Validating a rapid-update satellite precipitation analysis across telescoping space and time scales

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. E. Ebert

Centre for Australian Weather and Climate Research,

Bureau of Meteorology Research Centre, GPO Box 1289K,

Melbourne, VIC 3001, Australia


V. Levizzani

National Council of Research, Institute of Atmospheric Sciences

and Climate, ISAC-CNR, via Gobetti 101, 40129 Bologna, Italy


E. A. Smith

National Aeronautics and Space Administration, Goddard Space

Flight Center, Code 613.6, Greenbelt, MD 20771, USA


Present Address:F. J. Turk (&)

Jet Propulsion Laboratory, California Institute of Technology,

Pasadena, CA 91109, USA


123

Meteorol Atmos Phys (2009) 105:99–108

DOI 10.1007/s00703-009-0037-4

(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

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

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


123

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


123

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


123

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)


123

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


123

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


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



123

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


123

http://dx.doi.org/10.1029/2004JD004509

http://gpm.gsfc.nasa.gov/science.html

http://gpm.gsfc.nasa.gov/science.html

http://dx.doi.org/10.1175/2007JHM859.1

http://www.isac.cnr.it/~ipwg

http://dx.doi.org/10.1029/2009GL038823

Validating a rapid-update satellite precipitation analysis across telescoping space and time scales

Documents