ArticlePDF Available

Abstract and Figures

This paper presents an in-depth investigation of the error properties of two high-resolution global-scale satellite rain retrievals verified against rainfall fields derived from a moderate-resolution rain-gauge network (25-30-km intergage distances) covering a region in the midwestern U.S. (Oklahoma Mesonet). Evaluated satellite retrievals include the NASA Tropical Rainfall Measuring Mission multisatellite precipitation analysis and the National Oceanic and Atmospheric Administration Climate Prediction Center morphing technique. The two satellite products are contrasted against a rain-gauge-adjusted radar rainfall product from the WSR-88D network in continental U.S. This paper presents an error characterization of the Mesonet rainfall fields based on an independent small-scale, but very dense (100-m intergage distances), rain-gauge network (named Micronet). The Mesonet error analysis, although significantly lower than the corresponding error statistics derived for the satellite and radar products, demonstrates the need to benchmark reference data sources prior to their quantitative use in validating remote sensing retrievals. In terms of the remote sensing rainfall products, this paper provides quantitative comparisons between the two satellite estimates and the most definitive rain-gauge-adjusted radar rainfall estimates at corresponding spatial and temporal resolutions (25 km and 3 hourly). Error quantification presented herein includes zero- (rain detection probability and false alarm), first- (bias ratio), and second-order (root mean square error and correlation) statistics as well as an evaluation of the spatial structure of error at warm and cold seasons of 2004 and 2006.
Content may be subject to copyright.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010 1667
Benchmarking High-Resolution Global Satellite
Rainfall Products to Radar and Rain-Gauge
Rainfall Estimates
Emmanouil N. Anagnostou, Viviana Maggioni, Efthymios I. Nikolopoulos,
Tadesse Meskele, Faisal Hossain, and Anastasios Papadopoulos
Abstract—This paper presents an in-depth investigation of the
error properties of two high-resolution global-scale satellite rain
retrievals verified against rainfall fields derived from a moderate-
resolution rain-gauge network (25–30-km intergage distances)
covering a region in the midwestern U.S. (Oklahoma Mesonet).
Evaluated satellite retrievals include the NASA Tropical Rain-
fall Measuring Mission multisatellite precipitation analysis and
the National Oceanic and Atmospheric Administration Climate
Prediction Center morphing technique. The two satellite prod-
ucts are contrasted against a rain-gauge-adjusted radar rainfall
product from the WSR-88D network in continental U.S. This
paper presents an error characterization of the Mesonet rainfall
fields based on an independent small-scale, but very dense (100-m
intergage distances), rain-gauge network (named Micronet). The
Mesonet error analysis, although significantly lower than the
corresponding error statistics derived for the satellite and radar
products, demonstrates the need to benchmark reference data
sources prior to their quantitative use in validating remote sensing
retrievals. In terms of the remote sensing rainfall products, this
paper provides quantitative comparisons between the two satellite
estimates and the most definitive rain-gauge-adjusted radar rain-
fall estimates at corresponding spatial and temporal resolutions
(25 km and 3 hourly). Error quantification presented herein
includes zero- (rain detection probability and false alarm),
first- (bias ratio), and second-order (root mean square error and
correlation) statistics as well as an evaluation of the spatial struc-
ture of error at warm and cold seasons of 2004 and 2006.
Index Terms—Error, precipitation, rainfall estimation, remote
sensing.
Manuscript received February 16, 2009; revised July 24, 2009. First pub-
lished December 31, 2009; current version published March 24, 2010. This
work was supported in part by an EU Marie Curie Excellence Grant (project
PreWEC, MEXT-CT-2006-038331) and in part by NASA Precipitation Mea-
surement Mission Award NNX07AE31G. The work of E. I. Nikolopoulos was
supported by a NASA Earth System Science Graduate Fellowship. The work of
F. Hossain was supported by NASA New Investigator Award NNX08AR32G.
E. N. Anagnostou and E. I. Nikolopoulos are with the Department of Civil
and Environmental Engineering, University of Connecticut, Storrs, CT 06269-
2037 USA, and also with the Institute for Inland Waters, Hellenic Center for
Marine Research, 19013 Anavissos, Greece (e-mail: manos@engr.uconn.edu).
V. Maggioni is with the Department of Civil and Environmental Engineering,
University of Connecticut, Storrs, CT 06269-2037 USA.
T. Meskele is with Portland State University, Portland, OR 97207 USA.
F. Hossain is with the Department of Civil and Environmental Engineering,
Tennessee Technological University, Cookeville, TN 38505-0001 USA.
A. Papadopoulos is with the Institute of Inland Waters, Hellenic Center for
Marine Research, 19013 Anavissos, Greece.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TGRS.2009.2034736
I. INTRODUCTION
R
AINFALL exhibits high spatio-temporal variability that
affects the response of terrestrial hydrologic processes
such as soil moisture, evapotranspiration, and heat fluxes,
as well as the generation of runoff. The accurate prediction
of these dynamic surface hydrologic states requires accurate
data on rainfall distribution at the highest possible resolution.
Traditionally, rainfall observations have been limited to those
continental areas where rain gauges or weather radar systems
are available. However, measurements from these observational
systems have also been plagued with various problems that have
undermined their widespread and continuous use in hydrologic
modeling. For example, gage networks are typically very sparse
over important climatic regions like the tropical rain forests and
mountainous areas. Weather radars, on the other hand, bring
advancements to precipitation monitoring due to spatially dis-
tributed information but have limitations arising from signifi-
cant uncertainties associated with rain-path attenuation, the lack
of uniqueness in the reflectivity-to-rainfall (ZR) transforma-
tion, radar calibration and contamination by ground return prob-
lems, subresolution precipitation variability, and vertical profile
and complex terrain effects (discussions on those i ssues can be
found in [5], [27], [33], and others). In addition, radar and gage
monitoring systems require considerable financial and techno-
logical investment to operate and maintain on long-term and con-
tinuous basis. This probably explains the absence of historical
rainfall observations from such systems in developing nations.
Given the deficiencies in conventional surface networks for
hydrologic monitoring, observations from spaceborne instru-
mentation currently constitute the only truly viable mean to pro-
mote our understanding of terrestrial hydrology over the vast
regions on Earth [22]. Satellite observations have been used
regularly since the 1970s to extract precipitation/rainfall infor-
mation at global scale [17]. These observations cover a range of
wavelengths within the visible (VIS) and infrared (IR) spectrum
gathered from satellites in geostationary (GEO) and low earth
orbits (LEOs), and both passive (radiometers) and active (radars)
measurements gathered at microwave (MW) frequencies ob-
tained from LEO satellite sensors. With more GEO and LEO or-
biting satellites being launched steadily since the 1990s, global
satellite rainfall estimation has exhibited major growth and ad-
vancement during the last decade, particularly with the advent
of the Tropical Rainfall Measuring Mission (TRMM) [38].
0196-2892/$26.00 © 2009 IEEE
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
1668 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
VIS and IR sensors on GEO have a long history, and rain
estimation techniques based on those measurements generally
relate the cloud-top brightness temperatures (T
b
)’s to sur-
face and/or other satellite observations. Such techniques range
from simple time-integral techniques, such as the Geostation-
ary Operational Environmental Satellite Precipitation Index
(see, e.g., [3]), to neural networks [42]. Newly developed
multispectral s ensors, such as the Moderate Resolution Imaging
Spectroradiometer (in LEO) and Spinning Enhanced Visible
and InfraRed Imager (in GEO), now permit the microphysical
properties of clouds to be identified [37], leading to the prospect
of improving rainfall retrievals. GEO multispectral VIS/IR
images can be acquired at a nominal 15-min/3–4-km resolution
(1-km VIS), permitting the monitoring of precipitating cloud
systems and their evolution over time. However, VIS/IR tech-
niques cannot directly retrieve precipitation from the observa-
tions of cloud tops.
MW sensors on LEO can measure the modulation by hydro-
meteors (i.e., precipitation-sized ice and water particles) on
upwelling radiation from Earth more accurately: Techniques
have been developed to physically link the signal received by
the satellite sensors to the size and phase of the hydrometeors
present within the observed atmospheric column (foundations
of these techniques can be found in [29]–[31], [35], and [39]).
These algorithms use radiative transfer computations on the
basis of cloud model simulations to generate large databases
of simulated satellite observations and coincident precipitation
profiles. The inverse solution is based on a Bayesian framework
that identifies the database profile (or set of profiles) with
simulated brightness temperatures closest to the satellite obser-
vations. This approach is mainly used for over ocean precip-
itation retrievals where the uniformly cold ocean background
allows the use of lower frequency (10–19-GHz) channels. Over
land, these low-frequency channels cannot be used because of
the relatively warm and nonhomogeneous land background.
Overland retrievals are based on statistical algorithms [8], [11],
[16] that use brightness temperature depression at 85 GHz as
a primary input, which occurs due to scattering by ice above
the freezing level. These retrievals have a host of uncertainty
sources, including the spatio-temporal variability in the back-
ground land surface; the warm rain process that could give
significant amounts of surface rain but may not produce enough
ice aloft to be detected at 85 GHz; surface snow cover that
could have similar signature as ice aloft, consequently confused
with precipitation; the beam filling problem that is particularly
important in cases of small-scale convective systems; and the
complex and highly variable relationship between the ice aloft
and the rainfall rate at the surface (see [9] for a detailed
evaluation of MW overland retrieval techniques).
By using various combinations of observations from com-
plementary satellite sensors (MW and VIS/IR), a large number
of high-resolution global rainfall products are now available
in near real time. Some pertinent examples from the U.S.
science teams are PERSIANN, a neural-network-based estima-
tion technique based on IR data produced by the University
of California–Irvine [18], [41]; CMORPH, an IR-based mor-
phing of successive MW rainfall estimates developed at the
National Oceanic and Atmospheric Administration (NOAA)
Climate Prediction Center [25]; and TMPA (TRMM–Multi-
satellite Precipitation Analysis), which is a NASA Goddard
Space Flight Center TRMM radar-based calibration scheme
for IR and combined (IR–MW) rainfall estimates [24]. Several
other products exist from research groups worldwide, such
as the Global Satellite Mapping of Precipitation (GSMAP)
[28] product, which is the official high-resolution global pre-
cipitation product of Japan Aerospace Exploration Agency.
Other examples are the MW-calibrated I R techniques of the
U.S. Naval Research Laboratory [44] and the University of
Birmingham in the U.K. [26]. Ebert et al. [10] provides a more
encompassing list of the currently available global-scale high-
resolution satellite rainfall products.
With multiple satellite rainfall products currently available,
it is important that the overland retrieval uncertainty is bench-
marked as a function of region and season for advancing
their terrestrial applications. Herein, benchmarking refers to a
rigorous flagging of the quality of satellite estimates against the
assumption that a considerably higher quality data set is already
available, which may either be from a dense gauge network
or a calibrated ground radar system. This benchmarking is
important because a critical aspect of satellite data used in
hydrological applications is the need to resolve the precipitation
variability at high temporal ( 3 h) and spatial ( 25 km
2
)
scales. Hydrologists and other users, to varying degrees, need
to know the errors of the satellite rainfall data sets at those high
spatio-temporal scales.
Recognizing the need for quantifying the uncertainty, several
recent studies have been initiated to compare the accuracy
of various satellite rainfall products over land. For example,
the International Precipitation Working Group (IPWG) has
assessed (and continues to assess) six widely available satellite
data products over continental regions of the globe [10]. This
IPWG agenda has led to a more detailed comparison of uncer-
tainty of satellite products called the Pilot Evaluation of High
Resolution Precipitation Products (see [4]). Hong et al. [19] and
Gottschalck et al. [15] have evaluated different high-resolution
(0.25
spatial resolution) satellite rainfall products at daily to
seasonal time scales for hydrologic and land data assimilation
applications, respectively. Other examples of assessing satellite
rainfall uncertainty include the studies of Gebremichael
and Krajewski [13], [14] on sampling errors; the study of
McCollum et al. [34] on the assessment of bias; and the s tudy
of Ali et al. [1] on satellite error functions for the Sahel region.
The most recent satellite rainfall error s tudies are those of
Tian et al. [43] and Hossain and Huffman [21]. Those investiga-
tions presented important conclusions about various error
metrics determined at a range of scales for a few key global high-
resolution satellite data sets over the continental U.S. region.
With the planned Global Precipitation Measurement mission
(see [39] and http://gpm.gsfc.nasa.gov) and the continued pro-
gression toward smaller scales (5–10 km and hourly) relevant
to hydrology, there is a need to reevaluate satellite products
in a manner that reveals more insightful information on the
reliability of those data sets in distributed hydrologic and
land surface modeling, among other uses [23], [32]. Hossain
and Lettenmaier [22] have argued that a shift in paradigm is
needed to properly assess estimates of rainfall from satellite
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
ANAGNOSTOU et al.: BENCHMARKING HIGH-RESOLUTION GLOBAL SATELLITE RAINFALL PRODUCTS 1669
Fig. 1. Oklahoma Mesonet (black dots on the Oklahoma map) and Micronet (locations shown on the Little Washita River watershed in the expanded diagram)
networks. The stations marked with red color are retired since 2005. The 25-km grid cells shown in the same graph represent the locations where the Mesonet
atmospheric and soil moisture data are interpolated to match the spatial distribution of the satellite rainfall products. The rectangular grid defines the area of the
accumulation plots shown in subsequent figures.
sensors for modeling dynamic hydrologic processes such as the
rainfall-runoff transformation and associated energy and mois-
ture fluxes. Studies have shown that errors in rainfall estimation
propagate nonlinearly through the simulation of runoff (see,
e.g., [5], [20], and [36]). Understanding of this nonlinear error
propagation effect will depend on how well we can capture the
uncertainty structure to the level that matters for hydrologic
variability.
In this paper, we extend the works of Tian et al. [43] and
Hossain and Huffman [21] by focusing on a small-scale domain
in the midwestern U.S., capturing the State of Oklahoma. The
region is covered by a dense network of meteorological sta-
tions (that include sensors on soil moisture), named Oklahoma
Mesoscale Network (Mesonet [6]), and high-resolution radar
rainfall fields from the National Weather Service WSR-88D
network available over a long-term period (1997–2006). The
Mesonet station data can be used as a reference for evaluat-
ing satellite rainfall retrievals at high spatio-temporal scales
as well as land surface model simulations of soil moisture
and land surface heat fluxes. The satellite rainfall techniques
evaluated and compared on the basis of this investigation are the
TRMM 3B42 (V.6), which is a gauge-adjusted satellite product
on monthly basis, and CMORPH, an IR-based morphing of
sequential MW r ain snapshots. The techniques are evaluated
against the Oklahoma Mesonet station measurements. The error
statistics determined for the satellite rainfall products are com-
pared against those evaluated from comparisons of the Mesonet
gauges with rain-gauge-adjusted WSR-88D rainfall estimates
(stage IV). This paper presents a rigorous benchmarking of the
Mesonet station measurements as for its accuracy in deriving
area rainfall estimates at the resolution of satellite products
(0.25
/3 hourly) based on comparisons against a dense rain-
gauge network (Micronet: http://grl.ars.usda.gov/micronet/) lo-
cated over a small area (600 km
2
) covering the Little Washita
River Experimental Watershed in southwestern Oklahoma. In
Section II, we present the s tudy area, data, and evaluation
methodology devised in this paper. Results on error statistics
for the satellite retrievals and radar are presented in Section III,
while in Section IV, we discuss the implications of our compar-
ison and conclude with the major findings while closing with
future directions of this paper.
II. D
ATA AND METHODOLOGY
A. Study Area
Comparisons of rainfall estimates were made over the
Oklahoma region in the midwestern U.S., located in the latitude
range of 32
–39
N and longitude range of 93
to 102
W
associated with an area of about 158 000 km
2
. The region
is characterized by continental climate associated with cold
winters and hot summer seasons, while its topography rises
gently to the west from an altitude of 88 m.a.s.l. (meters
above sea level) in the southeastern corner to a height of 1515
m.a.s.l. on the tip of the “panhandle.” The region of Oklahoma
has been chosen as the study area because of its mild
terrain, good coverage by the WSR-88D National Weather
Service radar network, and dense in situ rainfall measure-
ments from the Oklahoma Mesonet and Micronet stations
(http://mesonet.ou.edu). The study area and locations of
Mesonet and Micronet stations are shown in Fig. 1. In Fig. 1,
we present a Cartesian grid showing the 25 × 25 km
2
common
domain grid over which radar, gauge, and satellite rainfall
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
1670 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
products are projected. The high density of Micronet stations
(Fig. 1) allows capturing the subgrid variability of rainfall in
one of the 25-km grid cells, which is used to assess error
statistics of the coarser Mesonet network in capturing the
3-hourly average rainfall over that grid cell. Details about the
data sets are provided next.
B. Data
Radar rainfall fields are extracted from the Stage IV National
Weather Service precipitation estimation algorithm prod-
uct that involves real-time adjustment of the radar rainfall
estimates based on mean-field radar rain-gauge hourly ac-
cumulation comparisons and merging of hourly radar with
gauge-interpolated rainfall fields [12]. The Oklahoma Mesonet
system provides meteorological observations with high spatial
resolution and temporal frequency [6]. It consists of 110 au-
tomated observing stations located throughout the state and
provides 5-min temporally aggregated surface meteorological
data. An additional data set available in this paper is the Little
Washita River Experimental Watershed Micronet, which is
5-min meteorological and 15-min soil temperature data from
42 stations covering an area of 610 km
2
. Both data sets are
quality controlled and flagged for bad-quality data, which is
very important because a significant component for a successful
research study is the requirement of high-quality data. Data are
available for the period 1997–2006 for the Mesonet stations
and for the period 2002–2004 for the Micronet. Radar rainfall
fields and rain-gauge data were interpolated to the satellite
grid domain having a 25-km grid resolution and aggregated to
3-hourly time scale.
Two different satellite rainfall products were used in this
paper for evaluation. The first one is the NOAA Climate Predic-
tion Center morphing method (CMORPH), which is a satellite
rainfall algorithm that uses motion vectors derived from half-
hourly interval GEO satellite IR imagery to propagate the
relatively high quality rainfall estimates obtained from LEO-
based MW sensors [25]. The dynamic morphological character-
istics (such as shape and intensity) of the precipitation features
are morphed at consecutive times between microwave sensor
samples by performing a time-weighted linear interpolation.
This process yields spatially and temporally continuous MW
rainfall fields that have been guided by IR imagery and yet is
independent of any IR temperature-based inversion to rainfall
rate. It can be argued that CMORPH uses a Lagrangian (system-
following) framework of reference for rainfall estimation.
The second technique is the TRMM Multi-satellite Precip-
itation Analysis (named as 3B42) that provides a calibration-
based sequential scheme for combining rainfall estimates from
various satellites, as well as gauge analyses where feasible, at
fine (0.25
× 0.25
and 3 hourly) space–time scales [24]. It is
available both as postanalysis (3B42 V.6) and in real time (3B42
RT), based on calibration by the TRMM Combined Instrument
and TRMM Microwave Imager precipitation products, respec-
tively. Only the postanalysis product (3B42 V.6) incorporates
a monthly gauge analysis (Global Precipitation Climatology
Center up to May 2005 and Climate Anomaly Monitoring
System from May 2005 forward) to create a monthly satellite–
Fig. 2. Sample sizes of the (blue) cold and (red) warm periods for (solid lines)
2004 and (dashed lines) 2006.
gauge estimate, which is then used to scale the individual
3-hourly values to approximately add up to the satellite–gauge
combined value. The 3B42 (both V6 and RT) data set covers
the latitude band 50
N–S for the period from 1998 to the
present. In this paper, we used 3B42 (V.6) to account for the
best possible satellite data set.
C. Methodology
This paper focuses on the performance assessment of the
two global high-resolution satellite rainfall products relative to
the nominally more accurate ground radar rainfall estimates
using as ground truth the area-interpolated fields from Mesonet
stations. In addition, the ability of these algorithms to determine
the rainfall intensity distributions as a function of seasons is
evaluated. The study period involves two distinct years in 2004
and 2006. In order to study the seasonal effect, the analy-
sis has been performed on the basis of two distinct periods:
1) cold season (November to April) and 2) warm season (May
to October).
To compare the different rainfall data sets, all products
needed to be scaled at the same spatial and temporal resolution.
The common resolution of 25 × 25 km
2
and 3 hourly has been
chosen to be the nominal resolution of the satellite products.
The data sets available at higher spatio-temporal (WSR-88D
radar data) or only temporal (gauges: 5 min) resolutions were
aggregated to this level (25 km/3 hourly). For instance, aggrega-
tion in space and accumulation in time has been performed for
the WSR-88D (i.e., NEXRAD Stage IV) rainfall fields being
at 4-km/hourly spatio-temporal resolution. The global satellite
rainfall products were simply cropped to the study area in the
latitude range of 34.5
–37
N and longitude range of 100
to
94.5
W and projected in the 25-km grid cells. Fig. 1 shows
the extent of the common grid domain and its grid spacing
relative to the Mesonet gauge network.
The ground meteorological stations of Mesonet provide ac-
curate estimates, but only at the point scale, which may not
accurately represent the mean rainfall over a 25 × 25 km
2
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
ANAGNOSTOU et al.: BENCHMARKING HIGH-RESOLUTION GLOBAL SATELLITE RAINFALL PRODUCTS 1671
Fig. 3. (Left) Mean and (right) standard deviation of Mesonet stations’ rainfall for the (blue) cold and (red) warm periods of (solid lines) 2004 and (dashed
lines) 2006.
grid area. On the other hand, radar and satellites provide an
area-average value. Thus, a spatial interpolation of ground mea-
surements was applied to moderate this area-to-point difference
effect. The inverse distance technique was chosen as the inter-
polation method. Alternative area interpolation schemes such
as spatial kriging was considered, but not deemed important
since the gage distances are at about the spatial correlation
length of 3-hourly rainfall as determined by the available gauge
rainfall data. The uncertainty of this interpolation method is
investigated in this paper on the basis of comparisons of the in-
terpolated Mesonet estimates with corresponding area-average
rainfall values derived from the dense Micronet stations con-
centrated in a single 25 × 25 km
2
grid cell in south Oklahoma.
D. Description of the Reference Data Set
The plot in Fig. 2 shows the number of data samples with
values greater than a threshold rainfall value G
th
for the two
seasons (cold and warm) and two years (2004 and 2006),
while Fig. 3 shows the mean and standard deviation of the
corresponding gauge data samples (Mesonet stations) in those
periods. In general, the sample size gets smaller as the rainfall
threshold increases. This means that statistics generated based
on these data would become less reliable as the threshold
value increases due to sample-size limitations. Both cold and
warm seasons of 2004 have similar sample sizes for all rainfall
thresholds, but the sample size of the cold season during 2006
is almost 50% lower than the warm season of the same year and
the cold season of the 2004. Taking into account that the mean
(Fig. 3) is similar for all seasons and both years, this indicates
that the cold season of 2006 was relatively dry compared to
the rest of the seasons examined in this paper. The standard
deviation shows higher values for the warm relative to the cold
season, which is associated with the higher frequency of con-
vective precipitation during the warm season. The difference in
variability between seasons is more pronounced for 2006.
A point to note is that the error statistics to be derived for
the remote sensing data sets depend on the accurate knowledge
of the “actual” grid-cell average rainfall intensity; therefore,
analysis of the reference data (i.e., the interpolated Mesonet
stations’ rainfall fields) is necessary. This is performed based on
comparisons against grid-cell average rainfall fields from a grid
cell in the southwestern Oklahoma that is densely covered by
the Micronet stations (i.e., 50 stations covering a 25 × 25 km
2
grid-cell area). The Micronet-derived grid-cell average rainfall
is used here as a reference to determine the error statistics
of the interpolated Mesonet estimates for that grid cell at
3-hourly accumulations. Fig. 4 shows the scatter plot of the
Mesonet versus Micronet grid-cell 3-hourly average rainfall
values and the cumulative density functions (CDFs) of falsely
detected Mesonet and nondetected Micronet rainfall values.
As noted from the scatter plot, the correlation of Mesonet
to the Micronet station estimates is very high for both warm
and cold seasons ( 0.95). The Mesonet-to-Micronet bias ratio
is one for cold season and 0.94 for the warm, indicating a
slight underestimation of the sparse Mesonet stations in the
case of spatially variable convective storms. The Mesonet rain
detection relative to Micronet is about 60%. Although this may
seem a low detection score, as noted from the CDF plot of
Fig. 4 (lower panel), the majority (> 99%) of the nondetected
Micronet rainfall values are below 0.1 mm/h, which correspond
to 4% (8%) of the overall rain volume for the warm (cold)
season. On the other hand, the false alarm rate (FAR) is very
low (1% and 3% for the cold and warm seasons accordingly).
The corresponding CDFs of Fig. 4 (middle) show that the
majority (> 99%) of the falsely detected Mesonet rainfall
values are below 0.1 mm/h (0.2 mm/h) for the cold (warm)
season, which associate with 4% of the rainfall volume for
both warm and cold seasons. The root mean square difference
between Mesonet and Micronet rainfall rates is 0.32 mm/h
(0.67 mm/h) in the cold (warm) season, which, in relative
terms, is 32% (39%) of the Micronet rainfall standard deviation.
The earlier analysis indicates that the Mesonet station grid-
cell average rainfall fields are associated with errors due to the
subgrid scale variability of rainfall that should be considered
in interpreting the quantitative error statistics of the remote
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
1672 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
Fig. 4. (Top) Scatter plot of Mesonet versus Micronet 3-hourly grid-cell
average rainfall values. (Middle) CDF of the falsely detected Mesonet rain-
fall (i.e., when Micronet is zero). (Bottom) CDF of the Micronet rainfall
not detected by Mesonet. Red color corresponds to warm season, and blue
corresponds to cold season.
sensing products presented in the following section. Past
ground validation studies have presented statistical techniques
for separating t he remote sensing rainfall error (assuming
TAB LE I
C
ONTINGENCY TABLES FOR THE EVA L UAT I O N O F
(a) POD AND (b) FAR STATISTICS
Gaussian error distribution) from the rain-gauge representative-
ness error caused by the natural variability of rainfall in the
subgrid scale (see, e.g., [2] and [7]). Our intent in this paper is to
use the Mesonet error statistics on a qualitative basis. Namely,
we will use them as a benchmark to argue that the error statistics
of the remote sensing products are significantly higher to justify
the use of Mesonet as a reference to those products.
E. Error Statistics
The basic descriptive statistics devised in this paper to pro-
vide information on the rain detection are through the con-
tingency table approach [see Table I(a) and (b)]. This binary
classification approach can help determine the probability of
detection (POD) and false alarm detection by the sensors (both
radar and satellite algorithms) as a function of a varying rainfall
threshold either by the reference or the remote sensor (R
thr
).
Specifically, in Table I(a) and (b), RS represents the remotely
sensed product (the estimate), which can be from satellite or
radar, and G is the ground truth (i.e., the Mesonet station). In the
case of Mesonet evaluation from Micronet stations, RS takes
values from Mesonet, while G is the area rainfall determined
from Micronet measurements. Variable A in Table I(a) and (b)
represents the number of hits or number of rainy grid points
correctly detected by the sensor, while B stands for the number
of misses or number of rainy grid points not detected by the
sensor. C is the number of false alarms or number of nonrainy
grid points estimated as rainy by the sensor, while D is the total
number of nonrainy grid points correctly detected by the sensor.
The POD of the sensors can be calculated from Table I(a), for
each rainfall threshold applied to the reference values, which is
given by
POD =
A
A + B
. (1)
The FAR (FA) can be retrieved from Table I(b), for each
rainfall threshold applied t o the RS values, which is given by
FA =
C
C + D
. (2)
In addition to t he POD and FAR binary detection statistics,
we evaluate the corresponding fractional rain volume ratios
of missed and falsely detected reference rain values by the
different RS estimates as a function of rain threshold.
Since the aforementioned statistical measures do not use
the magnitude of the rainfall differences between sensor and
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
ANAGNOSTOU et al.: BENCHMARKING HIGH-RESOLUTION GLOBAL SATELLITE RAINFALL PRODUCTS 1673
Fig. 5. Rainfall accumulation (in millimeters) maps for the Oklahoma region during the cold season months’ period of 2004 derived from (upper left) Mesonet,
(upper right) WSR-88D, (lower left) CMORPH, and (lower right) 3B42 (V.6).
reference, they are not strictly influenced by the variability of
estimation error. To measure the magnitude of the difference
between sensor estimates and reference rainfall, we calculated
the root mean square error (rmse) and bias ratio statistics as
follows:
RMSE
RS
i
=
N
t
t=1
N
u
u=1
(RS
i
(t, u) Ref(t, u))
2
N
t
× N
u
(3)
Bias Ratio
RS
i
=
N
t
t=1
N
u
u=1
RS
i
(t, u)
N
t
t=1
N
u
u=1
Ref(t, u)
(4)
where RS
i
(t, u) and Ref(t, u) are the different sensor-
estimated and reference rainfall values, respectively, at 3-hourly
time step t and grid location u, and N
t
× N
u
is the total num-
ber of observations reaching or exceeding a certain reference
rainfall threshold amount. Other statistical measure used here
to evaluate the consistency of the RS technique estimates to
reference rainfall is the correlation. Correlation is evaluated
in two ways. First, we evaluate the correlation between the
RS estimates and reference rainfall values for matched pairs
associated with reference rainfall values exceeding a varying
threshold. Second, we determine and compare the spatial cor-
relation of the rainfall maps estimated by the remote sensing
techniques and evaluated from the reference data sets.
Combining the aforementioned statistical criteria, we attempt
to provide a comprehensive evaluation of the satellite technique
performances relative to the radar rainfall estimates for the
region of Oklahoma. For example, a greater POD will represent
a technique improvement only if it is accompanied by a low
FAR and bias with a value close to one and a lowering rmse.
Furthermore, good performance of the RS technique would
imply spatial rainfall structures with similar correlation lengths.
The skill scores are presented for product resolutions at 25-km
and 3-hourly periods. Using rain accumulations from longer
periods (e.g., daily) and coarser resolutions, although could
provide improved statistics, would result i n smoothing the
local scale details occurring in high-resolution data. On the
other hand, evaluating the remote sensing estimates at finer
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
1674 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
Fig. 6. Same as in Fig. 5 but for the warm season months’ period of 2004.
space–time scales is not currently possible for some of the
global satellite rainfall products. Some of the techniques are
now producing rainfall fields at finer scales (e.g., CMORPH,
PERSIANN, and GSMaP). Hossain and Huffman [21] have in-
vestigated scale dependences of error metrics similar to the ones
presented herein for one of those techniques (i.e., PERSIANN).
III. R
ESULTS
A. Rainfall Accumulation Maps
The seasonal cumulative rainfall for each cell of the grid
was determined by aggregating the precipitation occurred in
the time steps where all the sensors had valid measurements
at the scale of 25-km grid cell. The cumulative rainfall maps
are shown in Figs. 5–8. Similar visual (qualitative) patterns
can be observed from all sensor retrieval maps. Specifically, all
sensors, for both seasons, capture a rainfall gradient from the
drier areas located in the western/northwestern part of the
region to the wetter areas in the eastern/southeastern part.
The WSR-88D exhibits the highest pattern correlation with
the reference rainfall map (MESONET) relative to the satellite
retrievals. For the satellite products, CMORPH exhibits sig-
nificant overestimation in the warm season months, which is
consistent in both years, while this systematic overestimation is
significantly lower in the cold season. The 3B42 (V.6) rainfall
patterns are very similar to the radar patterns and, generally,
in good agreement with the MESONET throughout the two
seasons and years. This is mainly due to the adjustment applied
to these rainfall estimates on the basis of rain-gauge rainfall
measurements. Finally, all rainfall products show lower accu-
mulated rainfall in the cold season relative to the warm season,
which is consistently depicted in the reference rainfall product
(Mesonet).
Comparisons of the various remote sensing products to
Mesonet are also presented in the form of s catter plots of
warm and cold season cumulated rainfall shown in Fig. 9. A
point to note is the higher values of rainfall accumulations in
both warm and cold seasons of 2004 consistently depicted on
all remote sensing products and Mesonet. Arguably, the radar
product (WSR88D) shows stronger correspondence with the
Mesonet than the satellite products, although, 3B42 (V6) scatter
is not much worse, particularly in the warm season. CMORPH
is significantly biased in the warm season, while in the cold
season, it gives similar scatter spread with 3B42 (V6). Although
the cumulated plots show good consistency between the various
products (at least for the 3B42 V.6 and WSR88D), this may not
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
ANAGNOSTOU et al.: BENCHMARKING HIGH-RESOLUTION GLOBAL SATELLITE RAINFALL PRODUCTS 1675
Fig. 7. Same as in Fig. 5 but for the cold season months’ period of 2006.
be true at the short time scales (3 hourly), which is the subject of
this paper. Quantitative error statistics for the 3-hourly/25-km
grid-cell average rainfall products are discussed next.
B. Error Statistics
In Fig. 10, we show the probability of rain detection (POD)
values of the different products for the two seasons (cold and
warm) and years (2004 and 2006) conditioned to reference
rainfall values greater than a threshold. The results presented
show that the 3B42 (V.6) product provides consistently the
lowest (highest) rain detection efficiency (fractions of missed
rainfall to the overall rain volume) across the whole range
of rainfall thresholds. The radar (WSR88D) estimates, on the
other hand, exhibit the highest probability of successful rain
detection and, correspondingly, the lowest missed rain volume
fractions during the cold season for both years, while during
the warm season, CMORPH and WSR-88D exhibit similar
rain detection behavior. Both of these products present PODs
(missed rain volume fractions) that exceed 0.8 (is below 8%)
for threshold rain rates greater than 0.2 mm/h. From the same
figures, one can note that the POD of 3B42 (V6) products
improved from 2004 to 2006, showing a higher POD, particu-
larly during the cold season, with correspondingly lower missed
rain volumes.
In Fig. 11, we show the probability of false rain detection
(i.e., FARs) described by the probability of the sensor retrieval
to be greater than a threshold rainfall value, when the reference
sensor (Mesonet) indicates zero rainfall. The figure shows that
CMORPH yields the highest FARs in the warm season of 2004,
reaching values close to 0.10 for low rain rates (< 0.1 mm/h).
The 3B42 (V.6) and WSR-88D products generally maintain
lower FARs across all seasons, with the radar product exhibiting
the lowest of the three. In terms of the fraction of rain volume
of the falsely detected rain values, the CMORPH exhibits
significantly higher values than the other two products in warm
season, but a significant improvement from 2004 to 2006 (re-
duction of nearly 32%). The 3B42 (V.6) product exhibits values
slightly higher than the radar product at threshold rainfall values
below 0.5 mm/h, while the two products converge at higher
(> 1.5 mm/h) rainfall thresholds.
The bias ratio statistic, defined as the ratio between the
sensor estimates to the Mesonet, is shown in Fig. 12 (upper
panels). WSR-88D is the product with the lowest bias ratio
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
1676 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
Fig. 8. Same as in Fig. 5 but for the warm season months’ period of 2006.
(close to one) for both seasons and years. Furthermore, we note
that the bias ratio is independent of rainfall threshold. In the
cold season, the satellite products show similarly low biases
(bias ratios close to one), which are relatively independent of
rainfall threshold. The 3B42 moderately underestimates rainfall
in both years at similar level (slightly increased bias ratio in
2006). CMORPH, on the other hand, shows a low underesti-
mation of rainfall in 2004, which t urns to moderate (20%)
overestimation in 2006. In the warm season, the picture for
the satellite products is very different. The 3B42 (V.6) product
exhibits stronger bias dependence on rain threshold, indicating
that the satellite product underestimation increases relative to
rainfall intensity. This is consistently shown for both years with
2006 exhibiting an overall stronger underestimation compared
to 2004. The CMORPH product shows overestimation, which is
stronger in low rainfall values and reduces as rainfall threshold
increases. Similar to 3B42 (V.6), the bias ratio of this product
also worsened from 2004 to 2006.
The correlation coefficient statistic, which indicates the
strength of the linear relationship between the sensor retrievals
and the reference, is shown in the same figure (middle panels).
As expected, the WSR-88D yields the strongest correlation
among the different remote sensing products, with a significant
difference in magnitude than the correlation values reported
for the other satellite products. The lowest correlation is of
the 3B42 (V6) product, which is consistently lower than the
CMORPH product. The difference between the 3B42 (V6) and
CMORPH correlations reduces from 2004 to 2006. The warm
season correlations are slightly higher that the cold season
correlations, which is consistently shown for all products and
both years. Strong magnitude dependence on correlation is
apparent, with lower rainfall thresholds (< 0.4 mm/h) exhibit-
ing higher correlation values. An interesting feature is that
the shape of the correlation as a function of rainfall threshold
transitions from a highly linear-like (cold season) to a more
exponential-like (warm season), indicating the seasonal depen-
dence of its functional form.
The last error statistic used in this paper is the rmse ( 3)
shown in t he bottom panels of Fig. 12. A first observation from
this figure is that radar rainfall product has the lowest r mse rel-
ative to the satellite products in both warm and cold seasons and
both years. Specifically, the radar rmse increases as a function
of the rain rate threshold to a value of 1.5 mm/h at rain rate val-
ues exceeding the 2-mm/h threshold, which is, in relative terms,
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
ANAGNOSTOU et al.: BENCHMARKING HIGH-RESOLUTION GLOBAL SATELLITE RAINFALL PRODUCTS 1677
Fig. 9. Scatter plots of Mesonet versus (upper panels) WSR88D, (middle panels) 3B42 V.6, and (lower panels) CMORPH warm and cold season rainfall
accumulations (in millimeters) at 25-km grid-cell resolution.
equal to 42% of the mean of reference rainfall, and remains at the
same level for both seasons and years. Comparing the two sat-
ellite products, 3B42 (V6) exhibits lower rmse than CMORPH,
particularly during the warm season where CMOPRH product
is associated with significant biases (shown on the top panels
of Fig. 12). Specifically, the 3B42 (V6) rmse values vary from
about 1.3 mm/h at 0.01-mm/h reference rainfall threshold
to about 3 mm/h at 2-mm/h threshold, which is, in relative
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
1678 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
Fig. 10. (Upper panels) Probability of rain detection and (lower panels) fraction of missed rainfall values to the overall rainfall volume plotted as a function
of the reference (Mesonet) rainfall for the different sensor estimates [WSR-88D, CMORPH, and 3B42 (V.6)]. Solid and dashed lines represent 2004 and 2006
statistics, respectively. Left and right panels correspond to cold and warm season months.
terms, equal to 85% of the mean of reference rainfall. The
corresponding rmse values for CMORPH are higher than that
for 3B42 (V.6) at all thresholds, which are up to 50% and
100% for 2004 and 2006, respectively. The two gauge-adjusted
estimates (WSR-88D and 3B42, V6) exhibit almost identical
behavior in terms of the rmse error statistic of 2004 and 2006.
The CMORPH product, on the other hand, that does not
include gage data shows a significant change (increase) in the
rmse values from 2004 to 2006 (in both seasons). In the warm
season particularly, the increase in the rmse reached 40%,
while in the cold season, the increase was up to 30%.
In Fig. 13, we show the spatial correlation of rainfall error
maps for the different sensor products. As expected, the spatial
correlation of the retrieval error decreases with increasing space
lag, and it is lower during the warm season, mainly character-
ized by spatially variable convective systems. The radar product
shows t he lowest spatial correlation of the error during the
warm period, with values smaller than 0.1 beyond an 80-km
spatial lag, whereas during the cold period, its behavior is very
similar to 3B42 (V6), with values of about 0.5 even beyond a
100-km spatial lag. No significant difference in the spatial cor-
relation of radar rainfall error is noticed between the two years.
The two satellite products exhibit higher spatial correlations
than the radar product during the warm season, particularly in
2006. In the cold season, the radar and 3B42 (V6) have al-
most identical spatial correlation patterns, while the CMORPH
product exhibits lower correlations. Overall, among the two
satellite products, CMORPH correlation patterns perform the
following: 1) They exhibit the greatest disagreement with the
radar (i.e., significant overestimation in the warm season and
underestimation in the cold season) and 2) do not carry a
strong signature of seasonal dependence (as in the case of
3B42 V6).
C. Representativeness of the Reference Data set
As discussed in Section III-A, the representativeness of
spatially interpolated Mesonet station measurements to the
grid-cell average rainfall is affected by the subgrid variability
of rainfall. The Mesonet-to-Micronet grid-cell average rainfall
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
ANAGNOSTOU et al.: BENCHMARKING HIGH-RESOLUTION GLOBAL SATELLITE RAINFALL PRODUCTS 1679
Fig. 11. As in Fig. 10 for (upper panels) FARs and (lower panels) fraction of rain volume of the falsely detected rain values for the different sensor estimates.
error statistics presented in Section III-A provide a qualitative
benchmarking to characterize the validity of Mesonet as a ref-
erence data set for the various remote sensing products. Com-
paring the error statistics of the three remote sensing products
(shown in Figs. 10–12) to the Mesonet error statistics (shown
in Fig. 4), one can argue that the grid-cell average rainfall
derived from Mesonet stations is superior to any remote sens-
ing product. Specifically, Mesonet (in reference to Micronet)
yields low FARs and high probability of rain detection. The
POD values are greater than 0.8 (0.9) for a rain threshold of
0.1 mm/h in warm (cold) season months. The FAR is below
0.005 for the same rainfall threshold (0.1 mm/h). Moreover,
the Mesonet estimates were shown to be nearly unbiased, while
the correlation with Micronet data was above 0.93. Finally, the
rmse of Mesonet is significantly lower than the rmse shown for
the different remote sensing products at the same space–time
resolution. Specifically, the radar product rmse is nearly 100%
(150%) higher than the rmse of the Mesonet for the warm
(cold) season months. Although this qualitative comparison
cannot be used to derive the actual radar product error variance,
which requires elaborate statistical approaches on the error
variance separation (see [7]), it does provide a qualitative basis
to justify the use of Mesonet-derived grid-cell average rainfall
as a reference for evaluating remote sensing rainfall products
(particularly those from satellite sensors).
IV. C
ONCLUSION
The error statistics of two high-resolution global-scale
satellite products were contrasted against rain-gauge-adjusted
ground-based radar rainfall estimates. Rain estimation error
in this paper was assessed at two levels, starting from the
reference rainfall fields derived from interpolating Mesonet
network rainfall measurements and moving up to the distributed
rainfall fields derived from radar and satellite techniques. The
evaluation of the representativeness of the “reference” data
source approached in this paper provides the most rigorous
benchmarking of satellite rainfall products currently not avail-
able in literature. We showed that 3-hourly and 25-km grid
space resolution rainfall fields derived from Mesonet stations
have high rain detection probability (> 0.8), high correlation
(> 0.9), and standard error below 30% at both warm and cold
seasons. These error statistics, although significantly lower (be-
tween 20% and 50% in terms of rmse) than the corresponding
error statistics of the satellite and radar products, demonstrate
the need to benchmark reference data sources prior to their
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
1680 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
Fig. 12. As in Fig. 10 for (upper panels) bias ratio, (middle panels) correlation, and (lower panels) rmse statistics.
quantitative use in validating remote sensing retrievals. In terms
of the satellite rainfall products, our main conclusions are the
following.
1) In the warm season, the version of CMORPH evaluated
in this paper exhibits significant overestimation of rain-
fall (> 50%), which also results in high rmse statistic.
The corresponding bias in the 3B42 (V6) product is
below 50% in the warm season. In the cold season, the
bias of both satellite estimates is low (below 20%). A
point to note is that the 3B42 (V6) product is adjusted
to gauge data at the monthly time scale, which may
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
ANAGNOSTOU et al.: BENCHMARKING HIGH-RESOLUTION GLOBAL SATELLITE RAINFALL PRODUCTS 1681
Fig. 13. Spatial correlation of rainfall error between sensor estimates (radar and satellite products) and Mesonet rainfall fields.
have contributed to the reduction of bias in the warm
season.
2) The probability of rain detection of CMORPH is high
relative to the 3B42 (V6), accompanied, though, with
higher FARs, which overall give a balanced difference
between lack of detected versus falsely detected rainfall
volumes.
3) There is distinct improvement in the 3B42 (V6) POD
from 2004 to 2006 and in the warm season FAR of
CMORPH (the cold season FAR worsens).
4) There is distinct worsening of the CMORPH rmse from
2004 to 2006, which is mainly attributed to the increased
positive bias.
5) In terms of correlations, CMORPH is higher than 3B42
(V6), but both rainfall products are significantly lower
than the radar rainfall estimates (WSR-88D) in both
warm and cold seasons. Overall, the radar rainfall esti-
mates have the lowest rmse, biases, and FARs; in terms
of POD, the radar rainfall estimates are similar to the
CMORPH.
6) In terms of the rainfall error structure, the CMORPH
product exhibits the highest spatial correlation in the
warm season and the lowest in the cold season. Radar
has significantly lower spatial error correlation in the
warm season than the two satellite products, indicat-
ing fast decorrelation of error, while in the cold sea-
son, radar and 3B42 (V6) products have similar error
correlations.
This paper has focused on two of the currently available high-
resolution global-scale satellite rainfall products but provided
detailed benchmarking in terms of their quantitative error sta-
tistics. A point to note is that the study in this paper was carried
out over a continental midlatitudinal region. Hence, the results
may not be applicable to other regions, particularly the tropics
and mountainous regions where the precipitation processes and
cloud types are generally different. Future continuation should
include those regions as well as benchmarking of additional
high-resolution satellite products, such as the PERSIANN and
GSMaP, discussed in the introduction session. Another impor-
tant aspect to consider is the fact that there are two primary
sources of error in satellite blending techniques: errors in the in-
stantaneous rain rate estimates derived from microwave sensors
and the interpolation between these snapshots using VIS/IR
information. It would be useful to distinguish and quantify the
relative effects to determine whether the emphasis in correcting
the problem should be on the former or the latter error source.
Finally, satellite rainfall error studies should also focus on the
error propagation in the prediction of land surface hydrolog-
ical variables such as runoff, soil moisture, and heat/water
fluxes.
R
EFERENCES
[1] A. Ali, T. Lebel, and A. Amani, “Rainfall estimation in the Sahel.
Part I: Error function,” J. Appl. Meteorol., vol. 44, no. 11, pp. 1691–1706,
Nov. 2005.
[2] E. N. Anagnostou, W. F. Krajewski, and J. Smith, “Uncertainty quantifi-
cation of mean-field radar-rainfall estimates,” J. Atmos. Ocean. Technol.,
vol. 16, no. 2, pp. 206–215, Feb. 1999.
[3] P. A. Arkin, R. Joyce, and J. E. Janowiak, “The estimation of
global monthly mean rainfall using infrared satellite data: The GOES
Precipitation Index (GPI),” Remote Sens. Rev., vol. 11, pp. 107–124,
1994.
[4] P. A. Arkin, J. Turk, and E. Ebert, “Pilot evaluation of high resolution
precipitation products (PEHRPP): A contribution to GPM planning,” in
Proc. 5th GPM Plan. Workshop, Tokyo, Japan, 2006.
[5] M. Borga, E. N. Anagnostou, and E. Frank, “On the use of real-time radar
rainfall estimates for flood prediction in mountainous basins,” J. Geophys.
Res., vol. 105, no. D2, pp. 2269–2280, 2000.
[6] F. V. Brock, K. C. Crawford, R. L. Elliott, G. W. Cuperus,
S. J. Stadler, H. L. Johnson, and M. D. Eilts, “The Oklahoma Mesonet: A
technical overview,” J. Atmos. Ocean. Technol., vol. 12, no. 1, pp. 5–19,
Feb. 1995.
[7] G. Ciach and W. F. Krajewski, “On the estimation of radar rain-
fall error variance,” Adv. Water Resour., vol. 22, no. 6, pp. 585–595,
Feb. 1999.
[8] M. D. Conner and G. R. Petty, “Validation and intercomparison of SSM/I
rain-rates retrieval methods over the continental Unites States,” J. Appl.
Meteorol., vol. 37, no. 7, pp. 679–700, Jul. 1998.
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
1682 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 4, APRIL 2010
[9] T. Dinku and E. N. Anagnostou, “Regional differences in overland rain-
fall estimation from PR-calibrated TMI algorithm,” J. Appl. Meteorol.,
vol. 44, no. 2, pp. 189–205, Feb. 2005.
[10] E. E. Ebert, J. Janowiak, and C. Kidd, “Comparison of near-real-
time precipitation estimates from satellite observations and numeri-
cal models,” Bull. Amer. Meteorol. Soc., vol. 88, no. 1, pp. 47–64,
Jan. 2007.
[11] R. R. Ferraro and G. F. Marks, “The development of SSM/I rain-rate
retrieval algorithms using ground-based radar measurements,” J. Atmos.
Ocean. Technol., vol. 12, no. 4, pp. 755–770, Aug. 1995.
[12] R. A. Fulton, J. P. Breidenbach, D. J. Seo, D. A. Miller, and T. O’Bannon,
“The WSR-88D rainfall algorithm,” Weather and Forecasting, vol. 13,
no. 2, pp. 377–395, 1998.
[13] M. Gebremichael and W. F. Krajewski, “Characterization of the tempo-
ral sampling error in space-time-averaged rainfall estimates from satel-
lites,” J. Geophys. Res., vol. 109, no. D11, p. D11110, Jun. 2004,
DOI: 10.1029/2004JD004509.
[14] M. Gebremichael and W. F. Krajewski, “The effect of temporal sampling
error on inferred rainfall spatial statistics,” J. Appl. Meteorol., vol. 44,
no. 10, pp. 1626–1633, Oct. 2005.
[15] J. Gottschalck, J. Meng, M. Rodell, and P. Houser, Analysis of mul-
tiple precipitation products and preliminary assessment of their im-
pact on global land data assimilation system land surface states,”
J. Hydrometeorol., vol. 6, no. 5, pp. 573–598, Oct. 2005.
[16] M. Grecu and E. N. Anagnostou, “Overland precipitation estimation
from passive microwave observations,” J. Appl. Meteorol., vol. 40,
pp. 1367–1380, 2001.
[17] C. G. Griffith, W. L. Woodley, and P. G. Grube, “Rain estimation
from geosynchronous satellite imagery—Visible and infrared studies,”
Mon. Weather Rev., vol. 106, no. 8, pp. 1153–1171, 1978.
[18] Y. Hong, K.-L. Hsu, S. Sorooshian, and X. Gao, “Self-organizing nonlin-
ear output (SONO): A neural network suitable for cloud patch–based rain-
fall estimation from satellite imagery at small scales,” Water Resour. Res.,
vol. 41, no. 3, p. W03008, Mar. 2005, DOI:10.1029/2004WR003142.
[19] Y. Hong, K.-L. Hsu, H. Moradkhani, and S. Sorooshian, “Uncertainty
quantification of satellite precipitation estimation and Monte Carlo assess-
ment of the error propagation into hydrologic response,” Water Resour.
Res., vol. 42, p. W08421, Aug. 2006, DOI:10.1029/2005WR004398.
[20] F. Hossain and E. N. Anagnostou, Assessment of current passive-
microwave- and infrared-based satellite rainfall remote sensing for
flood prediction,” J. Geophys. Res., vol. 109, no. D7, p. D07102,
Apr. 2004.
[21] F. Hossain and G. J. Huffman, “Investigating error metrics for satellite
rainfall at hydrologically relevant scales,” J. Hydrometeorol., vol. 9, no. 3,
pp. 563–575, Jun. 2008.
[22] F. Hossain and D. Lettenmaier, “Flood prediction in the future: Recogniz-
ing hydrologic issues in anticipation of the global precipitation measure-
ment mission—Opinion paper,” Water Resour. Res., vol. 42, p. W11301,
2006, DOI:10.1029/2006WR005202.
[23] G. J. Huffman, R. Adler, D. Bolvin, and E. Nelkin, “Uncertainty in fine-
scale MPA precipitation estimates and implications for hydrometeorolog-
ical analysis and forecasting,” in Proc. 18th Conf. Hydrol., Seattle, WA,
Jan. 11–18, 2004.
[24]G.J.Huffman,R.F.Adler,D.T.Bolvin,G.Gu,E.J.Nelkin,
K. P. Bowman, Y. Hong, E. F. Stocker, and D. B. Wolff, “The TRMM
multi-satellite precipitation analysis: Quasi-global, multi-year, combined-
sensor precipitation estimates at fine scale,” J. Hydrometeorol.,vol.8,
no. 1, pp. 38–55, Feb. 2007.
[25] R. J. Joyce, J. E. Janowiak, P. A. Arkin, and P. Xie, “CMORPH: A method
that produces global precipitation estimates from passive microwave and
infrared data at high spatial and temporal resolution,” J. Hydrometeorol.,
vol. 5, no. 3, pp. 487–503, Jun. 2004.
[26] C. Kidd, D. R. Kniveton, M. C. Todd, and T. J. Bellerby, “Satellite
rainfall estimation using combined passive microwave and infrared
algorithms,” J. Hydrometeorol., vol. 4, no. 6, pp. 1088–1104,
Dec. 2003.
[27] W. F. Krajewski, M. C. Anderson, W. E. Eichinger, D. Entekhabi,
B. K. Hornbuckle, P. R. Houser, G. G. Katul, W. P. Kustas, J. M. Norman,
C. Peters-Lidard, and E. F. Wood, A remote sensing observatory
for hydrologic sciences: A genesis for scaling to continental hydrol-
ogy,” Water Resour. Res., vol. 42, no. 7, p. W07301, Jul. 2006,
10.1029/2005WR004435.
[28] T. Kubota, K. Okamoto, S. Shige, T. Ushio, T. Iguchi, N. Takahashi,
K. Iwanami, K. Aonashi, M. Kachi, and R. Oki, “The Global Satel-
lite Mapping of Precipitation (GSMaP) Project,” in Proc. 7th GPM
Int. Plan. Workshop, Tokyo, Japan, Dec. 7, 2007. [Online]. Available:
http://www.eorc.jaxa.jp/GPM/ws7/pdf/7thDEC2007/7_am07.pdf
[29] C. D. Kummerow and L. Giglio, A passive microwave technique
for estimating rainfall and vertical structure information from space.
Part I: algorithm description,” J. Appl. Meteorol., vol. 33, no. 1, pp. 3–18,
Jan. 1994.
[30] C. D. Kummerow and L. Giglio, A passive microwave technique for
estimating rainfall and vertical structure information from space. Part II:
Applications to SSM/I data,” J. Appl. Meteorol., vol. 33, no. 1, pp. 19–34,
Jan. 1994.
[31] C. D. Kummerow, Y. Hong, W. S. Olson, S. Yang, R. F. Adler,
J. McCollum, R. Ferraro, G. Petty, D.-B. Shin, and T. T. Wilheit, “The
evolution of the goddard profiling algorithm (GPROF) for rainfall estima-
tion from passive microwave sensors,” J. Appl. Meteorol., vol. 40, no. 11,
pp. 1801–1820, Nov. 2001.
[32] K. H. Lee and E. N. Anagnostou, “Investigation of the nonlinear
hydrologic response to precipitation forcing in physically based land
surface modeling,” Can. J. Remote Sens., vol. 30, no. 5, pp. 706–716,
2004.
[33] F. S. Marzano, E. Picciotti, and G. Vulpiani, “Rain field and reflectiv-
ity vertical profile reconstruction from C-band radar volumetric data,”
IEEE Trans. Geosci. Remote Sens., vol. 42, no. 4, pp. 1033–1046,
May 2004.
[34] J. R. McCollum, W. F. Krajewski, R. R. Ferraro, and M. B. Ba, “Evalu-
ation of biases of satellite rainfall estimation algorithms over the conti-
nental United States,” J. Appl. Meteorol., vol. 41, no. 11, pp. 1065–1080,
Nov. 2002.
[35] A. Mugnai, E. A. Smith, and G. J. Tripoli, “Foundations for
statistical–physical precipitation retrieval from passive microwave
satellite measurements. Part II: Emission-source and general-
ized weighting-function properties of a time-dependent cloud-
radiation model,” J. Appl. Meteorol., vol. 32, no. 1, pp. 17–39,
Jan. 1993.
[36] B. Nijssen and D. P. Lettenmaier, “Effect of precipitation sampling error
on simulated hydrological fluxes and states: Anticipating the global pre-
cipitation measurement satellites,” J. Geophys. Res., vol. 109, p. D02103,
Jan. 2004.
[37] D. Rosenfeld, “Cloud top microphysics as a tool for precipitation
measurements,” in Measuring Precipitation From Space, EURAINSAT
and the Future, advances in Global Change Research, vol. 28,
V. Levizzani and P. Bauer, Eds. Berlin, Germany: Springer-Verlag, 2007,
pp. 61–78.
[38] J. Simpson, C. Kummerow, W. K. Tao, and R. F. Adler, “On the Tropical
Rainfall Measuring Mission (TRMM),” Meteorol. Atmos. Phys., vol. 60,
pp. 19–36, 1996.
[39] E. A. Smith, A. Mugnai, H. J. Cooper, G. J. Tripoli, and X. Xiang,
“Foundations for statistical-physical precipitation retrieval from passive
microwave satellite measurements. Part I: Brightness-temperature prop-
erties of a time-dependent cloud-radiation model,” J. Appl. Meteorol.,
vol. 31, no. 6, pp. 506–531, Jun. 1992.
[40] E. A. Smith, G. Asrar, Y. Furuhama, A. Ginati, A. Mugnai, K. Nakamura,
R. F. Adler, M.-D. Chou, M. Desbois, J. F. Durning, J. K. Entin,
F. Einaudi, R. R. Ferraro, R. Guzzi, P. R. Houser, P. H. Hwang, T. Iguchi,
P. Joe, R. Kakar, J. A. Kaye, M. Kojima, C. Kummerow, K.-S. Kuo,
D. P. Lettenmaier, V. Levizzani, N. Lu, A. V. Mehta, C. Morales, P. Morel,
T. Nakazawa, S. P. Neeck, K. Okamoto, R. Oki, G. Raju, J. M. Shepherd,
J. Simpson, B.-J. Sohn, E. F. Stocker, W.-K. Tao, J. Testud, G. J. Tripoli,
E. F. Wood, S. Yang, and W. Zhang, “International global precipitation
measurement (GPM) program and mission: An overview,” in Measuring
Precipitation From Space—EURAINSAT and the Future, V. Levizzani,
P. Bauer, and F. J. Turk, Eds. Berlin, Germany: Springer-Verlag, 2007,
pp. 611–653.
[41] S. Sorooshian, K. L. Hsu, X. Gao, H. V. Gupta, B. Imam, and
D. Braithwaite, “Evaluation of PERSIANN system satellite–based esti-
mates of tropical rainfall,” Bull. Amer. Meteorol. Soc., vol. 81, no. 9,
pp. 2035–2046, Sep. 2000.
[42] F. J. Tapiador, C. Kidd, V. Levizzani, and F. S. Marzano, A
neural networks-based fusion technique to estimate half-hourly rain-
fall estimates at 0.1
resolution from satellite passive microwave
and infrared data,” J. Appl. Meteorol., vol. 43, no. 4, pp. 576–594,
Apr. 2004.
[43] Y. Tian, C. D. Peters-Lidard, B. J. Choudhury, and M. Garcia, “Multi-
temporal analysis of TRMM-based satellite precipitation products for
land data assimilation applications,” J. Hydrometeorol., vol. 8, no. 6,
pp. 1165–1183, 2007.
[44] F. J. Turk and S. D. Miller, “Toward improved characterization of remotely
sensed precipitation regimes with MODIS/AMSR-E blended data tech-
niques,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 5, pp. 1059–1069,
May 2005.
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
ANAGNOSTOU et al.: BENCHMARKING HIGH-RESOLUTION GLOBAL SATELLITE RAINFALL PRODUCTS 1683
Emmanouil N. Anagnostou received the B.S. de-
gree in civil and environmental engineering from
the National Technical University of Athens, Athens,
Greece, and the M.S. and Ph.D. degrees in civil and
environmental engineering from The University of
Iowa, Iowa City.
He is currently a Professor with the Department
of Civil and Environmental Engineering, University
of Connecticut, Storrs, and a Team Leader of a
Marie Curie Excellence Grant with the Institute of
Inland Waters, Hellenic Center for Marine Research,
Anavissos, Greece. His primary research focuses on developing techniques for
remote sensing of precipitation parameters, including precipitation profiles and
surface rainfall from satellite, ground-based (weather radars), and underwater
(acoustic) sensors. His other research interests include the integration of rainfall
remote sensing products in hydrologic modeling systems for the prediction
of floods and studying regional water and energy cycle. He is the author or
coauthor of more than 90 journal papers and several book chapters in the areas
of precipitation remote sensing and hydrometeorological applications.
Dr. Anagnostou is a member of several international organizations and
scientific committees.
Viviana Maggioni received the B.S. and M.S. de-
grees in environmental engineering from the Politec-
nico di Milano, Milano, Italy, in 2003 and 2006,
respectively. She is currently working toward the
Ph.D. degree in the Department of Civil and Envi-
ronmental Engineering, University of Connecticut,
Storrs.
She studies error propagation from the precipita-
tion measurement to land surface simulations of soil
moisture and other parameters and investigates how
improved characterization of precipitation-modeling
uncertainty impacts a land data assimilation system. Her research interest
focuses on the hydrologic applicability of satellite rainfall observations in land
data assimilation systems.
Efthymios I. Nikolopoulos received the B.Eng. de-
gree in environmental engineering from the Tech-
nical University of Crete, Crete, Greece, and the
M.Sc. degree in environmental engineering from
The University of Iowa, Iowa City. He is currently
working toward the Ph.D. degree at the University of
Connecticut, Storrs.
He is also an early stage Researcher Member of
a Marie Curie Excellence team with the Institute of
Inland Waters, Hellenic Center for Marine Research,
Anavissos, Greece. His research interests include
remote sensing of precipitation, flash flood hydrology, and error propagation
of satellite rainfall through hydrologic models. His research is focused on
the understanding of runoff generation mechanism during flash floods and
the evaluation of the use of satellite rainfall observations to predict floods in
complex terrain basins.
Tadesse Meskele received the B.S. degree in civil
engineering from Arba Minch University, Arba
Minch, Ethiopia, and the M.S. degree in water re-
sources engineering from Katholieke Universiteit
Leuven, Leuven, Belgium, and Vrije Universiteit
Brussel, Brussels, Belgium. He is currently working
toward the Ph.D. degree at Portland State Univer-
sity, Portland, OR, working on the satellite rainfall
retrieval error translation to hydrological responses,
and developing and implementing various data as-
similation techniques for stream flow and soil mois-
ture assimilation in hydrologic models.
Faisal Hossain received the B.S. degree from the In-
dian Institute of Technology, Varanasi, India, in 1996
and the Ph.D. degree in environmental engineering
from the University of Connecticut, Storrs, in 2004.
He is currently an Associate Professor in civil
engineering with the Tennessee Technological Uni-
versity, Cookeville. He has over 50 peer-reviewed
publications in the fields of groundwater contam-
ination mapping, flood prediction, satellite precip-
itation, transboundary water resources issues, and
engineering education. He has been the Associate
Editor of the Journal of American Association of Water Association since 2006.
Dr. Hossain is the recipient of awards and recognitions such as NASA New
Investigator Program, American Society of Engineering Education Outstanding
New Faculty Research Award, and Top Performer Rating by NSF Alan T.
Waterman Award Committee.
Anastasios Papadopoulos received the B.Sc. degree in physics from the
Aristotle University of Thessaloniki, Thessaloniki, Greece, and the M.Sc.
and Ph.D. degrees in meteorology from the National Technical University of
Athens, Athens, Greece.
He is currently a Physicist–Meteorologist and Associate Researcher with
the I nstitute of Inland Waters, Hellenic Center for Marine Research, Anavis-
sos, Greece. His research interests and experience include the following:
regional–mesoscale numerical weather prediction; operational meteorology
and oceanography (design, development, and evaluation); air–land–sea in-
teraction and processes; extreme weather events and severe storms (theory
and modeling); soil erosion and mechanisms of production, transport, and
deposition of dust substance; and atmospheric stability and dispersion of air
pollution. He is the author or coauthor of more than 20 publications in peer-
reviewed journals, 70 peer-reviewed publications in international and national
conferences, and more than 90 other publications in the fields of meteorology,
mesoscale modeling, study of severe weather events, and air pollution.
Authorized licensed use limited to: Princeton University. Downloaded on June 07,2010 at 16:05:03 UTC from IEEE Xplore. Restrictions apply.
... As a solution for the abovementioned limitations and shortcomings identified in rain gauge data, multisatellite high-resolution precipitation products such as tropical rainfall measuring mission (TRMM) multisatellite precipitation analysis (TMPA), precipitation estimation from remotely sensed information using artificial neutral network (PERSIANN) system, multisatellite precipitation analysis, and multisatellite rainfall estimate with climate prediction center (CPM) and morphing technique (CMORPH) and weather radar observations are widely in use around the world [9][10][11]. ...
... However, these SbPPs are only used after a careful investigation in the desired study area. Since it was discovered that these SbPPs have certain uncertainties, such as accuracy that is affected by topographical features of the study area and precipitation mechanism due to seasonal and regional climate conditions, such accuracy evaluations of the products with respect to rain gauge data are done for each area of concern [7,9,10], which cannot be ignored if we plan to use them in hydrological applications [6]. A study done by et al., [13] in the Ganzi river basin of the Tibetan plateau to evaluate the impact of satellite data sets to be used in hydrological modeling for that area used CMORPH-CRT, PERSIANN-CDR, 3B42RT, and 3B42 satellite data sets against observed rainfall data using HEC-HMS model to find out that TRMM-3B42RT and CMORPH-CRT show good performance in the respective area and they also suggested that TRMM-3B42RT is a better choice overall for hydrological models in the Ganzi river basin of Tibetan plateau. ...
Article
Full-text available
Satellite-based Precipitation Products, (SbPPs) have piqued the interest of a number of researchers as a reliable replacement for observed rainfall data which often have limited time spans and missing days. The SbPPs possess certain uncertainties, thus, they cannot be directly used without comparing against observed rainfall data prior to use. The Kelani River Basin is Sri Lanka's fourth longest river and the main source of water for almost 5 million people. Therefore, this research study aims to identify the potential of using SbPPs as a different method to measure rain besides using a rain gauge. Furthermore, the aim of the work is to examine the trends in precipitation products in the Kelani River basin. Three SbPPs, Precipitation Estimation using Remotely Sensed Information using Artificial Neural Networks (PERSIANN), PERSIANN-Cloud Classification System (CCS), and PERSIANN-Climate Data Record (CDR) and ground observed rain gauge daily rainfall data at nine locations were used for the analysis. Four continuous evaluation indices namely, Root Mean Square Error (RMSE), (Percent Bias) Pbias, Correlation Coefficient (CC), and Nash-Shutcliffe Efficiency (NSE) were used to determine the accuracy by comparing against observed rainfall data. Four categorical indices including Probability of Detection (POD), False Alarm Ratio (FAR), Critical Success Index (CSI), and Proportional Constant (PC) were used to evaluate the rainfall detection capability of SbPPs. Mann-Kendall Test and Sen’s Slope Estimator were used to identify whether a trend was present while the magnitudes of these were calculated by Sen’s Slope. PERSIANN-CDR performed well by showing better performance in both POD and CSI. When compared to observed rainfall data, the PERSIANN product had the lowest RMSE value, while all products indicated underestimations. The CC and NSE of all three products with observed rainfall data were also low. Mixed results were obtained for the trend analysis as well. The overall results showed that all three products are not a better choice for the chosen study area.
... Besides, these stations are heterogeneously distributed around the earth (Kidd et al., 2016) and represent the rainfall in a limited area around each station (Kidd and Huffman, 2011). Therefore, rainfall monitoring through alternatives, including rainfall radars and satellite sensors, has been well-established (Anagnostou et al., 2010). ...
... However, large uncertainties might influence rain radars due to beam blockage and frozen hydrometeor (Villarini and Krajewski, 2010). In contrast, satellite rainfall products have been available since the 1970 s, and their accuracy and resolution (both temporal and spatial) have improved over time (Anagnostou et al., 2010;Zhang et al., 2020). Satellite products can measure various precipitation types such as snow, rainfall, and hail. ...
Article
Sparse distribution of rain gauge networks challenges the estimation of rainfall variability over space and time. The SM2RAIN algorithm was developed to estimate rainfall from the knowledge of soil moisture (SM) by inverting the soil‐water balance equation. The algorithm was simplified by neglecting the contribution of evapotranspiration and surface runoff rate during the rainfall event. A recent study developed an analytical model to estimate the net water flux (NWF) from SM data via inversion of analytical Warrick’s equation. In this study, the SM2RAIN-NWF algorithm was developed by integrating the SM2RAIN algorithm and the NWF model to improve the accuracy of rainfall estimation. The applicability of the SM2RAIN-NWF algorithm was evaluated based on observed rainfall data in the Lake Urmia basin, Iran. Satellite SM data was obtained from the Advanced Microwave Scanning Radiometer 2 (AMSR2). The algorithm calibrated based on the data from July 3, 2012, to December 31, 2017, was then used to estimate rainfall for two years extending from January 2018 to December 2019. Estimated rainfall through SM2RAIN-NWF algorithm improved compared to SM2RAIN by 14% and 37.4% increase in the average values of correlation coefficient (R) and Nash–Sutcliffe (NS), and 11.5% decrease in the Percentage Root Mean Square Error (PRMSE) over the calibration period. Validating the estimated rainfall showed a considerable improvement in the performance of the SM2RAIN-NWF algorithm compared to the SM2RAIN algorithm by 8.6% and 30.4% increase in the average values of R and NS, and 13.4% decrease in the PRMSE. It was also found that the SM2RAIN-NWF algorithm contributes to the improvement of error indices in rainfall estimation and simulates the rainfall variation trend in a better fashion than the SM2RAIN algorithm.
... Since satellite QPEs contain data information from multiple sensors, such as the Microwave Imager (TMI) on TRMM, Special Sensor Microwave Imager (SSM/I) on Defense Meteorological Satellite Program (DMSP) satellites, and Advanced Microwave Scanning Radiometer-Earth Observing System (AMSR-E) on Aqua, they are subject to biases and uncertainties in estimating regional PR including PR extremes (Anagnostou et al., 2010;Liu and Zipser, 2015;Maggioni et al., 2014). A comprehensive assessment in the ability of QPEs to estimate the intensity, frequency and spatial distributions of PR extremes at different spatio-temporal scales is essential for the accurate application of QPEs in the monitoring and forecasting of extreme PR events (Tan et al., 2018;Trenberth et al., 2017). ...
Article
The wide and consistent global coverage of satellite-based quantitative precipitation estimates (QPEs) has shown great potential for monitoring precipitation (PR) at large spatial scales. Evaluation of QPEs in estimating PR extremes is vital to forecasting hydrologic extremes. Here, we present a systematic evaluation of four commonly used QPEs: (1) the Climate Hazards Group InfraRed Precipitation with Stations (CHIRPS), (2) Tropical Rainfall Measuring Mission 3B42 Version 7 (TRMM 3B42 V7), (3) Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Climate Data Record (PERSIANN-CDR), and (4) Multi-Source Weighted-Ensemble Precipitation (MSWEP), in their abilities to detect PR extremes in mainland China at annual-seasonal-monthly scales. Four Standard Extreme Precipitation Indices (SEPIs) are chosen as the assessment metrics, including the maximum 1-day PR (Rx1day), the simple PR intensity index (SDII), the count of days with PR ≥ 20 mm (R20mm), and the consecutive dry days (CDD). Results indicate that PERSIANN-CDR performs best for all SEPIs, followed by CHIRPS and TRMM 3B42 V7. All QPEs (except MSWEP) perform better (worse) in capturing CDD (SDII) than other SEPIs at all three timescales. However, large differences among the performance of QPEs in estimated seasonal SEPIs are found - CHIRPS (PERSIANN-CDR) outperforms the other QPEs in spring (summer and autumn), while PERSIANN-CDR and CHIRPS outperform the other QPEs in winter. All QPEs perform better in detecting extreme PR occurrence in summer than in other seasons, and spatially most of them perform better in humid southeastern China than in arid northwestern China. CHIRPS and TRMM 3B42 V7 overestimate Rx1day, R20mm and SDII in most of China, while MSWEP notably underestimates CDD in most of China and overestimates Rx1day, R20mm and SDII in western China.
... However, rainfall gauging station networks are often unevenly distributed sparsely across space, which imposes difficulties for properly capturing the spatial variability of precipitation systems [2]. In addition, precipitation samples obtained from ground-based gauges are frequently corrupted by long periods of missing data, which may hinder their use for continuous rainfall-runoff modeling and, accordingly, for the indirect estimation of streamflow-related variables [3]. ...
Article
Full-text available
Precipitation products derived from satellites have emerged as a promising approach for obtaining precipitation estimates, enabling accurate long-term observations and describing the water cycle dynamics from a global scale to a local scale. The quality of these products has improved significantly in the last decades, especially with the emergence of TRMM missions and its successor GPM. The objective of this study was to evaluate the daily, monthly and annual precipitation estimates provided by IMERG version 05 of the GPM, with the data observed by the rainfall stations of the Brazilian Agency of Water and Sanitation (ANA) in the basins of the Brazilian midwest. In order to compare the data, the spatialization of the data of the rainfall stations was performed by means of the ordinary kriging technique, interpolating the data for grids of 0.1° × 0.1° that correspond to the specialized grids of the GPM satellite. The data were evaluated quantitatively by means of statistical metrics. The GPM satellite precipitation product performed relatively well on a daily scale for regions with smooth topography, and was able to describe the rainfall regime on larger time scales, regardless of the terrain conditions. However, the satellite retrievals were unable to reproduce rainfall extremes in virtually all situations, which may limit their application in frequency analyses.
... This difference in model performance between the two scenarios may be attributed to the difference in the value of satellite precipitation data with respect to the precipitation gauge data (Stisen and Sandholt, 2010). Past studies have shown that even rain gauge data are not free of error and contain uncertainties similar to SPP data (Ali et al., 2005;Anagnostou et al., 2010). Hence, such a difference may also be due to the poor quality, lack of spatial coverage and missing data, particularly for regions with sparse rain gauges. ...
Article
This work summarizes lessons learnt on using satellite precipitation products (SPPs) for flood simulation and prediction and proposes ways forward in this field of research. A meta-analysis was carried out to review: effect of climate zone, topographical features, selection of hydrological models, and calibration procedures on SPPs forced hydrological model performance. SPPs performance was shown to be higher in temperate and tropical than in dry climates. Low lying and moderate elevations areas exhibited high-performance accuracy compared to higher latitudes landscapes. SPPs that use microwave algorithms were found to outperform the others. The best simulation and prediction results were found after bias correction and model recalibration. From a general standpoint, SPPs offer great potential for flood simulation and prediction, but the performance of SPPs needs to be enhanced for operational purposes. The present study discusses bias correction techniques, model recalibration, the importance of interaction between different types of SPPs and hydrological models, and other lessons learned and future directions of using SPPs for future flood applications.
... Many different metrics to assess predictive skill can be defined, even when considering only two classes (Agha-Kouchak & Mehran, 2013;Wilks, 2006), however, the probabilities of a hit (known as probability of detection; POD) and of a false alarm (known as the false alarm ratio; FAR) are most commonly used in the literature (Anagnostou et al., 2010;Gourley et al., 2012;Hao et al., 2013). Here we used the critical success index (CSI), which combines the latter two metrics, as follows (Schaefer, 1990): ...
Article
Full-text available
Precipitation prediction at seasonal timescales is important for planning and management of water resources as well as preparedness for hazards such as floods, droughts and wildfires. Quantifying predictability is quite challenging as a consequence of a large number of potential drivers, varying antecedent conditions, and small sample size of high‐quality observations available at seasonal timescales, that in turn, increases prediction uncertainty and the risk of model overfitting. Here, we introduce a generalized probabilistic framework to account for these issues and assess predictability under uncertainty. We focus on prediction of winter (Nov–Mar) precipitation across the contiguous United States, using sea surface temperature‐derived indices (averaged in Aug–Oct) as predictors. In our analysis we identify “predictability hotspots,” which we define as regions where precipitation is inherently more predictable. Our framework estimates the entire predictive distribution of precipitation using copulas and quantifies prediction uncertainties, while employing principal component analysis for dimensionality reduction and a cross validation technique to avoid overfitting. We also evaluate how predictability changes across different quantiles of the precipitation distribution (dry, normal, wet amounts) using a multi‐category 3 × 3 contingency table. Our results indicate that well‐defined predictability hotspots occur in the Southwest and Southeast. Moreover, extreme dry and wet conditions are shown to be relatively more predictable compared to normal conditions. Our study may help with water resources management in several subregions of the United States and can be used to assess the fidelity of earth system models in successfully representing teleconnections and predictability.
Article
Study Region The Upper Taoer River Watershed (UTRW), Northeastern China Study Focus This study presents a comprehensive hydrometeorological evaluation of six satellite precipitation products (SPPs) over a sparsely gauged semi-arid watershed, including the adjusted and unadjusted versions of Tropical Rainfall Measuring Mission Multi-satellite Precipitation Analysis (TMPA), Global Precipitation Measurement (GPM) Integrated Multi-satellite Retrievals for GPM products (IMERG), and Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERISIANN) series. The effectiveness of monthly bias correction in improving SPPs estimation accuracy and in enhancing their feasibility in hydrological applications was also analyzed. New hydrological insights for the region The SPPs, especially bias-corrected SPPs, adequately provide complementary precipitation information for UTRW where ground measurement is insufficient. The adjusted SPPs after bias correction obtained a higher Nash-Sutcliffe Efficiency (NSE) than gauge data when driving streamflow simulation. The unadjusted SPPs hydrological applicability was significantly enhanced after bias correction, with NSE< 0 (except for IMERG product) and NSE> 0.34 before and after correction, respectively. Performance of the TMPA series improved the most through bias correction, making it the preferred choice for daily and monthly simulations. Additionally, SPPs were more applicable in wet and normal years and require improvement in dry years. This study provides valuable references for identifying better alternative precipitation sources for local water resource management and exploring an effective SPP utilization approach in practical applications.
Article
Full-text available
To understand and manage water systems under a changing climate and meet an increasing demand for water, a quantitative understanding of precipitation is most important in coastal regions. The capabilities of the Integrated Multisatellite Retrievals for Global Precipitation Measurement (IMERG) V06B product for precipitation quantification are examined over three coastal regions of the United States- the West Coast, the Gulf of Mexico, and the East Coast, all of which are characterized by different topographies and precipitation climatologies. A novel uncertainty analysis of IMERG is proposed, that considers environmental and physical parameters such as elevation and distance to the coastline. The IMERG performance is traced back to its components, i.e. passive microwave (PMW), infrared (IR), and morphing-based estimates. The analysis is performed using high-resolution, high-quality Ground Validation Multi-Radar/Multi-Sensor (GV-MRMS) rainfall estimates as ground reference at the native resolution of IMERG of 30 min and 0.1 deg. IMERG Final (IM-F) quantification performance heavily depends on the respective contribution of PMW, IR and morph components. IM-F and its components overestimate the contribution of light rainfall (<1 mm/h) and underestimate the contribution of high rainfall rates (>10 mm/h) to the total rainfall volume. Strong regional dependencies are highlighted, especially over the West Coast where the proximity of complex terrain to the coastline challenges precipitation estimates. Other major drivers are the distance from the coastline, elevation, and precipitation types, especially over the land and coast surface types, that highlight the impact of precipitation regimes.
Article
Full-text available
The availability of accurate spatiotemporal rainfall data is of utmost importance for reliable predictions from hydro climatological studies. Challenges and limitations faced due to the absence of dense rain gauges (RGs) networks are seen especially in the developing countries. Therefore, alternative rainfall measurements such as Satellite Rainfall Products (SRPs) are used when RG networks are scarce or completely do not exist. Noteworthy, rainfall data retrieved from satellites also possess several uncertainties. Hence, these SRPs should essentially be validated beforehand. The Mahaweli River Basin (MRB), the largest river basin in Sri Lanka is the heart of the country’s water resources contributing to a significant share of the hydropower production and agricultural sector. Given the importance of the MRB, this study explored the suitability of SRPs as an alternative for RG data for the basin. Daily rainfall data of six types of SRPs were extracted at 14 locations within the MRB. Thereafter, statistical analysis was carried out using continuous and categorical evaluation indices to evaluate the accuracy of SRPs. Non-parametric tests, including the Mann-Kendall and Sen’s slope estimator tests were used to detect the possibility of trends and the magnitude, respectively. Integrated Multi-satellite Retrievals for Global Precipitation Measurement (IMERG) outperformed among all SRPs, while Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) products showed dire performances. However, IMERG also demonstrated underestimations when compared to RG data. Trend analysis results showcased that the IMERG product agreed more with RG data in monthly and annual time scales while Tropical Rainfall Measurement Mission Multisatellite Precipitation Analysis – 3B42 (TRMM-3B42) agreed more on the seasonal scale. Overall, IMERG turned out to be the best alternative among the SRPs analyzed for MRB. However, it was clear that these products possess significant errors which cannot be ignored when using them in hydrological applications. The results of the study will be valuable for many parties including river basin authorities, agriculturists, meteorologists, hydrologists, and many other stakeholders.
Article
This study is a statistical and hydrological assessment of three high-spatial resolution precipitation products over Ouergha basin (Northern Morocco). PESIANN-CCS (Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks - Cloud Classification System) 0.04° × 0.04°, CHIRPS (Climate Hazards Group InfraRed Precipitation with Station data) 0.05°× 0.05°, and TRMM 3B42 (Tropical Rainfall Measuring Mission version 7) 0.25° × 0.25° are assessed using observed data from 11 rain gauge stations and 4 flow-gauge stations, over the period 2003–2010. This assessment is performed at different time and space scales using continuous and categorical statistical scores, and finally a hydrological modeling using GR4J Model. Results show that all the products perform poorly at daily time steps. PERSIANN and TRMM have a negative bias while CHIRPS has a positive one. CHIRPS and TRMM exhibit a significant improvement at monthly scale. In winter, all products have the best correlation but also the largest errors. PERSIANN generously overestimates the summer rain. CHIRPS shows the best annual correlation and captures very well the spatial pattern of mean annual precipitations. CHIRPS also shows a better performance over the PERSIANN and TRMM in reproducing stream flow. However, it’s still not enough since the Nash criterion is around 0.52. Results of this study show the necessity of a bias correction in order to increase the accuracy of the satellite precipitation products in estimating daily rainfall and reproducing streamflow. This correction may be the subject of a further study.
Article
Full-text available
The Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis (TMPA) provides a calibration-based sequential scheme for combining precipitation estimates from multiple satellites, as well as gauge analyses where feasible, at fine scales (0.25° × 0.25° and 3 hourly). TMPA is available both after and in real time, based on calibration by the TRMM Combined Instrument and TRMM Microwave Imager precipitation products, respectively. Only the after-real-time product incorporates gauge data at the present. The dataset covers the latitude band 50°N-S for the period from 1998 to the delayed present. Early validation results are as follows: the TMPA provides reasonable performance at monthly scales, although it is shown to have precipitation rate-dependent low bias due to lack of sensitivity to low precipitation rates over ocean in one of the input products [based on Advanced Microwave Sounding Unit-B (AMSU-B)]. At finer scales the TMPA is successful at approximately reproducing the surface observation-based histogram of precipitation, as well as reasonably detecting large daily events. The TMPA, however, has lower skill in correctly specifying moderate and light event amounts on short time intervals, in common with other finescale estimators. Examples are provided of a flood event and diurnal cycle determination.
Article
Full-text available
In this study, the recent work of Gottschalck et al. and Ebert et al. is extended by assessing the suitability of two Tropical Rainfall Measuring Mission (TRMM)-based precipitation products for hydrological land data assimilation applications. The two products are NASA's gauge-corrected TRMM 3B42 Version 6 (3B42), and the satellite-only NOAA Climate Prediction Center (CPC) morphing technique (CMORPH). The two products were evaluated against ground-based rain gauge-only and gauge-corrected Doppler radar measurements. The analyses were performed at multiple time scales, ranging from annual to diurnal, for the period March 2003 through February 2006. The analyses show that at annual or seasonal time scales, TRMM 3B42 has much lower biases and RMS errors than CMORPH. CMORPH shows season-dependent biases, with overestimation in summer and underestimation in winter. This leads to 50% higher RMS errors in CMORPH's area-averaged daily precipitation than TRMM 3B42. At shorter time scales (5 days or less), CMORPH has slightly less uncertainty, and about 10%-20% higher probability of detection of rain events than TRMM 3B42. In addition, the satellite estimates detect more high-intensity events, causing a remarkable shift in precipitation spectrum. Summertime diurnal cycles in the United States are well captured by both products, although the 8-km CMORPH seems to capture more diurnal features than the 0.25° CMORPH or 3B42 products. CMORPH tends to overestimate the amplitude of the diurnal cycles, particularl in the United States. Possible causes for the discrepancies between these products are discussed.
Article
Full-text available
A detailed description of the operational WSR-88D rainfall estimation algorithm is presented. This algorithm, called the Precipitation Processing System, produces radar-derived rainfall products in real time for forecasters in support of the National Weather Service's warning and forecast missions. It transforms reflectivity factor measurements into rainfall accumulations and incorporates rain gauge data to improve the radar estimates. The products are used as guidance to issue flood watches and warnings to the public and as input into numerical hydrologic and atmospheric models. The processing steps to quality control and compute the rainfall estimates are described, and the current deficiencies and future plans for improvement are discussed.
Article
Full-text available
An increasing number of satellite-based rainfall products are now available in near-real time over the Internet to help meet the needs of weather forecasters and climate scientists, as well as a wide range of decision makers, including hydrologists, agriculturalists, emergency managers, and industrialists. Many of these satellite products are so newly developed that a comprehensive evaluation has not yet been undertaken. This article provides potential users of short-interval satellite rainfall estimates with information on the accuracy of such estimates. Since late 2002 the authors have been performing daily validation and intercomparisons of several operational satellite rainfall retrieval algorithms over Australia, the United States, and northwestern Europe. Short-range quantitative precipitation forecasts from four numerical weather prediction (NWP) models are also included for comparison. Synthesis of four years of daily rainfall validation results shows that the satellite- derived estimates of precipitation occurrence, amount, and intensity are most accurate during the warm season and at lower latitudes, where the rainfall is primarily convective in nature. In contrast, the NWP models perform better than the satellite estimates during the cool season when nonconvective precipitation is dominant. An optimal rain-monitoring strategy for remote regions might therefore judiciously combine information from both satellite and NWP models.
Article
Full-text available
PERSIANN, an automated system for Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks, has been developed for the estimation of rainfall from geosynchronous satellite longwave infared imagery (GOES-IR) at a resolution of 0.25° × 0.25° every half-hour. The accuracy of the rainfall product is improved by adaptively adjusting the network parameters using the instantaneous rain-rate estimates from the Tropical Rainfall Measurement Mission (TRMM) microwave imager (TMI product 2A12), and the random errors are further reduced by accumulation to a resolution of 1° × 1° daily. The authors' current GOES-IR-TRMM TMI based product, named PERSIANN-GT, was evaluated over the region 30°S-30°N, 90°E-30°W, which includes the tropical Pacific Ocean and parts of Asia, Australia, and the Americas. The resulting rain-rate estimates agree well with the National Climatic Data Center radar-gauge composite data over Florida and Texas (correlation coefficient r > 0.7). The product also compares well (r ~ 0.77-0.90) with the monthly World Meteorological Organization gauge measurements for 5° × 5° grid locations having high gauge densities. The PERSIANN-GT product was evaluated further by comparing it with current TRMM products (3A11, 3B31, 3B42, 3B43) over the entire study region. The estimates compare well with the TRMM 3B43 1° × 1° monthly product, but the PERSIANN-GT products indicate higher rainfall over the western Pacific Ocean when compared to the adjusted geosynchronous precipitation index-based TRMM 3B42 product.
Article
This paper is concerned with the effect of precipitation forcing on land surface hydrological variables predicted by a physically based land surface scheme. The aspects considered are the differences in precipitation input across varying sensor measurements and temporal scales of aggregation. Precipitation accumulations at 1-, 2-, 3-, and 6-h time scales are derived on the basis of standard 5-min rain gauge rainfall measurements, hourly rain gauge calibrated WSR-88D radar rainfall estimates, and passive microwave calibrated half-hourly satellite infrared rain retrievals. The spatial resolution of the rainfall estimates is fixed to 1° grid boxes. The off-line community land model (CLM) is used to simulate land surface parameters on the basis of external meteorological forcing parameters. The study region and data consist of two vegetation-distinct (high and low vegetation cover) sites in Oklahoma. The data used include one warm season (May-August 2002) of in situ meteorological data from the Oklahoma Mesonet. The CLM is forced with the three different rainfall input datasets for varying temporal scales (1-6 h). Relative difference statistics in terms of rainfall and land surface parameters are presented between the two remote sensing rain retrievals and the gauge rainfall measurements used as reference. Results show that the hydrological response is nonlinear and strongly dependent on the error characteristics of the retrieval (e.g., more temporal correlated rainfall error results in higher error propagation in land surface parameters). We also investigate the temporal lag correlation of the error in rainfall with the error in the various land surface hydrological variables. Time resolution is shown to have an effect on the error statistics of the hydrologic variables. Coarse time resolutions are associated with errors of lower variance and higher correlation.
Article
We first develop the theory needed to interpret the vertically distributed radiative sources and the emission-absorption-scattering processes responsible for the behavior of frequency-dependent top-to-atmosphere brightness temperatures TB's. This involves two distinct types of vertical weighting functions for the TB's: an "emission-source weighting function' describing the origin of emitted radiation that eventually reaches a satellite radiometer, and "generalized weighting function' describing emitted-scattered radiation undergoing no further interactions prior to interception by the radiometer. The weighting-function framework is used for an analysis of land-based precipitation processes within a hail-storm simulation originally described in Part I. The individual roles of cloud drops, rain drops, graupel particles, ice crystals, and snow aggregates - as well as absorbing gases, the earth's surface, and cosmic background - on generating and modulating the frequency-dependent TB's are examined in detail. Finally, a summary of the various components of a hybrid statistical-physical rainfall algorithm used to produce liquid-ice profile information, as well as surface rain rates, is given. The algorithm employs the cloud model to provide a consistent and objectively generated source of detailed microphysical information as the underpinnings to an inversion-based perturbative retrieval scheme. -from Authors
Article
The most common rainfall measuring sensor for validation of radar-rainfall products is the rain gauge. However, the difference between area-rainfall and rain gauge point-rainfall estimates imposes additional noise in the radar-rain gauge difference statistics, which should not be interpreted as radar error. A methodology is proposed to quantify the radar-rainfall error variance by separating the variance of the rain gauge area-point rainfall difference from the variance of radar-rain gauge ratio. The error in this research is defined as the ratio of the 'true' rainfall to the estimated mean-areal rainfall by radar and rain gauge. Both radar and rain gauge multiplicative errors are assumed to be stochastic variables, lognormally distributed, with zero covariance. The rain gauge area-point difference variance is quantified based on the areal-rainfall variance reduction factor evaluated in the logarithmic domain. The statistical method described here has two distinct characteristics: first, it proposes a range-dependent formulation for the error variance, and second, the error variance estimates are relative to the mean rainfall at the radar product grids. Two months of radar and rain gauge data from the Melbourne, Florida, WSR-88D are used to illustrate the proposed method. The study concentrates on hourly rainfall accumulations at 2- and 4-km grid resolutions. Results show that the area-point difference in rain gauge rainfall contributes up to 60% of the variance observed in radar-rain gauge differences, depending on the radar grid size, the location of the sampling point in the grid, and the distance from the radar.