPP. The worst-case revisit time from the current PMW-
based satellite constellation currently hovers near 6 h in the
tropical latitudes, owing to the sun-synchronous orbits of
all platforms except TRMM (Negri et al. 2002). Sapiano
and Arkin (2009) validated several 3 h HRPPs over the
continental United States using radar and raingauge data.
Using a multiyear radar rainfall dataset, Steiner et al.
(2003) simulated the uncertainty of satellite-estimated
rainrate that would arise from finite sampling time inter-
vals, and repeated this analysis when the radar datasets
were averaged across various spatial scales and time
intervals. A similar study was undertaken by Gebremichael
and Krajewski (2004), who also concluded that the sam-
pling errors have seasonal and regional variations and
differ for land and ocean backgrounds. Even so, these
results have proven useful to establish approximate time
sampling uncertainty bounds for instantaneous overpass
satellite precipitation estimates. However, several factors
complicate the assessment of the time sampling error
uncertainty in a blended satellite HRPP. The final products
represent results that combine rapid, constant time sam-
pling from the geostationary datasets with the infrequent,
unequally spaced time sampling from the PMW datasets. In
this situation, the uncertainty associated with PMW time
sampling is highly variable; if the PMW revisit becomes
too long, HRPP techniques increasingly rely upon
increasingly older PMW data, and depend upon some
means to track the in-between rain evolution with the
geostationary datasets (Joyce et al. 2004; Ushio et al.
2004). At the same time, there are inter-sensor differ-
ences amongst the precipitation estimates from each LEO
100 F. J. Turk et al.
123
Page 3
satellite platform. In the GEO datasets, inter-sensor dif-
ferences include limb corrections for satellite zenith and
coverage overlap between adjacent geostationary imagers
(Joyce and Ferraro 2006), viewing parallax (Vicente et al.
2002), and the varying spatial resolution of across-track
scanners (Joyce et al. 2006), among others. In this manu-
script, we will not focus on the error associated with the
inter-sensor or time sampling underlying the HRPP; rather,
we will focus our attention on the skill of the final HRPP
product through direct comparison with a dense, fine time
sampling raingauge network.
While the gauge-comparison methods that will be dis-
cussed in this manuscript are applicable to many satellite
HRPPs, the blended-satellite HRPP used hereafter is the
blended-satellite technique developed at the Naval
Research Laboratory (NRL-Blend) (Turk and Miller 2005).
The core of the technique has been upgraded and used in
several research applications (Grose et al. 2002; Turk et al.
2009; Tian et al. 2007; Xian et al. 2009), and is discussed
in Sect. 2. Section 3 focuses upon the main effort of this
work, the validation and performance analysis of the NRL
HRPP at various space and time scale combinations, using
raingauge data from the densely spaced, 1-min updating
Korean Meteorological Administration (KMA) Automated
Weather Station (AWS) raingauge network (Oh et al.
2002). For short time scales, the rain may fall between but
not into individual gauges, or the rainfall pattern may
evolve and move between the gauge locations. Therefore,
the AWS and HRPP data are first collocated by time and
location, and at their highest resolution, prior to any sub-
sequent spatial and time scale averaging. The averaging is
done across spatial scales ranging from 0.1� to 3� box
sides, and from time scales ranging from 1 h to 30 days, in
order to construct the two-dimensional matrix of correla-
tion, mean bias, and root-mean-square error (RMSE). This
produces maps of the error metrics across varying tele-
scoping space and time scales. We note that the more
complex issue of physical validation is not addressed in
this paper, although such an approach would be the next
step in understanding the physical means (e.g., micro-
physical assumptions in the PMW rainfall algorithms)
behind the performance of this and other HRPPs.
2 Blended satellite technique description
In this section, we outline the design and implementation
of the NRL blended satellite precipitation technique
(NRL-Blend), which is based upon a real time, underlying
collection of time and space-matching pixels from all
operational geostationary (GEO) visible/infrared (VIS/IR)
imagers and PMW imagers onboard LEO satellites.
It operates in an autonomous, operational mode with a
steadily arriving stream of near real-time data from the
operational GEO and LEO satellites. As of mid-2009, the
current operational GEO satellites are GOES-11, GOES-
12, Meteosat-7, Meteosat-9 (MSG-2), and GMS-6
(MTSAT-1R). The current (mid-2009) LEO constellation
utilizes all 13 satellites for a total of 13 PMW sensors and 1
active radar system, including 3 Advanced Microwave
Sounding Units (AMSU-B) onboard NOAA-15/16/17,
three Microwave Humidity Sounders (MHS) onboard
NOAA-18/19 and MetOp-A, three Special Sensor Micro-
wave Imagers (SSMI) DMSP) F-13/14/15, two Special
Sensor Microwave Imager Sounder (SSMIS) onboard
DMSP F-16/17, one Advanced Microwave Scanning
Radiometer (ASMR-E) onboard the Earth Observing Sys-
tem (EOS) Aqua, the WindSat polarimetric radiometer
onboard Coriolis, and the Tropical Rainfall Measuring
Mission (TRMM) Microwave Imager (TMI) and its com-
panion Precipitation Radar (PR). All of these satellites orbit
sun-synchronously with the exception of TRMM, whose
local observing time repeats approximately every 28 days
at the equator (Negri et al. 2002). NOAA-15/17, Metop-A,
DMSP, and Coriolis are in orbit patterns with equator
crossing times in early-mid morning (AM) and evening
(within a few hours of the solar terminator), whereas
NOAA-18/19 and Aqua are in orbits which crossover in
early afternoon (PM) and early morning (NOAA-16 is an
operational backup which has drifted from its initial
afternoon 14:00 local equator crossing time to near 17:00
as of mid-2009). AMSU-B and MHS are cross-track
scanning sounders which can be used also for precipitation
estimation, PR is an across-track scanning Ku-band radar,
whereas SSMI, SSMIS, WindSat, and AMSR-E all scan
conically.
The operation of the NRL-Blend is essentially described
by three procedures (Turk and Miller 2005) (a) background
collocation of time/space intersecting pixels from all
operational GEO VIS/IR imagers and LEO PMW imagers,
which are used to dynamically build lookup tables relating
IR temperature to PMW-retrieved rainrates (b) use of these
collocated data for adjustment of VIS/IR data into instan-
taneous rainrates via bicubic interpolation of lookup table
values, and (c) updates of 3, 6, 12, and 24-h interval
accumulations updates at each 3-h synoptic time (00, 03,
…21 UTC). In the accumulations, each gridded GEO-
based instantaneous rainrate value is weighted according to
its time proximity to the nearest PMW overpass. The PMW
estimates are always fully weighted and the GEO estimate
receives a smaller weight the closer it occurs to a PMW
overpass. In step (b), there is an option to use numerical
weather prediction (NWP) forecast model 850 hPa wind
vectors from the Navy Operational Global Atmospheric
Prediction System (NOGAPS), combined with a high-res-
olution topographic database, to apply a correction factor is
Validating a rapid-update satellite precipitation analysis 101
123
Page 4
applied in regions of likely orographic effects on both the
upslope and downslope sides (for the time period used in
this study, this option was not yet available).
The NRL-Blend was operated intermittently beginning
as early as 1998, although longterm data collection of these
global precipitation accumulation products began in Janu-
ary 2004. Although the computations are done on a 0.1�latitude–longitude grid between ±60� latitude, the archived
products are averaged to a global 0.25� grid (480
lines 9 1,440 samples).
3 Comparisons with the Automated Weather Station
network
The KMA maintains an operational, densely spaced AWS
network over the southern Korean Peninsula, consisting of
nearly 500 tipping-buckets, uniformly spaced, 1-min
updating raingauges (approximately 40 gauges per 1� box).
Figure 1 depicts the AWS grid (not all stations are shown).
AWS data were collected during June to August 2000
along with the individual hourly, instantaneous rainfall
datasets produced by the blended satellite technique (the
GMS-5 satellite was the operational geostationary satellite
during this time and its routine refresh rate was hourly
beginning at 30 min after each hour, and the Korean
Peninsula was imaged about 8 min after the frame start
time). Figure 2a depicts the mean monthly rainrate over the
Southern Korean peninsula during each of these 3 months,
and Fig. 2b depicts the associated daily average rainrate
time series (i.e., each point represents the mean rainrate
during one 24-h period). Note that the rainfall during this
3-month period is dominated by four heavy rain events, one
each in June and July and two in August. During this time,
the AMSU-B, MHS, and Aqua AMSR-E rain products
were not yet incorporated into the NRL-Blend technique,
giving a LEO revisit over Korea of about 4 h on average
and 10 h worst-case. Considering that an individual
satellite observation represents a 0.1� (approximately
10 km 9 10 km area) area average whereas the gauge
measurement represents a small area less than 1 m2, South
Korea is divided into smaller boxes, ranging from 0.1� to
3� on a side, where relatively homogeneous gauge distri-
butions are found (Fig. 1 depicts the 1� box sizes). Because
of the inhomogeneity of the rain within the spatial aver-
aging box and very small areas represented by individual
gauges, a direct comparison of instantaneous (i.e., sensor
scan level) satellite-based retrievals and gauges is inher-
ently limited. For intermittent and sporadic rain events, the
rain may fall between but not into individual gauges, or the
rainfall pattern may evolve and move between the gauge
locations. Using the AWS network and various IR-based
rainfall techniques over a 1-month period, Oh et al., (2002)
investigated the impact of the spatial rain inhomogeneity
by analyzing the number of rain-detected gauges in indi-
vidual 1� boxes. They found that the satellite algorithm
validation was more likely to fail for sporadic and weak
rain events when the number of rain-detecting gauges per
1� box was less than 15. However, in some cases, isolated
convective rain events may be characterized by a small
number of rain-detecting gauges, so a simple minimum-
gauge criterion is not necessarily sufficient in all cases. In
the discussion to follow, it must be therefore remembered
that unambiguous interpretation is not always possible with
gauge-satellite comparisons, especially for sporadic rain
events.
For example, Fig. 3 depicts an example grid box (e.g.,
3� on each side) and the actual accumulated rain pattern
resulting from the passage of a precipitating cloud system.
There is a variable cloud-ground fall time (represented with
a time Dt) for all rain hydrometeors, ultimately forming the
pattern of the accumulated rainfall on the surface. The
surface rainfall is not necessarily associated with the cloud
vertically above it, since horizontal air motions reposition
the rain once it falls out of the cloud (represented by an
offset of Dd), and drops break up and/or coalesce. For
convective rain, the situation is both dynamically and
microphysically complex at a very fine scale, and evolving
with time. For comparison, Fig. 4 depicts a satellite over-
pass and its framework for retrieving and mapping the
rainfall onto a finer-scale inner grid (e.g., a 0.25� grid). The
Fig. 1 Depiction of the Automated Weather Station operational rain
gauge network operated by the Korean Meteorological Administra-
tion. The density is nearly homogeneous across the southern Korea
peninsula with approximately 40 gauges per 1� box, each reporting at
a 1-min time update
102 F. J. Turk et al.
123
Page 5
Time series of daily mean rainrate over South Korea
Time (day)
Rai
nrat
e (m
m/h
r)
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
Jun Jul Aug
(a)
(b)
Fig. 2 a Maps of the monthly
mean rainrate over the South
Korean peninsula for June, July,
and August 2000. b Time series
of daily rainrate over the South
Korean peninsula during June
to August 2000
Validating a rapid-update satellite precipitation analysis 103
123
Page 6
precipitation will be mapped onto the surface according to
the geolocation of the rain-flagged satellite pixels, here
shown for a satellite viewing along nadir. In reality, coni-
cally scanning PMW radiometers typically view along 53�from nadir, and geostationary imagers view along nadir at
the equator, to 70� from nadir at 60� latitude. Bauer et al.
(2002) and Vicente et al. (2002) have developed methods
to correct for the viewing-angle-induced parallax from
PMW and geostationary imagers, respectively; however,
the rainfall is still displaced due to horizontal air motions,
and advection of the cloud system as it moves across the
network. Ultimately, the surface rainfall estimates derived
from the satellite essentially represent conditions at or near
the physical cloud top height, and the pattern of the surface
rainfall more or less resembles the pattern of the cloud, as
shown in Fig. 4. Space and time averaging of the raingauge
analysis and/or the satellite-derived rainfall estimates are
necessary to make meaningful comparisons. One way to do
this is to choose a time interval, and then compute separate
averages of all raingauge and satellite-estimated rainfall
within the box, illustrated in Fig. 3. This may be a good
assumption for sufficiently long enough time intervals,
when many rain systems move across the gauges. How-
ever, for short time intervals, substantial rain may fall in
between gauges, whereas the satellite overpass and sensor
scan always captures the rain in the entire box.
Therefore, to account for the fallout time of the hy-
drometeors, the gauge data were first averaged over time
windows varying between ±1 and ±30 min, centered
about the GMS-5 observation time over Korea. Then, at
each spatial resolution, these data were time-integrated
over various time intervals ranging from 1 h to 30 days.
That is, the original data were fixed at one spatial scale and
were then integrated over the various time scales, then
repeated for the next spatial scale, and so on. This allows
two-dimensional matrix of performance scores to be con-
structed. Most importantly, the 1-min time resolution of the
dense KMA network allows the gridded instantaneous
satellite pixels to be closely aligned with the corresponding
raingauge locations by observation time and location, prior
to any space and time averaging.
Figure 5 shows the correlation, mean bias, and RMSE
for AWS time windows ranging from -30 to ?30 min. For
example, if the GMS-5 satellite overpass occurred at 12:00
UTC, then a value of -10 means than all raingauge data
between 11:50 and 12:00 UTC were used in the time
averaging for all space/time combinations. The different
time windows produce different results, owing to the var-
iable fallout times of the hydrometeors from within the
cloud, and the increased number of gauges in the average
as the window is widened. There is a sharp improvement
when the time window is widened from 0 to ±5 min, and
to ±10 min, which is a typical hydrometeor fallout time in
tropical clouds (Soman et al. 1995). The skill scores appear
largely independent of whether the gauge averaging is
performed before or after the GMS-5 observation time.
This suggest that any improvements in skill scores via
precise time-coordination between the satellite overpass
and gauge report times is still dominated by the total
number of gauges that go into the AWS time window
averaging. With a wider time window, the chances of one
or more gauges reporting a non-zero value are higher.
Figure 6a, b, c depict the results from the analysis at
this and other space and time scales in a two-dimensional
format, where the spatial average and averaging period
determine the abscissa and ordinate, respectively. The
Fig. 3 Illustration depicting the advection of a precipitating cloud
across a raingauge network inside of a latitude–longitude gridbox.
The red dots identify raingauge locations. Dd represents the offset
between the rainfall position within the cloud and where it falls on the
surface, and Dt represents the falltime between the cloud and surface.
The blue dotted lines represent the fall trajectories of the in-cloud rain
regions, which generate the rainfall pattern illustrated in blue
Fig. 4 Illustration depicting the resultant satellite-derived precipita-
tion pattern from the cloud structure shown in Fig. 3. The satellite-
derived rainfall is map registered onto an inner map grid inside of the
gridbox (not to scale). Since the satellite has no knowledge of the
underlying rain structure, Dd and Dt are assumed to be 0, and the rain
is assumed to be on the same position on the ground that it is within
the cloud
104 F. J. Turk et al.
123
Page 7
correlation, mean bias and RMSE are each contoured for
AWS time windows (gauge averaging time centered about
the GMS observation time) of ±1, ±5, and ±10 min. The
different time windows produce different results, owing to
the variable fallout times of the hydrometeors from within
the cloud, and the increased number of gauges in the
average as the window is widened. As expected, all three
parameters improve as either the averaging period is
increased or the grid size is coarsened. The contours do not
flow smoothly at the longest time intervals due to the
resultant small number of data points available from
the finite 3-month period. Likewise, the small size of the
Korean Peninsula produces a small number of data points
when the data are averaged over the coarsest space scales.
The blended technique is biased slightly negative (-0.1, or
10% low) once the time interval exceeds 3 days, falling to
about -0.35 (35% low) at 3-h/0.25� scales. One possible
explanation for the bias behavior is that as the time scale is
shortened, extreme heavy precipitation events are less
likely to be captured by a LEO overpass. Since the nature
of the NRL-Blend technique is to retain some of the most
recent rainfall history, there is a gradually increasing
(rather than sudden) negative bias. The RMSE is about
0.5 mm h-1 for time intervals exceeding 3 days, and
degrades to near 3.5 mm h-1 at 3 h/0.25� scales. The
correlation coefficient can be as high as 0.8 for 12 h
averages, but only when the grid size exceeds 2.5�. Most
notably, the correlation begins to fall off quickly once the
time average drops below 1 day, and/or the spatial scale
falls under 1�, and this same reduction in performance is
evident in the RMSE, and less so in the mean bias. This
could be because the RMSE is more affected by the rela-
tively few large precipitation events, whereas the very
large number of light rainrate points (less than 2 mm h-1)
affects the correlation. As the time and/or space scale
shrink, there are fewer heavy rain events, and the overall
correlation is dominated by the large number of near-zero
rainrates. From analysis of 24 h accumulated precipitation
from a collection of 12 HRPPs, Ebert et al. (2007) showed
that for daily rain thresholds under 1 mm h-1, the false
alarm ratio (FAR) is about 0.5 and 0.3 in the middle and
tropical latitudes, respectively. When analyzed by season,
all HRPPs exhibited a reduced FAR and a smaller proba-
bility of detection (POD) in the winter months compared to
the summer months.
Exact cause of the HRPPs FAR and POD characteristics
are difficult to pinpoint. The tendency for the technique to
assign rain to radiometrically cold, non-precipitating
clouds is a main contributor to a poor FAR performance for
high rainrate thresholds. For small rainrate thresholds, one
possibility is related to the PMW algorithm rain/no-rain
screening. Under certain conditions, light rain can be
misidentified over a variety of Earth surfaces that appear to
scatter radiation similar to a precipitating cloud (Bauer
et al. 2002). In the NRL-Blend histogram matching pro-
cedure, these (falsely identified) light rain pixels get paired
with their corresponding IR TB, which is often higher
(warmer) than other localized pixels that were correctly
identified as rainfall. The end result can at times be a very
light rainfall that incorrectly gets assigned to regions in
subsequent IR imagery, until these falsely identified PMW
points are discarded after the next PMW overpass update.
The opposite effect also occurs, when the PMW algorithm
rain screen fails to identify regions of light rain. That is,
these no-rain pixels get paired with a lower (colder) IR TB
Fig. 5 Correlation, mean bias, and RMSE as a function of the time
offset (seconds) between the satellite overpass time and the raingauge
time. For example, if the GMS-5 satellite overpass occurred at 12:00
UTC, then a value of -10 means than all raingauge data between
11:50 and 12:00 UTC were used in the time averaging for all space/
time combinations. Three space-time combinations are shown, 3-h
time and 0.25� spatial (solid line), 6-h and 0.5� (dashed) and 24-h and
1� (dotted)
Validating a rapid-update satellite precipitation analysis 105
123
Page 8
compared to other localized pixels. The result is that the
zero-rain IR temperature threshold gets assigned too small
of a value and the lookup table assigns 0 or a very small
value to subsequent IR imagery. Any characteristics of the
PMW instantaneous rain algorithms (bias, rain screening,
etc) are retained in the blended technique processing until
corrected (i.e., erased from the lookup table) by a sub-
sequent PMW overpass.
On the other hand, the rapid decay of the correlation and
RMSE statistics may be reflecting the fact that below a
certain time scale, the blended technique estimates are
tuned to rainfall from a somewhat earlier stage in a the
rainfall evolution (especially for long PMW revisit times),
which correlate poorly with more recent observational data.
At time scales greater than 1 day, the correlation remains
quite high even at the finest spatial scales, which has been
noted in evaluation of other HRPPs (Gebremichael et al.
2005).
4 Conclusions
We have presented the results of a validation study of a
blended meteorological satellite High Resolution Precipi-
tation Product (HRPP) using the AWS network operated by
the KMA over the South Korean peninsula, during a 3-
month summer interval indicative of summer monsoon
conditions. The motivation for this effort was the belief
that a validation of the performance of a HRPP should be
carried out with a validation system whose overall time
sampling can resolve the precipitation scale and intensity
over short accumulation windows (minutes), and be
appropriately centered about the observation time of indi-
vidual satellite overpasses. The space-time RMS error,
mean bias, and correlation matrices were computed using
various time windows for the gauge averaging, centered
about the satellite observation time. For ±10 min time
window, a correlation of 0.6 was achieved at 0.1� spatial
scale by averaging over 3 days; coarsening the spatial scale
to 1.8� produced the same correlation by averaging over
1 h. Finer than approximately 24 h and 1� time and space
scales, respectively, a rapid decay of the error statistics
were obtained by trading off either spatial or time resolu-
tion. Beyond a daily time scale, the blended estimates were
nearly unbiased and with an RMS error of no worse than
1 mm day-1.
The validation of HRPPs is fundamentally constrained
by the nature of sporadic and intermittent rain falling over
a limited number of gauges at short time scales. By ana-
lyzing the 3-month period, there are many short time-scale
periods that are averaged together, some with intermittent,
sporadic rainfall and others with more widespread rainfall,
therefore effects related to rain inhomogeneity across the
box size should be averaged to some extent. To fully
examine the overall characteristics and performance of this
or another HRPP, a longer validation time interval is
needed, and the analysis should be pooled into tropical and
mid-latitude rainfall regimes, summer and winter seasons,
and 3-h local time windows (to examine if the HRPP is
capturing a diurnal cycle). The efforts being coordinated by
the International Precipitation Working Group (IPWG) and
its Pilot Evaluation of High Resolution Precipitation
Products (PEHRPP; Turk et al. 2008) are focused upon not
only such pooled analyses, but also regional-scale valida-
tion, and comparisons against model-forecasted precipita-
tion from several NWP models (Ebert et al. 2007).
Although it was not investigated, it seems plausible that
there is dependence between the overall PMW constella-
tion revisit and the performance of the NRL-Blend tech-
nique, below some minimum combination of space and
time scales. However, this may not be true for the more
modern advection-type HRPPs (Joyce et al. 2004; Ushio
et al. 2004) (which use geostationary data strictly to
propagate rainfall in between successive PMW overpasses)
and which have exhibited generally improved performance
compared to the calibration-type HRPPs such as NRL-
Blend. In addition, advanced geostationary imagers such as
SEVIRI (Schmetz et al. 2002) and the future GOES-R
Advanced Baseline Imager (ABI) are increasingly sophis-
ticated and their expanded spectral capabilities have yet to
be fully exploited for cloud microphysical, cloud phase,
and three-dimensional dynamical information (Levizzani
et al. 2007).
The ongoing development of HRPPs through groups
such as the IPWG has direct implications for future envi-
ronmental satellite missions. The proposed GPM is envi-
sioned to include various blended and merged precipitation
estimates, produced by a LEO constellation containing
different types of satellite systems and sensors (GPM
2008). To do so require a reference satellite (e.g., the GPM
core satellite) and a means to transfer the information
gathered from intersection of the reference with the con-
stellation members. The transfer should be done across
various space-time boundaries so as to best produce a final
global rainfall rate product consistent between all con-
stellation members.
Fig. 6 a Space-time contour plots of the correlation coefficient, root
mean square error and mean bias for AWS time window average of
±1 min, centered about the time of the GMS satellite observation of
Korea. The abscissa and ordinate of each contour plot denotes the
spatial and temporal scales, respectively, used to average the gauge
data and the blended satellite technique estimated rain. b Same as a,
but for an AWS time window average of ±5 min, centered about the
time of the GMS satellite observation of Korea. c Same as a, but for
an AWS time window average of ±10 min, centered about the time
of the GMS satellite observation of Korea
c
106 F. J. Turk et al.
123
Page 9
Correlation
(a)
(b)
(c)
Grid Size (deg)
0.1 0.2 0.3 0.4 0.5 1.0 1.5 2.0 2.5 3.0
Ave
ragi
ng P
erio
d
1hr
2hr
3hr
6hr
12hr
24hr
3day
5day
10day
15day
30day0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.6
0.5
0.5
0.50.4
0.4
0.4
0.3
0.3
0.2
0.2
0.4
0.1
0.3
Mean Bias
Grid Size (deg)
0.1 0.2 0.3 0.4 0.5 1.0 1.5 2.0 2.5 3.0
-0.2
-0.2
-0.2
-0.3
-0.3
-0.3
-0.4
-0.4
-0.5
-0.5
-0.6-0.7
-0.8-0.9
-0.1
RMS Error
Grid Size (deg)
0.1 0.2 0.3 0.4 0.5 1.0 1.5 2.0 2.5 3.0
2
2
4
68
1012
1816
14
Contour Map(AWS time window = +1 min.)
Correlation
Grid Size (deg)0.1 0.2 0.3 0.4 0.5 1.0 1.5 2.0 2.5 3.0
Ave
ragi
ng P
erio
d
1hr
2hr
3hr
6hr
12hr
24hr
3day
5day
10day
15day
30day0.9
0.8
0.8
0.7
0.7
0.7
0.6
0.6
0.6
0.6
0.5
0.5
0.5
0.4
0.4
0.3
0.30.2
0.4
0.1
Mean Bias
Grid Size (deg)0.1 0.2 0.3 0.4 0.5 1.0 1.5 2.0 2.5 3.0
-0.1
-0.2
-0.2
-0.2
-0.1
-0.1
-0.1-0.1
-0.3
-0.3
-0.4
-0.5
-0.6
-0.7
RMS Error
Grid Size (deg)0.1 0.2 0.3 0.4 0.5 1.0 1.5 2.0 2.5 3.0
1
1
2
3
4
5
7 68109
Contour Map(AWS time window = +5 min.)
Correlation
Grid Size (deg)
0.1 0.2 0.3 0.4 0.5 1.0 1.5 2.0 2.5 3.0
Ave
ragi
ng P
erio
d
1hr
2hr
3hr
6hr
12hr
24hr
3day
5day
10day
15day
30day0.9
0.8
0.8
0.8
0.8
0.7
0.7
0.7
0.7
0.6
0.6
0.6
0.5
0.5
0.4
0.4
0.6
0.3
0.2
Mean Bias
Grid Size (deg)
0.1 0.2 0.3 0.4 0.5 1.0 1.5 2.0 2.5 3.0
-0.1
-0.2
-0.1
-0.1
-0.1
-0.2
-0.2
-0.3
-0.3
-0.4
-0.5
RMS Error
Grid Size (deg)
0.1 0.2 0.3 0.4 0.5 1.0 1.5 2.0 2.5 3.0
1
1
2
3
45
6
Contour Map(AWS time window = +10 min.)
Validating a rapid-update satellite precipitation analysis 107
123
Page 10
Acknowledgments The first author acknowledges the support of the
research sponsors, the Office of Naval Research, Program Element
(PE-0602435N) and the National Aeronautics and Space Adminis-
tration (NASA) under grant NNG04HK11I. We acknowledge the
efforts of the Microwave Surface and Precipitation Products System
(MSPPS) at NOAA/NESDIS for the AMSU-B and MHS rainfall
datasets, and the TRMM Precipitation Processing System (PPS) for
the TMI and PR rainfall datasets.
References
Adler RF, Kidd C, Petty G, Morrissey M, Goodman HM (2001)
Intercomparison of global precipitation products: the third
Precipitation Intercomparison Project (PIP-3). Bull Am Meteorol
Soc 82:1377–1396
Arkin PA, Xie P (1994) The global precipitation climatology project:
first algorithm intercomparison project. Bull Am Meteorol Soc
75:401–419
Barrett EC, Dodge J, Goodman HM, Janowiak J, Kidd C, Smith EA
(1994) The First WetNet Precipitation Intercomparison Project.
Remote Sens Rev 11:49–60
Bauer P, Burose D, Schulz J (2002) Rain detection over land surfaces
using passive microwave satellite data. Meteorol Z 11:37–48
Ebert EE, Manton MJ, Arkin PA, Allam RJ, Holpin CE, Gruber A
(1996) Results from the GPCP Algorithm Intercomparison
Programme. Bull Am Meteorol Soc 77:2875–2887
Ebert EE, Kidd C, Janowiak J (2007) Comparison of near real-time
precipitation estimates from satellite observations and numerical
models. Bull Am Meteorol Soc 88:47–64
Gebremichael M, Krajewski WF (2004) Characterization of the
temporal sampling error in space-time averaged rainfall esti-
mates from satellites. J Geophys Res 109:D11110. doi:
10.1029/2004JD004509
Gebremichael M, Krajewski WF, Morissey ML, Huffman GJ, Adler
RF (2005) A detailed evaluation of GPCP 1-degree daily rainfall
estimates over the Mississippi river basin. J Appl Meteorol
44:665–681
GPM (2008) GPM science objectives. Global Precipitation Measure-
ment, NASA Goddard Spaceflight Center, http://gpm.gsfc.
nasa.gov/science.html
Grose A, Smith EA, Chung HS, Ou M-L, Sohn BJ, Turk FJ (2002)
Possibilities and limitations for QPF using nowcasting methods
with infrared geosynchronous satellite imagery. J Appl Meteorol
41:763–785
Huffman GJ, Adler RF, Bolvin DT, Gu G, Nelkin EJ, Bowman KP,
Hong Y, Stocker EF, Wolff DB (2007) The TRMM Multisat-
ellite Precipitation Analysis (TMPA): quasi-global, multiyear,
combined-sensor precipitation estimates at fine scales. J Hydro-
meteorol 8:38–55
Joyce R, Ferraro RR (2006) Improvements of CMORPH resulting
from limb adjustments and normalization of AMSU-B rainfall.
14th AMS Conference on Satellite Meteorology and Oceanog-
raphy, 29 Jan to 3 Feb, Atlanta
Joyce RJ, Janowiak JE, Arkin PA, Xie P (2004) CMORPH: a method
that produces global precipitation estimates from passive
microwave and infrared data at high spatial and temporal
resolution. J Hydrometeorol 5:487–503
Joyce R, Janowiak JE, Arkin PA, Xie P (2006) The combination of a
passive microwave based satellite rainfall estimation algorithm
with an IR-based algorithm. 14th AMS Conference on Satellite
Meteorology and Oceanography, 29 Jan to 3 Feb, Atlanta
Kummerow CD, Simpson J, Thiele O, Barnes W, Chang ATC,
Stocker E, Adler RF, Hou A, Kakar R, Wentz F, Ashcroft P,
Kozu T, Hong Y, Okamoto K, Iguchi T, Kuroiwa H, Im E,
Haddad Z, Huffman G, Ferrier B, Olson WS, Zipser E, Smith
EA, Wilheit TT, North G, Krishnamurti T, Nakamura K (2000)
The status of the Tropical Rainfall Measuring Mission (TRMM)
after two years in orbit. J Appl Meteorol 39:1965–1982
Levizzani V, Bauer P, Turk J (2007) Measuring precipitation from
space: EURAINSAT and the future. Springer, Berlin, p 722
Negri AJ, Bell TL, Xu L (2002) Sampling of the diurnal cycle of
precipitation using TRMM. J Atmos Ocean Technol 19:1333–1344
Oh HJ, Sohn BJ, Smith EA, Turk FJ, Seo AS, Chung HS (2002)
Validating infrared-based rainfall retrieval algorithms with 1-
minute spatially dense raingauge measurements over the Korean
peninsula. Meteorol Atmos Phys 81:273–287
Sapiano MRP, Arkin PA (2009) An intercomparison and validation of
high-resolution satellite precipitation estimates with 3-hourly
gauge data. J Hydrometeorol 10:149–166
Schmetz J, Pili P, Tjemkes S, Just D, Kerkmann J, Rota S, Raiter A
(2002) An introduction to Meteosat Second Generation (MSG).
Bull Am Meteorol Soc 83:977–992
Smith EA, Lamm JE, Adler R, Alishouse J, Aonashi K, Barrett E,
Bauer P, Berg W, Chang A, Ferraro R, Ferriday J, Goodman S,
Grody N, Kidd C, Kniveton D, Kummerow C, Liu G, Marzano
F, Mugnai A, Olson W, Petty G, Shibata A, Spencer R, Wentz F,
Wilheit T, Zipser E (1998) Results of WetNet PIP-2 Project. J
Atmos Sci 55:1483–1536
Soman VV, Valdes JB, North GR (1995) Satellite sampling and the
diurnal cycle statistics of Darwin rainfall data. J Appl Meteorol
34:2481–2490
Sorooshian S, Hsu K, Gao X, Gupta HV, Imam B, Braithwaite D
(2000) Evaluation of PERSIANN system satellite-based esti-
mates of tropical rainfall. Bull Am Meteorol Soc 81:2035–2046
Steiner M, Bell TL, Zhang Y, Wood EF (2003) Comparison of two
methods for estimating the sampling-related uncertainty of
satellite rainfall averages based on a large radar dataset. J Clim
16:3759–3778
Tian Y, Peters-Lidard CD, Choudhury B, Garcia M (2007) Multi-
temporal analysis of TRMM based satellite precipitation prod-
ucts for land data assimilation applications. J Hydrometeorol
8:1165–1183. doi:10.1175/2007JHM859.1
Turk J, Bauer P (2006) The International Precipitation Working
Group and its role in the improvement of quantitative precip-
itation measurements. Bull Am Meteorol Soc 87:643–647
Turk FJ, Miller SD (2005) Toward improving estimates of remotely-
sensed precipitation with MODIS/AMSR-E blended data tech-
niques. IEEE Trans Geosci Remote Sens 43:1059–1069
Turk FJ, Arkin P, Ebert E, Sapiano M (2008) Evaluating high
resolution precipitation products. Bull Am Meteorol Soc
89:1911–1916
Turk FJ, Mostovoy G, Anantharaj V (2009) Satellite rainfall
applications for surface hydrology. In: Gebremichael M, Hossain
F (eds) The NRL-Blend high resolution precipitation product and
its application to land surface hydrology, Chapter 6. Springer,
Berlin, p 445. ISBN 978-90-481-2914-0
Ushio T, Okamoto K, Isoda F, Iida Y, Aonashi K, Inoue T, Takahashi
N, Iguchi T, Hanado H, Iwanami K (2004) The Global Satellite
Mapping of Precipitation (GSMaP) Project: integration of
microwave and infrared radiometers for a global precipitation
map. Proceedings of 2nd international precipitation working
group, 25–28 October, Monterey. EUMETSAT, Darmstadt, p 44,
ISBN 92-9110-070-6. Online at http://www.isac.cnr.it/*ipwg
Vicente GA, Davenport JC, Scofield RA (2002) The role of
orographic and parallax correction on real time, high resolution
satellite rain rate observation. Int J Remote Sens 23:221–230
Xian P, Reid J, Turk FJ, Hyer EJ, Westphal D (2009) Impact of
modeled versus satellite measured tropical precipitation on
regional smoke optical thickness in an aerosol transport model.
Geophys Res Lett 36. doi:10.1029/2009GL038823
108 F. J. Turk et al.
123