JANUARY 2008
LAMPTEY
185
Comparison of Gridded Multisatellite Rainfall Estimates with Gridded Gauge Rainfall
over West Africa
BENJAMIN L. LAMPTEY
National Center for Atmospheric Research,* Boulder, Colorado
(Manuscript received 12 September 2006, in final form 18 April 2007)
ABSTRACT
Two monthly gridded precipitation datasets of the Global Precipitation Climatology Project (GPCP; the
multisatellite product) and the Global Precipitation Climatology Centre (GPCC) Variability Analysis of
Surface Climate Observations (VASClimO; rain gauge data) are compared for a 22-yr period, from January
1979 to December 2000, over land areas (i.e., latitudes 4°–20°N and longitudes 18°W–15°E). The two
datasets are consistent with respect to the spatial distribution of the annual and seasonal rainfall climatology
over the domain and along latitudinal bands. However, the satellite generally overestimates rainfall. The
inability of the GPCC data to capture the bimodal rainfall pattern along the Guinea coast (i.e., south of
latitude 8°N) is an artifact of the interpolation of the rain gauge data. For interannual variability, the
gridded multisatellite and gridded gauge datasets agree on the sign of the anomaly 15 out of the 22 yr (68%
of the time) for region 1 (between longitude 5° and 18°W and north of latitude 8°N) and 18 out of the 22
yr (82% of the time) for region 2 (between longitude 5°W and 15°E and north of latitude 8°N). The datasets
agreed on the sign of the anomaly 14 out of the 22 yr (64% of the time) over the Guinea Coast. The
magnitudes of the anomaly are very different in all years. Most of the years during which the two datasets
did not agree on the sign of the anomaly were years with El Niño events. The ratio of the seasonal
root-mean-square differences to the seasonal mean rainfall range between 0.24 and 2.60. The Kendall’s tau
statistic indicated statistically significant trends in both datasets, separately.
1. Introduction
The semiarid Sahelian climate, the fact that agriculture is the backbone of the economy of most of the
West African countries, and the current drought (e.g.,
Nicholson 1980; Giannini et al. 2003; Zeng 2003) make
a study of rainfall patterns over West Africa important.
A comparison of rainfall estimates from satellite and
rainfall amounts from gauges (i) could be of value in
assessing and improving the usefulness of satellite imagery in rainfall estimation; (ii) is relevant to the use of
real-time rainfall data for flood warnings and river flow
estimates or water management, and for pinpointing
* The National Center for Atmospheric Research is sponsored
by the National Science Foundation.
Corresponding author address: Dr. Benjamin L. Lamptey, National Center for Atmospheric Research, 3450 Mitchell Lane,
Boulder, CO 80301.
E-mail: [email protected]
DOI: 10.1175/2007JAMC1586.1
© 2008 American Meteorological Society
areas of likely agricultural impacts (i.e., locust warnings, drought monitoring, crop production and food security, and land-use changes); (iii) will add to our understanding of the limitations and advantages of the use
of satellites to overcome problems associated with global precipitation measurements (Hulme 1995); and (iv)
will enable estimation of the extent to which satellite
observations can be relied upon to give a good database
needed for the analysis of rainfall variability. An additional benefit will be obtained if a quantitative relationship between satellite estimates and surface rainfall is
established. This will give an indication of where and
when the spatially and temporally regular satellite data
can be used to estimate rainfall in areas with sparse rain
gauges. In a wider context, the study will also help
tackle the economic impacts of climate variability.
For instance, climate variability (or for that matter
weather) has economic consequences relating to health,
agriculture (e.g., variability in crop yield), forest fires
(forest products are economically important), energy
(energy demands and alternative energy sources), outdoor activities such as construction, the design of farm
186
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY
machinery and its efficient operation, transportation
(including aviation), and recreation.
A variety of techniques exist for retrieving rainfall
estimates from satellite data. For instance, satellite observations of infrared and microwave radiance have
been used to retrieve precipitation information over
many parts of the globe (Barrett and Martin 1981; Arkin and Ardanuy 1989; Huffman et al. 1995, 1997).
Over the Sahelian region of West Africa alone, satellite
rainfall estimation techniques include the Tropical Agricultural Meteorology using satellite and other data
(TAMSAT), the Polar-orbiter Effective Rainfall Monitoring Integrative Technique (PERMIT; e.g., Grimes et
al. 2003), the Agricultural Drought Monitoring Integrated Technique (ADMIT), and the Environmental
Analysis and Remote Sensing methods (EARS-e,
EARS-m and EARS-c; Snijders 1991) and the Tropical
Rainfall Measuring Mission (TRMM) satellite (e.g.,
Kummerow et al. 2000; Sealy et al. 2003). A review of
satellite rainfall climatology is contained in Kidd
(2001).
For hydrometeorological purposes, satellite data refer to a pixel area and can easily be integrated over a
catchment or subcatchment. Satellite data is useful in
locating the position of the intertropical convergence
zone (ITCZ) and enables the position of the ITCZ over
land and water to be estimated (e.g., Ba et al. 1995).
However, satellite data are mainly useful for convective
rain over land, where passive microwave methods are
limited to the scattering of ice particles to affect the
higher frequency (⬃85 GHz) channels, they need local
calibration against rain gauges, and the estimates they
provide are in some cases contaminated by high-level
cirrus clouds.
While rain gauges contain sampling errors, satellite
rainfall estimates over land contain nonnegligible random error and bias because of the indirect nature of the
relationship between observations and the precipitation, the inadequate sampling, and algorithm imperfections (e.g., Xie and Arkin 1997; Ali et al. 2005a,b).
Three major deficiencies that exist in the different
datasets of large-scale precipitation are incomplete global coverage (for some datasets), significant random
error, and nonneligible bias. The individual data
sources, however, present similar distribution patterns
of large-scale precipitation, but with differences in
smaller-scale features and in amplitudes (Xie and Arkin 1995). The different sources have been combined in
various instances to take advantage of the strengths of
each, to produce the best possible analysis (gridded
field) of precipitation. Examples of such datasets include those of the Global Precipitation Climatology
Project (GPCP; Xie and Arkin 1995, 1997; Huffman et
VOLUME 47
al. 1997), the Global Precipitation Climatology Centre
(GPCC; Rudolf et al. 1994), the Climatic Research Unit
(New et al. 1999, 2000), and the dataset of Dai et al.
(1997). Note that for gauge datasets, combination here
refers to the use of different sources of gauge rainfall to
form a gauge-only database.
Previous satellite-based studies of the mean state and
variability of rainfall over West Africa include those by
Chapa et al. (1993), Ba et al. (1995), Laurent et al.
(1997), Mathon et al. (2002), Sealy et al. (2003), and
Nicholson (2005). Other related studies include those
that validate the satellite rainfall estimates (e.g.,
Snijders 1991; Laurent et al. 1998a; Nicholson et al.
2003), estimate areal rainfall using both satellite- and
ground-based data (e.g., Grimes et al. 1999), and combine both satellite- and ground-based data to study
rain-producing systems (e.g., Laurent et al. 1998b). Satellite data (sometimes combined with other types of
data) have also been used in studies over other parts of
Africa (e.g., Nicholson 1994; Xie and Arkin 1995; Grist
et al. 1997; McCollum et al. 2000; Thorne et al. 2001;
Grimes et al. 2003). The aforementioned studies made
significant contributions to our understanding of the
use of satellite estimates. However, in comparing satellite observations with rain gauge data, they have used
station rainfall (not gridded gauge rainfall). Also, the
data used in most of the above studies have covered
relatively short periods of time in comparison with the
22 yr of both gridded gauge and satellite data used in
this study.
This work will provide a comparison between the
GPCP multisatellite product and the GPCC gauge
datasets. The objective of this study is to examine how
the gridded satellite-only and gridded gauge-only data
compare with regard to the mean rainfall as well as
spatial and interannual variability. Specific questions to
be addressed are how do the satellite data compare
with the gauge data with regard to
1) mean annual, seasonal, and monthly rainfall over
the whole domain, along latitudinal bands, and in
three distinct regions over the domain;
2) spatial variability over the domain at the annual and
seasonal time scales;
3) interannual variability in the three regions;
4) the mean state and fluctuations during El Niño and
La Niña events within the study period; and
5) trends in either dataset. If there are trends, are the
trends statistically significant?
The study area is between 4° and 20°N and 15°E and
the west coast, 18°W (boxed area in Fig. 1). The period
of comparison is from 1979 to 2000. The study domain
is divided into three regions as shown in Fig. 2. In re-
JANUARY 2008
LAMPTEY
187
al. (2003). A brief description of the datasets is given in
this section.
a. Gauge precipitation data
FIG. 1. Study domain marked with the box.
gions 1 and 2, rainfall is governed by squall lines and
thunderstorms. The distinction between regions 1 and 2
is the oceanic influence in the former. The boundary
between the two regions is 5°W. Region 3 (located
south of 8°N) is mainly affected by the localized convection of the monsoon, with minimal influence of the
synoptic and mesoscale systems of the Sahel (Ba et al.
1995). Note that because the GPCC gauge dataset covers only land areas, region 3 in this study, is smaller
than what it should have been.
2. Data
The gridded rain gauge data used are that of Beck et
al. (2005) and the satellite estimates are from Adler et
The rain gauge dataset used in this work is from the
Variability Analysis of Surface Climate Observations
(VASClimO; more information available online at
http://www.dwd.de/en/FundE/Klima/KLIS/int/GPCC/
Projects/VASClimO/VASClimO.htm) 50-yr precipitation climatology. These data are gridded at 0.5°, 1.0°,
and 2.5° latitude–longitude resolutions. A detailed description of the construction of the monthly gridded
dataset of precipitation is given in Beck et al. (2005).
For this work the 2.5° resolution dataset was used because the satellite estimates have 2.5° resolution.
The gridded monthly terrestrial precipitation dataset
was developed on the basis of the comprehensive database of monthly observed precipitation data that resides with the GPCC. Only station time series with a
minimum of 90% data availability during the analyzed
period (1951–2000) are used for interpolation to a regular 0.5° ⫻ 0.5° grid in order to minimize the risk of
generating temporal inhomogeneities in the gridded
data due to varying station densities. The method of
ordinary kriging was used for the interpolation. This is
because it yielded the least interpolation error of several methods tested (including the method of Shepard,
which is the standard method used in GPCC; Beck et al.
2005). To estimate grid averages of the precipitation,
the interpolated values at the four corner points of a
0.5° ⫻ 0.5° grid cell are averaged. Global maps with 1°
and 2.5° resolution were produced by area-weighted
averaging of the 0.5° grids.
The GPCC has the following gridded rain gauge
products (available online at http://gpcc.dwd.de):
FIG. 2. Suggested rainfall zones in West Africa.
188
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY
1) Reanalysis of the GPCC full data product (current
version 3) for the period 1951–2004. The full data
product is based on all available monthly in situ precipitation data and includes national collections, regional datasets and data from the Global Telecommunication System (GTS) of the World Meteorological Organization (WMO). The full data product
provides the best spatial data coverage for each individual month and is recommended for water budget studies.
2) VASClimO (current version 1.1) 50-yr climatology.
This new analysis product of monthly precipitation,
developed within the project “Variability Analysis
of Surface Climate Observations” of the German
Climate Research Program is for the period 1951–
2000. This product is preferred for analysis of temporal climate variability, in particular the spatial distribution of climate change with respect to precipitation. The GPCC VASClimO product is an input to
the Intergovernmental Panel on Climate Change
(IPCC) Fourth Assessment Report (AR4).
GPCC maintains the following two quasi-operational
analysis products:
1) First-guess product (or land surface precipitation
anomaly)—This is the monthly global land surface
precipitation. The data are available within 5 days
after the end of the month, based on the globally
disseminated synoptic weather reports (SYNOP).
This product is used by the United Nations Food
and Agricultural Organization (FAO) for drought
monitoring.
2) Monitoring product—This product is available
about 2 months after the end of a month, based on
the synoptic reports received at the GPCC and the
National Oceanic and Atmospheric Administration
(NOAA) and data from globally disseminated
monthly climate bulletins (CLIMAT). This product
is used by the Global Energy and Water Cycle Experiment (GEWEX) and the GPCP as a near-realtime in situ reference for adjustment of satellitebased precipitation estimates.
b. Satellite data
The satellite estimates used in this work are the
GPCP monthly multisatellite rainfall estimate with a
latitude–longitude resolution of 2.5° by 2.5°. This
dataset is for the period 1979 to the present, and a
description can found in Adler et al. (2003). This
dataset combines the strengths of the different satellite
rainfall estimation systems, and removes biases based
on hierarchical relations in a stepwise approach.
VOLUME 47
It is worth noting that the version of the data used for
the study is the latest that was released in June 2006.
This was after an inhomogeneity in the dataset, which
occurred around 1987 (the period when SSMI/I data
started to become available), was removed or corrected
(G. Bolvin 2006, personal communication).
The microwave estimates are based on Special Sensor Microwave Imager (SSM/I) data from the Defense
Meteorological Satellite Program (DMSP, United
States) satellites, which fly in sun-synchronous lowEarth orbits. The infrared precipitation estimates are
obtained primarily from geostationary satellites operated by the United States, Europe, and Japan, and secondarily from polar-orbiting satellites operated by the
United States. Additional precipitation estimates are
obtained based on Television and Infrared Observation
Satellite (TIROS) Operational Vertical Sounder
(TOVS). The merging of these products utilizes the
higher accuracy of the low-orbit microwave observations to calibrate, or adjust, the more frequent geosynchronous infrared observations. The dataset is extended back into the premicrowave era (before mid1987) by using infrared-only observations calibrated to
the microwave-based analysis of the later years. Microwave- and TOVS-based estimates were obtained by utilizing separate relationships over ocean and land. The
geosynchronous IR-based estimates employ the Geostationary Operational Environmental Satellite (GOES)
Precipitation Index (GPI) technique, which relates cold
cloud-top area to rain rate. The geostationary satellite
operators that contributed to the data are GOES
(United States); the Geosynchronous Meteorological
Satellite (GMS; Japan); and the Meteorological Satellite [Meteosat; the European community (now Eumesat)].
3. Methods
The statistical measures used are the following.
a. The bias
This is the difference between the mean reference (or
gridded rain gauge) value and the mean gridded satellite rainfall estimate. The bias describes whether the
multisatellite data has over- or underestimated rainfall.
It is assumed here that the gridded gauge rainfall data
are the reference, although there may be biases associated with the rain gauge estimates. In this study the bias
was modified by dividing it by the average gauge rainfall. This modified bias is the fractional difference field,
which reflects the nature of the differences with respect
JANUARY 2008
189
LAMPTEY
to the magnitudes of the precipitation values. This
quantity is more useful than the bias field alone.
b. The standardized departure or anomaly
This standardizes deviations or anomalies with respect to the standard deviation (Gregory 1978; Nicholson 1980; Wilks 2006). Standardized anomalies reduce
the influence of the mean and facilitate comparison between different areas or nonhomogeneous regions. The
standardized anomaly, z, is computed as follows:
z⫽
共x ⫺ x兲
,
sx
共1兲
where x is the raw data, x is the sample mean of the raw
data, and sx is the corresponding sample standard deviation.
square of the correlation coefficient. In short, in regression, a relationship between a dependent variable ( y)
and one or more independent variables (xs) is being
sought. Correlation seeks to examine the degree of association between two variables. Actually, it only quantifies the degree of linear association. Even, with a very
low r there may still be a relationship, but it is just not
linear. The values of the correlation coefficient range
from ⫺1 to ⫹1 inclusive, with ⫹1 indicating a perfect
linear combination. The coefficient of determination
has values ranging from 0 to 1 inclusive. Murphy (1995)
gives a good discussion of the correlation coefficient (or
coefficient of correlation) and the coefficient of determination.
The sample correlation coefficient (Pearson linear
correlation coefficient) r is given by
r⫽
c. The root-mean-square difference (rmsd)
兺共x ⫺ x兲共 y ⫺ y兲
公兺共x ⫺ x兲 兺共 y ⫺ y兲
i
i
2
2
i
The root-mean-square difference of all elements of
the time dimension (year, season, month) at all grid
points is computed. The rmsd is used because both variables, gridded rain gauge data and gridded satellite
data, may contain errors:
rmsd ⫽
冑兺
1
n
n
共ei ⫺ oi兲2,
共2兲
i
where e is the multisatellite value and o is the gauge
value.
The rmsd was modified in this study. The ratio of the
rmsd to the average gauge precipitation was used. The
rmsd is roughly the square root of rain and will give a
wide range of values. The modified rmsd is a more
powerful statistic.
d. Correlation coefficient (r)
The term correlation coefficient is usually used to
mean the Pearson product-moment coefficient of linear
correlation between two variables x and y (Wilks 2006).
Simple linear regression simply consists of estimating a
line through a set of points when relationship appears
linear. The correlation coefficient measures the
strength of the linear association between two variables. Whereas in regression it matters which variable is
x and which is y, for the correlation coefficient, it does
not. The coefficient of determination specifies the proportion of the variability of one of the two variables
that is linearly accounted, or described by the other. It
is not the amount of the variance of one variable “explained” by the other (Wilks 2006). For simple linear
regression, the coefficient of determination r 2 is the
i
⫽
Sxy
公Sxx Syy
,
共3兲
where Sxy is the corrected sum of the products of x and
y, Sxx is the corrected sum of squares of x, and Syy is the
corrected sum of squares of y. Note that x and y are the
mean of x and y, respectively.
The linear correlation coefficient r can also be
given by
n
兺 关共x ⫺ x
i
r⫽
ave 兲共 yi
⫺ yave兲兴
i⫽1
n␴x␴y
,
共4兲
where xi ( yi) is the set of n reference (estimated) values
and xave ( yave) and ␴x and ␴y are the mean and standard
deviations of the reference (estimated) values, respectively (Laurent et al. 1998a).
Another way to view the Pearson correlation coefficient is the ratio of the sample covariance of the two
variables to the product of the two standard deviations:
r⫽
cov共x, y兲
,
sx sy
共5兲
where sx is the standard deviation of variable x and sy is
the standard deviation of variable y (Wilks 2006; Murphy 1995). It is worth noting that the linear correlation
coefficient is not sensitive to a bias (Laurent et al.
1998a).
1) KENDALL’S TAU CORRELATION COEFFICIENT (␶)
Kendall’s tau correlation coefficient is also known as
Kendall rank-order correlation coefficient (Siegel and
190
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY
Castellan 1988). Correlation coefficients measure the
strength of association between two continuous variables. When one of the variables is a measure of time or
location, the correlation becomes a test for temporal or
spatial trend. Data may be correlated in either a linear
or nonlinear fashion. When y generally increases or
decreases as x increases, the two variables are said to
possess a monotonic correlation. This correlation may
be nonlinear, with exponential patterns or some other
patterns. Kendall’s tau is an example of a measure of
correlation that measures all monotonic correlations
(linear and nonlinear). Kendall’s tau is based on ranks
and is resistant to outliers, invariant to monotonic
power transformation of one or both variables (Helsel
and Hirsch 1991), insensitive to the distribution of the
data (Karl and Knight 1998), and is powerful in rejecting the hypothesis of no trend in the data. The Kendall
tau statistic is used to assess significance of linear
trends. Tau generally has lower values than the traditional correlation coefficient r for linear associations of
the same strength. Its large sample approximation produces p values very near-exact values, even for small
sample sizes.
The test statistic S is given by
S ⫽ P ⫺ M,
共6兲
where P is the number of (x, y) pairs where y increases
with increasing x (i.e., concordant pairs or “number of
pluses”) and M is the number of (x, y) pairs where y
decreases as x increases (i.e., discordant pairs or “number of minuses”).
Kendall’s tau correlation coefficient is given by
␶⫽
S
,
n共n ⫺ 1兲Ⲑ2
共7兲
where n is the number of data pairs. The significance of
␶ is tested by comparing S with what would be expected
when the null hypothesis H0 is true. If it is further from
0 than expected, H0 is rejected. For n ⱕ 10 an exact test
should be computed. The critical values should be
found from a table of the normal distribution.
2) LARGE-SAMPLE
APPROXIMATION OF
␶
For n ⬎ 10 the test statistic can be modified to be
closely approximated by a normal distribution. The
Kendall’s tau large sample approximation is of the
same form as that of the rank-sum test (also called
Wilcoxon rank-sum test or Mann–Whitney test or Wilcoxon–Mann–Whitney rank-sum test) but with a slight
modification. The standardized form of the rank-sum
test statistic Zs is given by
Zs ⫽
S⫺1
if S ⬎ 0,
␴s
Zs ⫽ 0
Zs ⫽
if S ⫽ 0,
S⫺1
if
␴s
VOLUME 47
or
共8兲
or
共9兲
S ⬍ 0.
共10兲
The null hypothesis is rejected at a significance level ␣
if | Zs | ⬎ Zcrit, where Zcrit is the value of the standard
normal distribution with a probability of exceedance of
␣/2. In the case where some of the x and (or) y values
are tied, the formula for ␴s needs to be modified
(Helsel and Hirsch 1991).
e. Pattern correlation
This measures the overall agreement between two
grids.
A high pattern correlation indicates that both the
amplitude and phase of the two grids are in agreement.
The “pattern correlation” used here is a special case of
the linear correlation coefficient when dimensions of
both grids are identical.
The previous statement was made because the Pearson correlation coefficient between two fields that have
been interpolated onto the same grid of n points, is
sometimes called a pattern correlation. This is because
pairs of maps with similar patterns will yield high correlations (Wilks 2006). Pattern correlation is also defined as the Pearson product-moment coefficient of linear correlation between two variables that are, respectively, the values of the same variables at corresponding
locations on two different maps (Glickman 2000). The
two different maps can be for example different times,
for different levels in the vertical direction, or forecast
and observed values. This use of the pattern correlation
is occasionally referred to as map correlation.
A commonly used measure of association that operates on pairs of gridpoint values in the forecast and
observed fields, is the anomaly correlation (AC). It is
designed to detect similarities in the patterns of departure (i.e., anomalies) from the climatological field, and
is sometimes referred to as pattern correlation. Like the
Pearson correlation, it is bounded by ⫾1 and is not
sensitive to bias in the forecast. The anomaly correlation (as defined by Glickman 2000) is a special case of
pattern correlation for which the variables being correlated are the departure from some appropriately defined mean, most commonly a climatological mean.
f. Coefficient of variation
This is also called relative variability (or dispersion).
It gives the relative or proportional variability (in a
number of observations). It is given by
JANUARY 2008
191
LAMPTEY
cv ⫽
␴
100,
x
共11兲
where ␴ is the standard deviation and x is the mean
(Gregory 1978). This statistical parameter can be considered as interannual variability expressed as the percentage of average values (Ba et al. 1995). It is convenient if the variation of different datasets having differing means is to be compared. Thus, it is an appropriate
measure for comparing variability in a nonhomogeneous geographical region.
4. Results
The consistency between the gauge and satellite data
is examined in terms of (i) the mean annual, seasonal
and monthly rainfall fields (i.e., the climatologies), (ii)
the spatial variability at the annual and seasonal time
scales (including effects on zonal averages), and (iii) the
temporal variability at the annual (year-to-year fluctuations) time scales for the study domain as a whole and
separately for the three regions. The consistency between the satellite and gauge data is also examined for
a composite of El Niño and La Niña events (see Table
1). Recall that the season here is from the months of
May–October. The results presented must be viewed
bearing in mind the fact that the distribution of the
precipitation stations in the gridded dataset is such that
the rain gauges are sparse in the northern latitudes
(north of 10°N) as compared with the southern latitudes. (A map of rain gauge locations used in constructing the gridded rain gauge product can be found online
at http://www.dwd.de/en/FundE/Klima/KLIS/int/
GPCC/Projects/VASClimO/VASClimO.htm.)
For the regions or rainfall zones (Fig. 2) used in the
study, the analysis software NCAR Command LanTABLE 1. Listing of El Niño and La Niña events between 1979
and 1999 as defined by SSTs in the Niño-3.4 region and exceeding
⫾0.4°C threshold. The starting and ending month of each is given
along with the duration in months. [Source: Trenberth (1997) and
online at http://www.cgd.ucar.edu/cas/papers/clivar97/table_2.
html.]
El Niño events
La Niña events
Begin
End
Duration
Begin
End
Duration
Oct 1979
Apr 1982
Aug 1986
Mar 1991
Feb 1993
Jun 1994
Apr 1997
Apr 1980
Jul 1983
Feb 1988
Jul 1992
Sep 1993
Mar 1995
Apr 1998
7
16
19
17
8
10
13
Sep 1984
May 1988
Sep 1995
Jul 1998
Jun 1985
Jun 1989
Mar 1996
Dec 1999
10
14
7
18*
* Extended beyond Dec 1999.
guage (NCL) selects areas greater (or less) than the
specified value. For example, for a region between latitudes 8° and 20°, the software selects all latitudes
greater than or equal to 8° and less than or equal to
20°N. This implies there will be some issues (common
to both datasets) where that boundary is not at the edge
of a grid box. This could have some effect on the results
even though the same thing was done for both the
gauge and multisatellite datasets. It is also known that
satellite estimates are complimentary to gauge data, but
this analysis seeks to find out what information is common to both a satellite-only database and a gauge-only
database. In fact, the GPCP multisatellite data (an intermediate product) used here has a climatological adjustment to the GPCP satellite–gauge data (the final
product), which is dominated by the GPCC gaugemonitoring product (G. J. Huffman 2006, personal
communication). Thus, it is worth investigating what
information can be obtained from the satellite alone.
The difference between the gridded gauge data used in
this study (i.e., VASClimO) and the GPCC monitoring
product has already been described under the data section.
a. Climatology
1) MEAN
RAINFALL FIELD
The gauge and satellite datasets are fairly consistent
with respect to the percentage contribution of individual months to the mean annual rainfall along the
specified 2.5° latitudinal bands (Tables 2 and 3). The
two datasets are also consistent in capturing the decreasing rainfall amount from south to north over the
study domain (i.e., last column in the two tables). The
unweighted pattern correlation between monthly gauge
and satellite rainfall climatologies for the study domain
as a whole indicated that the amplitude and phase of
both grids are in agreement (i.e., comparable order of
magnitude). The lowest correlation value, 0.60, was in
TABLE 2. Percentage contribution of individual months to mean
annual rainfall (gridded gauge rainfall) for 2.5° latitudinal sectors
in the study domain. Except for the last column, entries display
the ratio of monthly mean to annual mean rainfall expressed as a
percentage. The last column displays the mean annual rainfall
(mm).
Lat
May
Jun
Jul
Aug
Sep
Oct
Mean
17.5°–20.0°N
15.0°–17.5°N
12.5°–15.0°N
10.0°–12.5°N
7.5°–10.0°N
5.0°–7.5°N
3
3
5
8
10
10
9
10
12
13
13
14
28
27
26
22
17
15
36
37
33
27
21
15
19
20
19
19
17
15
3
3
4
7
10
12
143
257
545
1050
1448
1819
192
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY
VOLUME 47
TABLE 3. Percentage contribution of individual months to mean
annual rainfall (gridded satellite rainfall) for 2.5° latitudinal sectors in the study domain. Except for the last column, entries display the ratio of monthly mean to annual mean rainfall expressed
as a percentage. The last column displays the mean annual rainfall
(mm).
Lat
May
Jun
Jul
Aug
Sep
Oct
Mean
17.5°–20.0°N
15.0°–17.5°N
12.5°–15.0°N
10.0°–12.5°N
7.5°–10.0°N
5.0°–7.5°N
4
3
5
8
10
13
5
8
12
14
14
15
16
24
25
21
17
12
32
37
32
25
19
12
19
19
19
18
16
13
6
4
5
8
10
12
95
250
613
1155
1535
1642
January while the highest value, 0.98, was in May. Both
datasets indicate that the percentage contribution of
the seasonal rainfall (May–October) to total annual
rainfall is relatively higher (80% or more), in latitudinal
bands to the north of the 10° latitude. In both datasets,
the month of August shows up as the wettest month in
most parts of the domain (i.e., mostly north of the 7.5°
latitude) followed by July and September. The little dry
season (a dip in August rainfall) occurs around latitude
7° or 8°N along most parts of the West African coast
(e.g., Hayward and Oguntoyinbo 1987; Hartmann
1994). That is, the annual cycle of rainfall to the south
of 8°N is bimodal.
The spatial distributions of rainfall on the annual and
seasonal time scales are very similar (Figs. 3 and 4), so
both time scales will not be considered in all cases. The
satellite captures the mean annual state reasonably well
(Figs. 3a,b) even though it overestimates rainfall in
most parts of the region as indicated by the difference
plot (Fig. 3c). The satellite overestimates annual and
seasonal rainfall over the highlands of Cameroon but
underestimates rainfall over the highlands of Guinea.
This is due to the influence of the Atlantic Ocean over
Guinea. There are relatively more stratiform clouds
over the highlands of Guinea than over the highlands of
Cameroun because of the influence of the Atlantic
Ocean (although both highlands prevent rain from occurring with deep convection; Figs. 3c and 4c). Over
land, the stratiform clouds have a negative impact on
the satellite algorithms, which are mostly suitable for
convective clouds. The less rainfall estimated by the
satellite could also be due to a preset threshold to discard ambiguous data. The gauge data over the highlands of Guinea also benefited from the strong degree
of rainfall enhancement by coastal effects (Ba et al.
1995; McCollum et al. 2000; Nicholson et al. 2003). The
multisatellite data do not appear to capture the coastal
effects. It is worth noting that passive microwave estimates along coast lack proper physics, where ambiguity
FIG. 3. Spatial distribution of 22-yr average rainfall (a) gauge,
(b) satellite, and (c) gauge ⫺ satellite data.
over the water or land contribution to each footprint
becomes problematic from a methodological standpoint, because of the ambiguity in determining whether
a given footprint is land, water, or a combination of
both (McCollum and Ferraro 2005).
The scatterplots of mean annual and seasonal rainfall
for the three regions indicate a relatively more linear
relationship in region 2 than in region 1. The satellite
estimates are biased in proportion to the gauge estimates in region 2 (Fig. 5). The scatterplot for region 3
(which has fewer points than regions 1 and 2) does not
give an indication of a linear relationship. The nonlin-
JANUARY 2008
LAMPTEY
FIG. 4. Spatial distribution of 22-yr mean seasonal (May–
October) rainfall (a) gauge, (b) satellite, and (c) gauge – satellite
data.
ear relationship between gauge and satellite rainfall in
region 3 has to do with (i) the fact that the region is
smaller so the statistical fluctuations do not average out
as nicely as in the larger region, (ii) the mesoscale features that definitely affect the satellite dataset, and (iii)
the inability of the satellite to capture the coastal effects. The complex topography along the coast (highlands and lowlands), different vegetation types (e.g.,
forest and savannah) coupled with the influence of the
ocean create difficult conditions for the calibration algorithms of the satellite as the region is relatively non-
193
homogenous in comparison with the northern latitudes.
The gauge data are also affected, as the monthly rainfall
amounts have to be obtained over complex terrain. Recall that region 3 is mainly affected by the local convection of the monsoon, with minimal influence of the
synoptic and mesoscale systems of the Sahel.
The annual cycle of rainfall is shown in Fig. 6. The
multisatellite data captures the cycle well, including the
double maximum (July and September) and the dip in
August rainfall in region 3. The gauge data do not capture the dip in August rainfall which is a known feature
in region 3 (e.g., Hayward and Oguntoyinbo 1987;
Hartmann 1994) but captures the unimodal nature of
the cycle in regions 1 and 2. The inability of the gauge
dataset to capture the dip in August rainfall is an artifact of the interpolation scheme used, because the
higher-resolution version of the gauge data (i.e., 0.5° by
0.5°) captures the bimodal nature of the rainfall in region 3 (Fig. 7). To further investigate the inability of the
coarse gauge data to capture the bimodal nature of
rainfall in region 3, the 0.5° data was regridded to 2.5°
resolution using the analysis (i.e., NCL) software. This
regridded data (i.e., to 2.5°) captured the bimodal rainfall in region 3 (Fig. 8). The inability of the coarseresolution gauge data to capture the bimodal rainfall in
region 3 is unexpected, as the spatial averaging of fine
resolution data to a coarser resolution is not expected
to affect a temporal signal.
For zonal mean annual rainfall, there is good agreement between the satellite and gauge in most parts of
the region, as shown by the latitudinal profile of zonal
mean annual rainfall (see Fig. 9). The relatively good
agreement between the two datasets in the central latitudes of the domain underscores the fact that the satellite algorithms are more efficient in a homogenous
region. The disagreement of the two datasets in the
extreme northern latitudes is attributed to the sparse
rain gauge distribution in the gauge dataset over that
area. The disagreement between the two datasets in the
extreme southern latitudes is because of the mesoscale
features and coastal effects in that area, mentioned earlier. To show this, a plot of the Climatic Research Unit
(CRU) gridded rain gauge data (at 2.5° resolution) and
the multisatellite data (GPCP at 2.5° resolution) was
made (Fig. 10). The results indicate the region of greatest disagreement between the CRU data and the multisatellite data to be in latitudes south of 10°N. This
shows that the influence of the mesoscale features and
coastal effects is robust. Note that the interpolation
scheme used to construct the CRU dataset is completely different from that used to construct the GPCC
dataset. The difference between CRU and GPCC is
194
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY
VOLUME 47
FIG. 5. Scatterplot of (a),(c),(e) 22-yr mean and (b),(d),(f) seasonal multisatellite rainfall vs
gauge rainfall for (top to bottom) region 1–3. A 1:1 line is fitted to each plot.
due to the two datasets having different climatologies.
Note that neither the CRU nor GPCC dataset address
the issue of the mesoscale features. The issue of mesoscale features will definitely require a denser gauge net-
work. It is worth noting that the station data used by
GPCC included stations from the CRU databases
(Beck et al. 2005). Both CRU and GPCC interpolated
rainfall anomalies in creating their datasets. However,
JANUARY 2008
LAMPTEY
195
FIG. 7. Areal monthly mean rainfall (mm day⫺1) from 0.5°resolution gauge data for region 3. The bimodal rainfall in region
3 is seen.
CRU used an exponential weighting with distance,
while GPCC used kriging.
2) COMPARISON
FIG. 6. Areal monthly mean rainfall (mm day⫺1) from gauge
(G) and multisatellite (S) for regions (a) 1, (b) 2, and (c) 3.
OF OTHER RAINFALL STATISTICS
The correlations are examined in order to conduct a
simple analysis of the random and systematic differences between the gauge and satellite monthly rainfall
data. Temporal correlations are used to study the similarity between the GPCC (gauge) and the GPCP (multisatellite) data during their time history. The spatial
distribution of the temporal correlation show a high
correlation (greater than 0.8), especially north of 10°N
(Fig. 11).
Spatial correlation is used to define the similarity between the spatial patterns of the GPCC gauge and the
GPCP multisatellite datasets at particular times. A time
series of the GPCC–GPCP spatial correlation for selected regions is obtained by calculating spatial correlations for all the time frames (Yin et al. 2004). The
month-by-month behavior of the pattern correlations in
time series form indicate that the correlations for regions 1 and 2 oscillate within a relatively smaller range
than that of region 3 over the 264-month period. Figure
12 shows that the range for the correlations is smaller in
these two regions (between ⫺0.3 and 1.0) relative to
region 3 (between ⫺0.7 and 0.9).
The unweighted pattern correlations between seasonal totals ranged from 0.86 to 0.97 for the 22 annual
values. The pattern correlation (one value) between the
gauge and satellite grids for the seasonal climatologies
196
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY
VOLUME 47
FIG. 9. Latitudinal profile of zonal average (1979–2000) rainfall
for gauge (solid) and satellite (dashed) observations.
was 0.73 and 0.94 for the monthly climatologies. The
unweighted pattern correlation for the monthly climatologies ranged from 0.60 to 0.98. The highest value was
in May. The values of the pattern correlations indicate
that there is a strong linear association between the
gauge and satellite data at the seasonal and monthly
scales over most of the domain. Recall that the pattern
FIG. 8. Areal monthly mean rainfall (mm day⫺1) from 0.5° data
regridded to 2.5° for regions (a) 1, (b) 2, and (c) 3. The bimodal
rainfall in region 3 for the gauge data is evident.
FIG. 10. Latitudinal profile of zonal average (1979–98) rainfall
for CRU gauge (solid) and satellite (dashed) observations.
JANUARY 2008
197
LAMPTEY
FIG. 11. Gridpoint correlation (or temporal correlation at monthly time scales) between
gauge and multisatellite data at lag 0.
correlation, as used here, is a special case of the linear
correlation coefficient.
The spatial distributions of the modified bias, modified rmsd, and the coefficient of determination for seasonal rainfall are shown in Fig. 13. This is to enable the
spatial distribution of these measures to be observed,
instead of having a single value for an area average. The
modified rmsd (for the season) ranged from 0.24 to 2.60
mm day⫺1, while the coefficient of determination
ranged from 0.00 to 0.5. The range of values obtained is
reasonable. The figure shows which geographical areas
have smaller rmsd values (which is desirable) and larger
coefficient of determination values (which is also desirable).
3) INTERANNUAL
VARIABILITY
Figure 14 shows the annual rainfall fluctuations expressed as regionally averaged standardized departures. The two datasets agree on the sign (⫹ or ⫺) of
the anomaly 68% of the time (i.e., 15 out of the 22 yr)
in region 1, 82% of the time (i.e., 18 out of the 22 yr) in
region 2, and 64% of the time (i.e., 14 out of the 22 yr)
in region 3 (see Table 4). The correlation coefficient
between the gauge and multisatellite standardized
anomalies were 0.28 for region 1, 0.54 for region 2, and
0.46 for region 3. It is worth noting that most of the
years in the table are years with some months of El
Niño events. This is not surprising, as El Niño–
Southern Oscillation (ENSO) events occur 55% of the
time (Trenberth 1997).
The algorithm of Press et al. (1996) was used to compute Kendall’s tau to test for temporal trends in either
datasets. Using the terminology in the algorithm, small
values of prob (i.e., the two-sided significance level)
indicate a significant correlation (tau positive) or anticorrelation (tau negative). The value of prob for the
gridded gauge data was 3.04 exp (⫺05) while that for
the gridded satellite data was 2.83 exp (⫺02). Thus, the
Kendall’s tau statistic indicated statistically significant
trends in each dataset. The trend in the multisatellite
data is however, less significant than that in the gauge
data.
4) SPATIAL
VARIABILITY
Spatial variability is expressed in terms of the coefficient of variation. Both datasets indicate greater variability in the 22-yr average rainfall in the northern areas
of the study domain (Fig. 15). The difference plot (Fig.
15c) showed greater variability in the satellite data over
most of the region, except the extreme northwestern
and extreme northeastern parts of the domain. A similar feature is observed on the seasonal time scale although the detailed patterns are different (Fig. 16).
b. Rainfall during El Niño and La Niña events
When stratified by phase of the ENSO cycle, the two
datasets are consistent in the rainfall pattern (Figs.
17a,c,e). The main differences are over the highlands of
Guinea and Cameroun, as in the climatology. Thus, the
performance of the satellite during El Niño years is not
different from non–El Niño years for spatial distribution. The agreement between the gauge and satellite is
also evident during La Niña years (Figs. 17b,d,f).
The satellite estimates are more than the gauge val-
198
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY
VOLUME 47
FIG. 13. Seasonal rainfall statistics: (a) fractional difference
field, (b) ratio of rmsd to mean gauge rainfall, and (c) coefficient
of determination.
ues over the highlands of Cameroon but less over the
highlands of Guinea, just as was observed in the climatology. This is true during both El Niño (Fig. 17e) and
La Niña (Fig. 17f) events.
c. Standardized anomalies during El Niño and La
Niña events
FIG. 12. Monthly time series of the pattern correlation for
regions (a) 1, (b) 2, and (c) 3.
The gauge rainfall indicates mainly positive standardized anomalies over the extreme northwestern and
northeastern parts of the domain, while the satellite
JANUARY 2008
LAMPTEY
FIG. 14. Annual rainfall fluctuations expressed as regionally averaged standardized departure from
(a), (c), (e) gauge and (b), (d), (f) multisatellite observations for (top to bottom) regions 1–3.
199
200
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY
VOLUME 47
TABLE 4. The years during which the gauge and satellite disagreed
on the sign of the anomaly (i.e., interannual variability).
Region
1
2
3
1980
1984
1986
1988
1993
1998
2000
1981
1991
1993
1997
1980
1981
1988
1990
1991
1993
1994
1997
indicates negative standardized anomalies during El
Niño events (Figs. 18a,c). The magnitude of the anomalies from the gauge data is generally greater than those
from the satellite data (Fig. 18e) during El Niño events.
During La Niña events, there is a general agreement
between the two datasets in the northeastern parts of
the domain (Figs. 18b,d). The anomalies from the satellite tend to be generally greater than those from the
gauge (Fig. 18f).
5. Summary and conclusions
This work showed that the multisatellite- and gaugebased analyses were generally consistent in the spatial
distribution of rainfall. An obvious difference between
the two datasets is the difference in rainfall amount.
These differences are due to the reasons discussed by
McCollum et al. (2000), which include the interpolation
technique (or area-averaging technique) used for the
surface data. Area averaging of rain gauge data introduces spatial sampling errors that depend primarily on
the number and distribution of gauges within the 2.5°
latitude–longitude grid box. The spatial sampling error
will be high over the study domain because gauges
are sparse. Other reasons for the multisatellite versus
gauge biases include aerosol and drop size distribution, water vapor and moisture flux (i.e., moisture distribution of the environment), and the presence of
warm versus ice-phase rain processes. Note that the
multisatellite product has a climatological adjustment
to the GPCP satellite–gauge product, which is dominated by the GPCC monitoring product. It is known
that satellite and gauge data are complimentary, so the
purpose of this study is to explore what common information can be obtained from the GPCP multisatellite
product and the GPCC VASClimO (not monitoring)
product.
The fact that the multisatellite and gauge mean annual and seasonal rainfall agreed reasonably well (Figs.
FIG. 15. Spatial distribution of coefficient of variation for mean
(22 yr) rainfall (a) gauge, (b) satellite, and (c) gauge ⫺ satellite
data.
3 and 4) for spatial distribution patterns, is an indication
that the multisatellite can be relied upon for the climatological spatial patterns of rainfall in the domain. Inferences about the latitudinal bands can also be made
from the multisatellite estimates. For example, the satellite and gauge agreed on the percentage contribution
of individual months to mean annual rainfall along the
2.5° latitude–longitude grid boxes, despite large differences in rainfall amounts.
A feature that was present in the annual and seasonal
climatologies, as well as during El Niño and La Niña
JANUARY 2008
LAMPTEY
201
data agree more in the central latitudes of the study
domain than in the extreme northern latitudes and
southern latitudes, for the latitudinal profile of zonal
mean annual rainfall. This is because although rainfall
in the middle and northern latitudes of the domain is
relatively more homogeneous, the rain gauge distribution over the extreme northern latitudes was sparse.
Recall that the satellite performs well in areas with homogeneous rainfall. The disagreement between the two
datasets over the extreme southern latitudes is due to
the different vegetation types and topography along the
coast, coupled with the influence of the ocean, which
affect the performance of the satellites.
The following points summarize the results of the
study:
• The gridded rain gauge and gridded multisatellite
•
•
•
FIG. 16. Spatial distribution of coefficient of variation for mean
seasonal rainfall (a) gauge, (b) satellite, and (c) gauge ⫺ satellite.
events, is the fact that the gauge rainfall amount over
the highlands of Guinea is more than that from the
multisatellite, while over the highlands of Cameroon
the gauge values are less than the multisatellite values.
The effect of the topography is that rain in the highlands occurs without deep convection and hence is not
captured by the multisatellite, which is more suited for
convective rain detection. Also, frequent cirrus cloud
cover and large soil moisture values due to frequent
rain affect the performance of the satellite over land
(Ba et al. 1995).
Figure 9 indicates that the multisatellite and gauge
•
data are fairly consistent for the percentage contribution of each month to the annual rainfall along
latitudinal bands, and in capturing the rainfall gradient over the region (Tables 2 and 3). The pattern
correlation values (the unweighted pattern correlation between monthly gauge and multisatellite climatologies) indicate that the amplitude and phase of
both grids are in agreement (i.e., comparable order of
magnitude).
Although the multisatellite and gauge datasets agree
on the spatial distribution of annual and seasonal
rainfall reasonably well over most of the domain,
there are large differences over the highlands of
Cameroon and Guinea (Figs. 3 and 4) for reasons
given earlier.
The relationship between multisatellite and gauge
rainfall is relatively better in regions 1 and 2 than in
region 3. The linear relationship in region 3 (Fig. 5) is
much noisier than the others.
The gauge dataset captures the unimodal nature of
the annual cycle of rainfall in regions 1 and 2, but is
unable to capture the bimodal nature of the rainfall
in region 3 (Fig. 6). The inability of the gauge dataset
to capture the bimodal rainfall in region 3 is due to
the averaging technique used to obtain the 2.5° resolution gauge data from the 0.5° resolution data since
the 0.5° resolution gauge data as well as the 2.5° resolution data obtained by regridding the 0.5° resolution
data with the analysis software both capture the annual cycle (Figs. 7 and 8). This is very unusual as the
spatial averaging should not affect a temporal signal.
The multisatellite data capture the annual cycle in all
three regions well.
There is good agreement between the multisatellite
and gauge zonal mean annual rainfall in the central
latitudes (Fig. 9) of the domain.
202
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY
VOLUME 47
FIG. 17. Spatial distribution of composite precipitation of seven El Niño events from (a) gauge, (c) multisatellite, and (e) gauge –
satellite and four La Niña events from (b) gauge, (d) multisatellite, and (f) gauge ⫺ satellite data.
• There is a statistically significant trend in the gauge
data as well as in the multisatellite data.
For the El Niño and La Niña events, the following
observations were made.
• During El Niño and La Niña events, the multisatellite
and gauge data are consistent on the spatial distribution of the rainfall, although the multisatellite underestimates rainfall over the highlands of Guinea and
overestimates rainfall over the highlands of Cameroon (Fig. 17).
• The gauge data indicate mainly positive standardized
anomalies over the northwestern and northeastern
parts of the domain, while the multisatellite indicates
negative standardized anomalies during El Niño
events (Figs. 18a,c). The magnitude of the anomalies
from the gauge data is generally greater than those
from the multisatellite data (Fig. 18e) during El Niño
events.
In summary, the GPCP multisatellite data can be
used to monitor precipitation in West Africa for spatial
JANUARY 2008
203
LAMPTEY
FIG. 18. Spatial distribution of standardized anomalies of the composite of El Niño events from (a) gauge, (c) multisatellite, and (e)
gauge ⫺ satellite and La Niña events from (b) gauge, (d) multisatellite, and (f) gauge ⫺ satellite data.
distribution of mean annual and seasonal rainfall, inferences about latitudinal bands, zonal mean annual
rainfall over the study domain as a whole, and to some
degree interannual variability (i.e., the sign of the
anomaly but not the magnitude) until the more reliable
gauge data become available.
ganization, Rome, Italy; David Gochis, Andrea Hahmann, and Tom Warner of the National Center for
Atmospheric Research; and the three anonymous reviewers whose comments greatly improved the clarity
of this work.
REFERENCES
Acknowledgments. I thank George Huffman of the
Global Precipitation Climatology Project; Jurgen
Grieser formerly of the Global Precipitation Climatology Centre and now of the Food and Agricultural Or-
Adler, R. F., G. J. Huffman, A. Chang, R. Ferraro, 2003: The
version-2 Global Precipitation Climatology Project (GPCP)
monthly precipitation analysis (1979–present). J. Hydrometeor., 4, 1147–1167.
204
JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY
Ali, A., T. Lebel, and A. Amani, 2005a: Rainfall estimation in the
Sahel. Part I: Error function. J. Appl. Meteor., 44, 1691–1706.
——, A. Amani, A. Diedhiou, and T. Lebel, 2005b: Rainfall estimation in the Sahel. Part II: Evaluation of rain gauge networks in the CILSS countries and objective intercomparison
of rainfall products. J. Appl. Meteor., 44, 1707–1722.
Arkin, P. A., and P. E. Ardanuy, 1989: Estimating climatic-scale
precipitation from space: A review. J. Climate, 2, 1229–1238.
Ba, M. B., R. Frouin, and S. E. Nicholson, 1995: Satellite-derived
interannual variability of West African rainfall during 1983–
88. J. Appl. Meteor., 34, 411–431.
Barrett, E. C., and D. W. Martin, 1981: The Use of Satellite Data in
Rainfall Monitoring. Academic Press, 340 pp.
Beck, C., J. Grieser, and B. Rudolf, 2005: A new monthly precipitation climatology for the global land areas for the period
1951 to 2000. German Weather Service Climate Status Rep.
(Klimastatusbericht) 2004, 181–190.
Chapa, S. R., J. R. Milford, and G. Dugdale, 1993: Climatology of
satellite-derived cold cloud duration over sub-Saharan Africa. Proc. Ninth Meteosat Scientific Users’ Meeting, Locarno,
Switzerland, EUMESAT, 245–250.
Dai, A., I. Y. Fung, and A. D. Del Genio, 1997: Surface observed
global land precipitation variations during 1900–1988. J. Climate, 10, 2943–2962.
Giannini, A., R. Saravanan, and P. Chang, 2003: Oceanic forcing
of Sahel rainfall on interannual to interdecadal time scales.
Science, 302, 1027–1030.
Glickman, T., Ed., 2000: Glossary of Meteorology. 2nd ed. Amer.
Meteor. Soc., 855 pp.
Gregory, S., 1978: Statistical Methods and the Geographer. 4th ed.
Pearson, 256 pp.
Grimes, D. I. F., E. Pardo-Iguzquiza, and R. Bonifacio, 1999:
Optimal areal rainfall estimation using raingauges and satellite data. J. Hydrol., 222, 93–108.
——, E. Coppola, M. Verdecchia, and G. Visconti, 2003: A neural
network approach to real-time rainfall estimation for Africa
using satellite data. J. Hydrometeor., 4, 1119–1133.
Grist, J., S. E. Nicholson, and A. Mpolokang, 1997: On the use of
NDVI for estimating rainfall fields in the Kalahari of
Botswana. J. Arid Environ., 35, 195–214.
Hartmann, D. L., 1994: Global Physical Climatology. Academic
Press, 411 pp.
Hayward, D. F., and J. S. Oguntoyinbo, 1987: Climatology of West
Africa. Barnes & Noble, 271 pp.
Helsel, D. R., and R. M. Hirsch, 1991: Statistical Methods in Water
Resources: Techniques of Water-Resources Investigations.
Book 4, Hydrologic Analysis and Interpretation, United
States Geological Survey, 510 pp.
Huffman, G. J., R. F. Adler, B. Rudolf, U. Schneider, and P. R.
Keehn, 1995: Global precipitation estimates based on a technique for combining satellite-based estimates, rain gauge
analysis, and NWP model precipitation information. J. Climate, 8, 1284–1295.
——, and Coauthors, 1997: The Global Precipitation Climatology
Project (GPCP) combined precipitation dataset. Bull. Amer.
Meteor. Soc., 78, 5–20.
Hulme, M., 1995: Estimating global changes in precipitation.
Weather, 50, 34–42.
Karl, T. R., and R. W. Knight, 1998: Secular trends of precipitation amount, frequency, and intensity in the United States.
Bull. Amer. Meteor. Soc., 79, 231–241.
VOLUME 47
Kidd, C. K., 2001: Satellite rainfall climatology: A review. Int. J.
Climatol., 21, 1041–1066.
Kummerow, C., and Coauthors, 2000: The status of the Tropical
Rainfall Measuring Mission (TRMM) after two years in orbit. J. Appl. Meteor., 39, 1965–1982.
Laurent, H., T. Lebel, and J. Polcher, 1997: Rainfall variability in
Soudano–Sahelian Africa studied from rain gauges, satellite
and GCM. Preprints, 13th Conf. on Hydrology, Long Beach,
CA, Amer. Meteor. Soc., 17–20.
——, J. Jobard, and A. Toma, 1998a: Validation of satellite and
ground-based estimates of precipitation over the Sahel. Atmos. Res., 47–48, 651–670.
——, N. D’Amato, and T. Lebel, 1998b: How important is the
contribution of the mesoscale convective complexes to the
Sahelian rainfall? Phys. Chem. Earth, 23, 629–633.
Mathon, V., H. Laurent, and T. Lebel, 2002: Mesoscale convective
system rainfall in the Sahel. J. Appl. Meteor., 41, 1081–1092.
McCollum, J. R., and R. F. Ferraro, 2005: Microwave rainfall estimation over coasts. J. Atmos. Oceanic Technol., 22, 497–
512.
——, A. Gruber, and M. B. Ba, 2000: Discrepancy between
gauges and satellite estimates of rainfall in equatorial Africa.
J. Appl. Meteor., 39, 666–679.
Murphy, A. H., 1995: The coefficients of correlation and determination as measures of performance in forecast verification.
Wea. Forecasting, 10, 681–688.
New, M., M. Hulme, and P. Jones, 1999: Representing twentiethcentury space–time climate variability. Part I: Development
of 1961–90 mean monthly terrestrial climatology. J. Climate,
12, 829–856.
——, ——, and ——, 2000: Representing twentieth-century
space–time climate variability. Part II: Development of 1901–
96 monthly grids of terrestrial surface climate. J. Climate, 13,
2217–2238.
Nicholson, S. E., 1980: The nature of rainfall fluctuations in subtropical West Africa. Mon. Wea. Rev., 108, 473–487.
——, 1994: On the use of the normalized difference vegetation
index as an indicator of rainfall. Global Precipitations and
Climate Change, M. Desbois and F. Désalmand, Eds.,
Springer-Verlag.
——, 2005: On the question of the “recovery” of the rains in the
West African Sahel. J. Arid Environ., 63, 615–641.
——, and Coauthors, 2003: Validation of TRMM and other rainfall estimates with a high-density gauge dataset for West Africa. Part II: Validation of TRMM rainfall products. J. Appl.
Meteor., 42, 1355–1368.
Press, W. H., S. A. Teukolsky, W. T. Vetterling, and B. P. Plannery, 1996: Numerical Recipes in Fortran 90: The Art of Parallel Scientific Computing. 2nd ed. Vol. 2, Fortran Numerical
Recipes, Cambridge University Press, 576 pp.
Rudolf, B., H. Hauschild, W. Ruth, and U. Schneider, 1994: Terrestrial precipitation analysis: Operational method and required density of point measurements. Global Precipitation
and Climate Change, M. Dubois and F. Désalmand, Eds.,
Springer-Verlag, 173–186.
Sealy, A., G. S. Jenkins, and S. C. Walford, 2003: Seasonal/
regional comparisons of rain characteristics in West Africa
using TRMM observations. J. Geophys. Res., 108, 4306,
doi:10.1029/2002JD002667.
Siegel, S., and N. J. Castellan Jr., 1988: Nonparametric Statistics
for the Behavioral Sciences. 2nd ed. McGraw-Hill, 399 pp.
JANUARY 2008
LAMPTEY
Snijders, F. L., 1991: Rainfall monitoring based on Meteosat
data—A comparison of techniques applied to the western
Sahel. Int. J. Remote Sens., 12, 1331–1347.
Thorne, V., P. Coakeley, D. Grimes, and G. Dugdale, 2001: Comparison of TAMSAT and CPC rainfall estimates with raingauges, for southern Africa. Int. J. Remote Sens., 22, 1951–
1974.
Trenberth, K. E., 1997: The definition of El Niño. Bull. Amer.
Meteor. Soc., 78, 2771–2777.
Wilks, D. S., 2006: Statistical Methods in the Atmospheric Sciences.
2nd ed. Academic Press, 627 pp.
Xie, P., and P. A. Arkin, 1995: An intercomparison of gauge ob-
205
servations and satellite estimates of monthly precipitation. J.
Appl. Meteor., 34, 1143–1160.
——, and ——, 1997: Global precipitation: A 17-year monthly
analysis based on gauge observations, satellite estimates, and
numerical model outputs. Bull. Amer. Meteor. Soc., 78, 2539–
2558.
Yin, X., A. Gruber, and P. Arkin, 2004: Comparison of the GPCP
and CMAP merged gauge–satellite monthly precipitation
products for the period 1979–2001. J. Hydrometeor., 5, 1207–
1222.
Zeng, N., 2003: Drought in the Sahel. Science, 302, 999–1000.