Sie sind auf Seite 1von 14

The Comparison of Two Merged Rain GaugeSatellite Precipitation Datasets

Arnold Gruber,* Xiujuan Su,+ M. Kanamitsu,+ and J. Schemm+

ABSTRACT
Two large-scale precipitation datasets, one produced by the Global Precipitation Climatology Project (GPCP) and the other by the Climate Prediction Center of the National Weather Service, and called Climate Prediction Center Merged Analysis of Precipitation (CMAP), were compared. Both datasets blend satellite and gauge estimates of precipitation. And while the latter has its heritage in the GPCP, different analysis procedures and some additional types of input data used by CMAP yielded different values. This study used the error characteristics of the data to assess the significance of the observed differences. Despite good spatial and temporal correlations between the two fields some of the observed differences were significant at the 95% level. These were traced to the use of some different input data such as the use by CMAP of atoll gauges in the tropical Pacific and gauges uncorrected for wetting evaporation and aerodynamic effects. The former impacts the tropical ocean rain amounts and the latter is particularly noticeable in the Northern Hemisphere land areas. Also, the application of these datasets to the validation of atmospheric general circulation models is discussed.

1. Introduction The Global Precipitation Climatology Project (GPCP) produced monthly mean precipitation data on a global 2.5 2.5 latitudelongitude grid for the period of July 1987December 1998 (now available) by combining gauge measurement and satellite estimates (Huffman et al. 1997). Independently, Xie and Arkin (1997) of the National Weather Service, Climate Prediction Center, developed a merged satellitegauge dataset with the same grid resolution. This dataset, named for the Climate Prediction Center Merged Analysis of Precipitation (CMAP) has its heritage in the GPCP in that it shares some of the same input data. CMAP, however, utilizes a different analysis procedure and adds other satellite data. Since both datasets

*NOAA/NESDIS Office of Research and Applications, Camp Springs, Maryland. + NOAA/NWS, National Centers for Environmental Prediction, Camp Springs, Maryland. Corresponding author address: Dr. Arnold Gruber, NOAA/ NESDIS, Office of Research and Applications, Camp Springs, MD 20746-4304. In final form 21 April 2000.

will ultimately be used in a variety of scientific studies such as seasonal to interannual variability studies and model validation (e.g., Kondragunta and Gruber 1997; Janowiak et al. 1998), it is important to understand the differences between them. This paper represents such an attempt emphasizing the large time- and space scales and using the error characteristics of the monthly mean rainfall estimates as reported by the developers to assess the significance of any differences. The study will compare the period July 1987 December 1998, which is a common period between the two datasets. It should be noted, however, that the datasets are continually being extended and that CMAP actually begins in 1979. Also, while the two datasets share much of the same input data, the inclusion of additional satellite data and the different analysis methodology used by CMAP means that we may not be able to unequivocally identify the cause of the differences in all cases. Nevertheless we hope that one of the results of the comparison may give some feedback for improving the methodology in merging gauge and satellite measurements. Keeping these goals in mind, the paper focuses on the spatial and temporal distribution comparison and causes of the difference between the two datasets.
2631

Bulletin of the American Meteorological Society

2. The data and merge methods in GPCP and CMAP a. Data The gauge and satellite estimates used in GPCPand CMAP-merged precipitation datasets are summarized in Table 1 adapted in part from Janowiak et al. (1998). The gauge data that was used by both GPCP and CMAP from 1987 to 1998 are from the Global Precipitation Climatology Center (GPCC; Rudolf et al. 1996). The GPCC collected 6700 rain gauge stations worldwide, subjected them to a rigid quality control, and prepared rainfall analyses on a 2.5 2.5 latitude longitude grid. However, the GPCP uses the version of the GPCC gauge analysis that is corrected for systematic errors caused by wetting, evaporation, and aerodynamic effects, while CMAP used a version of the analyses that is uncorrected. The climatological estimate of systematic error is from Legates, which is based on Sevruks method (Huffman et al. 1997). The

correction factor is between 1 and 1.5 and between 60S and 60N for most areas. In addition, the CMAPmerged analysis uses more than 100 gauges located on atolls and small islands (Morrissey and Greene 1991) for correction of merged satellite precipitation estimates over the tropical ocean (Xie and Arkin 1997), but the GPCP does not. The impacts of those 100 stations and the use of uncorrected gauge data will be discussed later. The common satellite precipitation estimates used by both GPCP and CMAP are based on the Geostationary Operational Environmental Satellite (GOES) operated by the United States; the Geostationary Meteorological Satellite operated by Japan; the Meteorological Satellite, Meteosat, operated by the European Community; and the National Oceanic and Atmospheric Administrations (NOAA) polar-orbiter infrared (IR) measurements in the region 40N40S, and microwave scattering and emission measurements from the Special Sensor Microwaver/Imager (SSM/I)

TABLE 1. Some characteristics of the satellite and rain gauge data used by the GPCP and CMAP analyses. Sensor Geostationary infrared (GPI) Polar infrared Sampling Eight times per day Coverage 40N40S (land/ocean) 40N40S (land/ocean) 60N60S (ocean only) 60N60S (land only) 60N60S (land, ocean) 90N90S (land/ocean) 60N60S (ocean only) Global land Usage GPCP CMAP GPCP CMAP GPCP CMAP GPCP References Arkin and Meisner (1987)

Four times per day

Janowiak and Arkin (1991)

SSM/I microwave emission SSM/I microwave scattering SSM/I microwave scattering OLR Precipitation Index (OPI) Microwave Sounding Unit (MSU) Rain gauge

Up to two times per day Up to two times per day Up to two times per day Four times per day

Wilheit et al. (1991)

Ferraro et al. (1996); Ferraro and Marks (1995) Ferraro et al. (1996); Ferraro and Marks (1995) Xie and Arkin (1998)

CMAP

CMAP

Four times per day

CMAP

Spencer (1993)

Continuous at stations

GPCP (corrected) CMAP (uncorrected) CMAP

Rudolf et al. (1996)

Rain gauge

Continuous at stations

Tropical atolls

Morrissey and Greene (1991)

2632

Vol. 81, No. 11, November 2000

on the polar-orbiting Defense Meteorological Satellite Program satellites. The emission estimates are from Wilheit et al. (1991) and are available over oceans only. The scattering estimates are from Ferraro et al. (1996) and Ferraro and Marks (1995) and are available over land and ocean; however, they are used by the GPCP over land only, while CMAP uses the scattering estimates over both land and ocean. The geostationary infrared precipitation estimates are obtained from the GOES Precipitation Index (GPI; Arkin and Meisner 1987), which has been adapted for use on all the geostationary satellites. It is applicable only between 40N and 40S where it is most reliable. This technique is also applied to histograms of the polarorbiting data or when histograms are not available to GPI-calibrated estimates of the outgoing longwave radiation (OLR). Each of the satellite estimates used in the GPCP and CMAP analyses has its advantages and disadvantages. For example, the geostationary GPI estimates, which are based on cloud-top temperatures equal to or colder than fixed threshold temperature (235 K), have the most frequent sampling, but may suffer from inaccuracy in estimating precipitation amounts. Clearly this is the case for rain clouds with tops warmer than 235 K and nonraining clouds with cold top clouds such as cirrus. Microwave estimates from the SSM/I are more closely related to precipitation processes in clouds than the IR and provide more accurate instantaneous rain estimates but are available only up to two times per day for most areas. Infrared polar satellite data are used only when the GPI is not available. In addition to using SSM/I-scattering estimates over the ocean, the CMAP-merged analysis uses Microwave Sounding Unit (MSU)-based precipitation estimates from Spencer (1993) and OLR-based precipitation index (OPI) data from Xie and Arkin (1998). b. Merging methods In general, rain gauge measurements provide relatively accurate point estimates of precipitation but suffer from sampling errors in representing area means and are not available over most oceans and remote areas. Satellite measurements can cover most of the globe; however, they suffer from errors due to lack of a direct relationship between observation parameters and precipitation. The major goal of merging gauge measurements and satellite estimates is to retain each datasets advantage while reducing overall errors. As described above, the GPCP and CMAP analyses use some common datasets; however, the merging procedures are quite different.
Bulletin of the American Meteorological Society

Huffman et al. (1995, 1997) described the GPCP merging procedure. The first step is to merge the SSM/I microwave emission estimates used over water with the SSM/I microwave scattering estimates used over land. Then the magnitudes of IR-based estimates obtained using the GPI (Arkin and Meisner 1987), which is used for all the geostationary IR satellites, are adjusted by the ratio of SSM/I- and GPIbased estimates that are roughly coincident in space and time. Polar-orbiting IR estimates are adjusted using a smoothly varying interpolation of the SSM/IGPI adjustment ratio where the GPI is not available. Adjusted GPI and polar IR estimates are limited to the area between 40N and 40S. The SSM/I estimates are used alone outside the 40N40S region. The combination of adjusted GPI with SSM/I estimates forms a multisatellite estimate. Over land the bias of the multisatellite estimate is adjusted according to the largescale (five grid boxes five grid boxes) average gauge analysis. Random errors are estimated for each dataset: multisatellite and gauge estimate. The final step is to merge multisatellite estimates and gauge estimates using a maximum likelihood estimate in which both values are combined with inverse error-variance weighting. The CMAP merging analysis is summarized in Xie and Arkin (1996, 1997). The first step is to produce a base period of merged data. The GPI, SSM/I-scattering, and the SSM/I emission precipitation estimates; the OPI; and the ocean precipitation estimates based on the MSU are combined in the first step by using a maximum likelihood estimate in which the weighting coefficients are inversely proportional to the squares of the individual random errors. The errors of each satellite estimate are determined by comparing with GPCC gauge measurements over the land and with atoll gauge measurements (Morrissey and Greene 1991) for an estimate of the errors over the ocean. The second step is the removal of possible biases by combining the results of the first step with gauge analyses. Over the land the blend with gauge measurements uses the methods developed by Reynolds (1988). In this blending method, the assumption is made that the combined satellite estimate can represent the structure of precipitation distribution and there is no bias in gauge estimates. The structure of precipitation can be described by a Poisson equation in which the boundary conditions are determined by gauge estimates. Then the merged base products are obtained by solving the Poisson equation. Over the ocean the bias removal is accomplished by comparing with atoll gauges over the Tropics and by subjective assumptions about
2633

the bias structure in the extratropics. Since the atolls are in the tropical western Pacific it is recognized that the oceanic bias adjustment may differ in other parts of the oceans (Xie and Arkin 1997). 3. The comparison of the GPCP and the CMAP datasets and discussion In the comparisons to follow we have tried to assess the statistical significance of some of the differences by using the random error estimates provided in the datasets. The procedures for estimating the number of degrees of freedom of monthly mean precipitation estimates that are further averaged in space and time were adapted from Janowiak et al. (1998). That procedure calculates the number of independent samples of space- and time-averaged data by accounting for the serial correlation typically found for monthly mean precipitation estimates. This information is used to perform a t test of significance of the differences at the 95% confidence level. The null hypothesis that the differences are not significantly different from zero at the 95% level is given as
2 1 2 P + 2 , 1 P 2 t 0 . 95 N i1 N i 2 12

subtropical subsidence regions of the South Pacific along the coast of Peru and Chile, the South Atlantic between Brazil and South Africa, and the Sahara Desert. The differences in the annual means are shown in Fig. 2; the upper panel is GPCP CMAP and the lower panel is the percent difference. Looking at the upper panel it is seen that the differences are mostly between 0.5 mm day1 over most of the domain especially over land where gauge measurements play an important role in determining the final magnitude of the merged precipitation estimates. Notable exceptions are in the equatorial Tropics where CMAP is 0.5 2.5 mm day1 higher than GPCP and in the North Atlantic storm tracks where GPCP is higher than CMAP by up to 2.5 mm day1. The relative difference (lower panel) provides a more sensitive measure of the differences. It is calculated as the difference between the GPCP and CMAP estimates divided by their average,

(1)

where P1 and P2 are the sample mean, 1 and 2 are the mean standard errors, and Ni1 and Ni2 are the independent numbers of sample 1 and 2, respectively. For large samples t0.95 1.96 at the 95% level. a. Spatial pattern The spatial pattern of the GPCP and the CMAP annual mean precipitation for the period January 1988December 1998 (Fig. 1) are nearly identical over the domain 60N60S. Indeed the spatial correlation between them is 0.92. Both datasets depict the intense precipitation in the intertropical convergence zone (ITCZ) over the Indian, Pacific, and Atlantic Oceans, the South Pacific convergence zone (SPCZ), the South Atlantic convergence zone, the primary storm tracks over the North Pacific Ocean along the Asian coast and North Atlantic Ocean along the North American coast, and the active convective areas over the continents of South America, Africa, and Southeast Asia. Also, both portray the dry areas with annual mean precipitation less than 0.5 mm day1 in the
2634

FIG. 1. Annual mean precipitation in mm day1 for the period Jan 1988Dec 1998 for (top) CMAP and (bottom) GPCP.

Vol. 81, No. 11, November 2000

FIG. 3. The relative difference between corrected gauge (CG) and uncorrected gauge (UC) analyses from the GPCC. The relative difference is (CG UG)/(CG + UG)/2 and is expressed in percent.

FIG. 2. Difference maps of annual mean precipitation: (top) GPCP CMAP in mm day1; (bottom) the relative difference, (GPCP CMAP)/(CMAP + GPCP)/2, in percent.

(GPCP CMAP)/(GPCP + CMAP)/2. Note that in the very dry areas such as the subtropical dry zones and deserts it is not very meaningful since small differences have a relatively large signal. The distribution of the relative difference is similar to the upper panel; however, an important feature that now shows more clearly is the large positive differences of about 10%50% over central North America and Europe and Asia in mid- and high latitudes. Since both merge procedures utilize gauge data in setting the magnitude, this broad pattern might be considered puzzling except that the GPCP uses the corrected gauge data while the CMAP does not, as mentioned earlier. Since the corrections always increase the rainfall the larger values for GPCP are quite reasonable and in fact agree very well with a map of the relative differences of the corrected and uncorrected gauges (Fig. 3). Also noticeable in the relative difference field are large negative differences along coastal areasmost noticeably the west coastal oceans of South America, Africa, and North America. There are several factors that may influence these differences. First, in ocean
Bulletin of the American Meteorological Society

coastal areas the GPCP assigns the gauge analyses proportionally to the area of ocean and land in the grid box whereas CMAP considers the grid as either water or land depending on a threshold value. Second, the GPCP uses only SSM/I emission estimates over the oceans with a fairly broad boundary near the coast in order to avoid land contamination in the estimate. CMAP uses both SSM/I scattering and emission over oceans and is thus less influenced by land/water boundaries. Finally over land areas another factor that may contribute to the observed differences is the methodology used when there is no gauge measurement in a grid box. In the GPCP analysis, the precipitation values in the grid boxes where no gauge measurements are available are determined by interpolating the gauge measurements from surrounding areas. In the

FIG. 4. The frequency of disagreement (%) between CMAP and GPCP. The fields are considered in disagreement if they exceeded 1.96 times the average standard errors of monthly precipitation (the 95% confidence interval). 2635

CMAP-merged analysis, the values of precipitation are determined by the modified satellite estimates. This difference in analysis methods can contribute to some of the observed differences. One example is over central equatorial Africa where the relative difference approaches 50%. Recently, McCollum et al. (2000) examined the difference between the GPCP-merged

and multisatellite estimates (GPI and SSM/I) in equatorial central Africa and demonstrated that both GPI and SSM/I overestimate the monthly precipitation in central Africa relative to the gauge and the GPCPmerged analysis. This is an area where the gauge density is very low. The GPCP-merged analysis, which utilizes surrounding gauge information from a large area, is more influenced by the gauges whereas the CMAP is more influenced by the satellite estimates in areas with low-gauge density. This may be why CMAP is up to 50% higher than GPCP in that area. Clearly this affects primarily land areas and may not be a factor in coastal regions except in mixed land/ water grids. Over the oceans the differences may be explained in large measure by the different data sources used in each climatology. In the Tropics the biggest influence comes from the use of the atoll gauge dataset that is used to adjust the final values in CMAP but is not used in GPCP because of the very low density of gauges. The adjustment used in CMAP is applied throughout the Tropics with decreasing weight with latitude phasing out to zero at 40N,S. GPCP uses only the GPI and SSM/I emission estimates and further adjusts the IR to the SSM/I estimates. Studies have shown that all satellite estimates are lower than the atoll gauge measurements (Xie and Arkin 1995). Thus, precipitation from CMAP with the atoll gauge correction probably results in higher values than from GPCP over the tropical oceans. At higher latitudes over the oceans (4060S, 4060N), CMAP merges SSM/I scattering and SSM/I emission while GPCP depends only on SSM/I emission estimates. The SSM/I emission precipitation estimates are almost 6 times higher than SSM/I-scattering estimates according to Xie and Arkin (1995). This may explain in part why the precipitation from CMAP is less than from GPCP there. FREQUENCY OF DISAGREEMENT The previous discussion looked at the average differences over the entire period. Clearly as one averages over space and time the random error reduces significantly and the differences can become significant, as will be shown later. However, it is of interest to see how well individual monthly mean estimates agree or disagree at each grid point. We have utilized the error estimates provided in the datasets to calculate the frequency of disagreement between the two datasets at each grid point for the period 198897. The estimates were judged to disagree if the monthly mean
Vol. 81, No. 11, November 2000

FIG. 5. Zonal average profiles of annual mean precipitation: (top) CMAP and GPCP (CMAP is solid, GPCP dashed), their difference; (middle) CMAP GPCP; and (bottom) significant differences at the 95% level, shaded black. Units are in mm day1. 2636

difference between them exceeded 1.96 times the average standard errors of monthly precipitation at a grid point from both datasets. This is the range of values for the 95% confidence interval for the null hypothesis that these two datasets are from the same population. If it exceeds this interval the null hypothesis is rejected and disagreement is significant. This is based on the assumption that the monthly precipitation estimates fall in a normal distribution. Although the distribution of the population from which the monthly precipitation is calculated is not known (daily or hourly precipitation), it is assumed normal based on the central limit theorem. Another assumption related to the test is that both monthly precipitation estimates are unbiased and the difference between FIG. 6. Average seasonal cycle for land, ocean, and combined land and ocean them is caused by random errors, esti- for the Northern Hemisphere, 060N, and the Southern Hemisphere, 060S, mated by Huffman et al. (1997) and Xie and 60N60S. The middle group of low amplitude lines is the 60N60S avland, ocean, and combined. CMAP is solid, GPCP is dashed. and Arkin (1996). However, as discussed erage for each month: Units are in mm day1. by Janowiak et al. (1998) these assumptions may not always be assured. Thus, the confidence limits should not be considered abso- b. Temporal variability lute but the test is a useful diagnostic that can provide The mean annual cycle of both CMAP and GPCP insights into the comparison of the datasets. are displayed in Fig. 6 for land and ocean for the NorthThe spatial distribution of the frequency of dis- ern Hemisphere (060N), Southern Hemisphere (0 agreement is shown in Fig. 4. Over the study period, 60S), and the globe (60N60S). For a more the frequencies of disagreement between two datasets complete assessment of the significance of the differare less than 10% over most of the area between 60N ences, reference should be made to Table 2. There the and 60S. Over the equatorial tropical oceans frequen- differences and their 95% confidence intervals cies as high as 50%60% are observed in isolated ar- grouped by season for land and ocean for each hemieas, with the values typically being 10%30% in areas sphere and for the globe are displayed. The seasons are where the mean differences (Fig. 2) were large. The DecemberJanuaryFebruary (DJF), MarchApril largest frequency of disagreement between the two May (MAM), JuneJulyAugust (JJA), and Septemdatasets occurs over the coast line of northern South berOctoberNovember (SON). If the magnitude of America, Southeast Asia, Africa, and over the north the confidence interval is less than the magnitude of European land (20%80%) with the higher values oc- the difference the differences are judged to be significurring at isolated locations. cant at the 95% level. Looking first at the time series This pattern is reflected in the profile of zonal av- of the land areas (Fig. 6) it is seen that both CMAP erages of the annual mean values (Fig. 5). The zonal and GPCP show nearly identical seasonal variability means are in the upper panel where it is seen in the in each hemisphere, as should be expected in view of region 15S to about 20N that the GPCP is less than the gauge influence. An interesting feature is the largeCMAP and in the region from 35N,S to about 60N,S amplitude variation in the Southern Hemisphere, but GPCP is greater than CMAP. The largest difference, because of its limited area the 60N60S average about 1 mm day1, is located at 5N. The lower panel (plotted in the figure and distinguished by its low shows in black where the differences are significant amplitude) is dominated by the Northern Hemisphere at the 95% confidence level. In between the differences values. It is also seen that GPCP tends to be slightly are statistically negligible. higher than CMAP, especially in the Northern HemiBulletin of the American Meteorological Society
2637

sphere, which is consistent with our previous discussion on the use of corrected and uncorrected gauges. Looking at Table 2 it is interesting to note that the significant differences between GPCP and CMAP are during the JJA and SON season in the Northern Hemisphere and for the 60N60S area and for the annual average 060N and 60N60S. The oceans present a somewhat different picture. The variabilities between the two datasets are quite similar however; as seen in Table 2 there are significant magnitude differences between them in the annual averages and all seasons except for SON in the Southern Hemisphere and 60N60S in DJF. Further note that in the Northern Hemisphere CMAP is greater than GPCP from April to November and less from about December to March, undoubtedly the influence of the tropical differences during the summer months and the storm track differences in the winter months. The Southern Hemisphere shows a reverse difference to the Northern Hemisphere, that is, CMAP greater than GPCP in JanuaryApril and slightly less than GPCP JuneOctober, but not as prominently. The combined land and ocean for 60N60S show vanishing small differences with the MAM and JJA seasons exhibiting a barely statistically significant difference.
FIG. 7. Temporal correlation between CMAP and GPCP for (top) monthly precipitation and (bottom) anomaly fields. The anomalies are calculated based on the period Jul 1987Dec 1998 for both datasets with the annual cycle removed. Units are in mm day1.

c. Temporal correlations A good way to examine how the two datasets agree in time is to look at the temporal correlations between them at each grid point. We have calculated correlations for both the mean fields and the anomaly fields. The anomalies are calculated for each month of the

TABLE 2. The mean precipitation difference (mm day1) between CMAP and GPCP (CMAP GPCP) and the 95% confidence intervals for the period 198794. The entries are stratified by region: 060N, 060S, 60N60S, and for land and ocean separately and combined and for seasons as defined in the text. Differences judged to be significant at the 95% level are in bold type. Latitude range 060S 060N 60S60N 060S 060N 6060N 60N60S DJF 0.19 0.30 0.11 0.31 0.10 0.25 0.12 0.06 0.09 0.07 0.03 0.05 0.01 0.06 MAM 0.15 0.24 0.10 0.17 0.11 0.14 0.11 0.13 0.11 0.05 0.07 0.05 JJA 0.03 0.14 0.15 0.10 0.12 0.08 0.07 0.04 0.40 0.10 0.12 0.05 0.06 0.05 SON 0.03 0.22 0.13 0.10 0.10 0.09 0.03 0.04 0.28 0.09 0.09 0.04 0.04 0.05 Annual 0.10 0.12 0.11 0.03 0.18 0.08 0.12 0.10 0.08 0.02 0.04 0.03

Land

Ocean

Land and ocean

0.06 0.05

0.03 0.03

2638

Vol. 81, No. 11, November 2000

dataset (annual cycle removed) for the period July 1987December 1998 for both the GPCP and CMAP datasets. They are shown in Fig. 7. For the mean fields (upper panel) correlation coefficients are greater than 0.9 for most regions over land where the gauge data plays a dominant role and for the tropical oceans where satellite estimates are more accurate. Correlation coefficients less than 0.5 are evident for small areas in the southeast Pacific and Atlantic subsidence region and the Sahara Desert where the monthly precipitation is less than 0.5 mm day1. The anomaly correlations show some striking differences. There is considerably less area of greater than 0.9 correlation than for the mean fields, most notably over the tropical and subtropical FIG. 8. Comparison of CMAP and GPCP time series of monthly mean precipiland areas and in the high-precipitation tation over land for 20-wide latitude belts from 60N to 60S. CMAP is solid, areas of the ITCZ, SPCZ, and the Indian GPCP is dashed. A single solid line indicates overlap. Units are mm day1. Ocean, although those areas still exhibit correlations greater than 0.7. To further examine the time series of these two values of CMAP, as seen on the average maps, are datasets, we examined time series of mean precipita- evident in the 020N,S regions where the differtion and monthly anomaly for six zones; 020N,S; ences are largest in the respective summer months 2040N,S; and 4060N,S, over land and ocean with better agreement in the winter season. In the separately. Figures 8 and 9 contain the time series of ocean zone 4060N,S the GPCP is systematically monthly means for land and ocean respectively, and Figs. 10 and 11 show the anomalies for land and ocean, respectively. The monthly mean and anomaly correlations of the time series of CMAP and GPCP are summarized in Table 3. For the monthly mean the two datasets fit each other better over land than over ocean although the correlations over both land and ocean are mostly greater than 0.9, the exception being the ocean areas 4060N and 4060S. The best fit occurs over the land areas of 0 40N,S (correlation coefficients are between 0.9960.999), and the poorest fit occurs over 4060S ocean (correlation coefficient 0.680). The largest difference (< 0.5 mm day1) between two time series over land occurs between 4060N in each winter, which further confirms the influence of the corrected and uncorFIG. 9. Same as Fig. 8 but for ocean areas. rected gauge data. Over the oceans high
Bulletin of the American Meteorological Society
2639

FIG. 10. Same as Fig. 8 but for anomalies over land.

anomaly correlation is less than 0.7 in the three Northern Hemisphere zones and is somewhat higher in the Southern Hemisphere. This suggests that the two datasets may exhibit some differences in the oceanic anomaly structures, although it may be somewhat subtle since the correlations are relatively high. In fact we compared the anomaly structures for the major El Nio warm event of 1997/98 (Fig. 12) and indeed find for all practical purposes there is no significant difference in the anomaly pattern and only subtle differences in magnitude. This was also reflected in anomaly time series (not shown) over Nio-4 (5N5S, 160E150W) and Nio-3.4 (5N5S, 170120W), which exhibited excellent agreement depicting not only the 1997/98 warm episode but also the 1991/92 warm and 1988/89 and 1995/96 cold events, with only minor differences in magnitude. 4. Validation of GCM precipitation outputs One important use of these climate-scale estimates of precipitation is to validate the tropical precipitation produced by atmospheric general circulation models (AGCMs). It has been known that although the long-range predictability is the greatest in the Tropics (Shukla and Fennessy 1988), AGCMs in general do not simulate precipitation very well because of imperfectness in the model, particularly the convective parameterization (Sperber and Palmer 1995). Improvements in tropical precipitation is an imperative task of numerical modelers for providing credible guidance for short- and seasonal-scale weather predictions. In this study, the CMAP and GPCP precipitation analyses have been utilized to verify tropical precipitation in two atmospheric general circulation model simulations and in the National Centers for Environmental Prediction National Center for Atmospheric Research (NCEPNCAR) reanalysis. As was discussed earlier, these estimates
Vol. 81, No. 11, November 2000

higher than CMAP in most of the months. There are small or no differences between CMAP and GPCP over the 2040N,S ocean. For the monthly anomaly, the correlations between the GPCP and the CMAP resemble those of the monthly precipitation time series in that they are higher over land than over oceans. However, over the oceans the

FIG. 11. Same as Fig. 10 but for anomalies over ocean. 2640

while capturing the mean structures quite well have somewhat lower anomaly correlation with each other in oceanic areas. In an attempt to utilize these datasets for validating the large-scale precipitation structure of AGCMs, both analyses and AGCM-simulated precipitation were filtered to capture only large-scale features. The filtering of the small scale is a well-known procedure for verification of forecasts, since simulation of the large scale is the primary interest of modelers (and simulation of the small scale is much more difficult). The spectral smoothing of triangular wavenumber 20 (T20) is commonly used to verify 500-hPa geopotential height field for comparison of various short- and medium-range forecast models. The validation of forecasts using raw data produces misleading results since errors in small-scale features contaminate the verification of large scales. Following the common procedures used in geopotential height verifications, and considering that the precipitation has larger spatial variability, the following smoothing procedures are employed to the precipitation field. First, the logarithm of base 2 was applied to precipitation amount to temper spatial gradients and then spectral truncation was imposed. Figure 13 illustrates an example of a filtered precipitation field compared to the original. It demonstrates that the logarithmic manipulation does not distort the spatial distribution and the spectral truncation retains the large-scale features without abnormal expansion of precipitation area. When the same filter was applied to the two analyses, the difference between them diminishes considerably and they correlate much better to each other. This is shown in Fig. 14, which compares the weighted anomaly correlation (AC) between the two analyses over the Tropics (20S20N) for the 198897 period. The average AC over the period increases to 0.94 from 0.86 with filtered analyses. The variability of the AC is much smaller for filtered analyses. In computing weighted AC, the anomalies are weighted proportional

FIG. 12. Depiction of the precipitation anomaly for El Nio conditions during Dec 19971998: (top) GPCP and (bottom) CMAP. Units are mm day1.

to the inverse of the climatological root-mean-square difference (rmsd) between the filtered analyses. This procedure is introduced to maximize the AC between the two analyses by incorporating uncertainties in measurements. Therefore, AC of 0.94 might be the upper bound of agreement between the two analyses over the Tropics.

TABLE 3. The correlation between the GPCP and the CMAP for the indicated latitude zones for anomaly and monthly mean precipitation and separated by land and ocean. 020N Monthly Land Ocean Land Ocean 0.999 0.964 0.967 0.680 2040N 0.997 0.905 0.969 0.672 4060N 0.980 0.865 0.978 0.631 020S 0.996 0.976 0.946 0.817 2040S 0.998 0.922 0.987 0.824 4060S 0.958 0.680 0.823 0.708

Anomaly

Bulletin of the American Meteorological Society

2641

98 period and that of ECHAM was for a much longer period of 195094. The observed sea surface temperature (SST) was given as the lower boundary condition for both simulations. The simulated precipitation underwent the same filtering procedure as in the analyses. Because of the longer overlapping data period, the skill estimates of the AGCM simulations shown in this paper are with the CMAP analysis. Both models exhibit the same level of skill in simulating the tropical precipitation despite several differences in the model configurations. The mean AC over the 197994 period is 0.19 for MRF and 0.21 for ECHAM (Figs. 15 and 16). Mean AC scores using unfiltered FIG. 13. Comparison of (left) original and (right) filtered precipitation for Jun simulation and analysis are 0.010.02 1998 from the CMAP analysis. (top) The total precipitation and (bottom) anoma- lower as expected. Close inspection of lous precipitation, mm day1. the figures reveals that the AC scores are high in some years when strong El Nio/ The filtered analyses were used to evaluate the La Nia events are observed (e.g., years 1983, 1987, tropical precipitation simulations with two AGCMs. 1989, and 1992). The filtered scores for those events One is an experimental version of the Medium Range are much higher than the unfiltered ones, and compariForecast (MRF) AGCM of NCEP and the other is son of the two models becomes much more meaningECHAM AGCM of the Max Plank Institute, Ham- ful with these scores. In comparison, the mean AC of burg, Germany. Both of the AGCM simulations were the NCEPNCAR reanalysis (Kalnay et al. 1996) preperformed in the horizontal spatial resolution of T40. cipitation is at the level of 0.54 for the same period The simulation with the NCEP MRF was for the 1979 (Fig. 17), and the score with unfiltered data is nearly

FIG. 14. Anomaly correlation coefficient between the CMAP and GPCP analyses over the Tropics (20N20S). Solid line corresponds to AC with the original analyses and dashed line to the filtered analyses. Solid and dashed straight lines indicate average AC for the 198897 period.

FIG. 15. Tropical (20N20S) anomaly correlation coefficients of the MRF GCM simulations against the CMAP analysis. Solid line corresponds to AC between the original analysis and simulation and the dashed line to that of the filtered analysis and simulation. Solid and dashed straight lines indicate average AC for the 197994 period.

2642

Vol. 81, No. 11, November 2000

FIG. 16. Same as Fig. 15 but for the ECHAM GCM.

FIG. 17. Same as Fig. 15 but for reanalysis vs CMAP.

0.1 lower. The tendency for the higher scores to be magnified with filtered data is also apparent. The same AC scores were computed using GPCP analysis (for shorter period) and comparable skill levels for both AGCMs and reanalysis precipitation were observed (not shown). These examples indicate that smoothed analysis provides excellent means to compare skill of different models, but model simulation itself is still far from perfect and the choice of analysis is still not an important issue for model verifications. 5. Conclusions The two observation datasets, the GPCP1 and the CMAP, derived from combining several satellite estimates and gauge measurements are compared. Both datasets are compatible in overall space and temporal pattern and in depicting important events such as the El NioSouthern Oscillation (ENSO) episodes. The intensity patterns of annual global distribution from both datasets match each other, as evidenced by the mean spatial correlation of annual global distribution being 0.92. The position of important intensive pre-

The GPCP and CMAP data are available online. For GPCP the address is http://www.ncdc.noaa.gov/pub/data/gpcp/. The complete merged dataset as well as individual components as well as some experimental datasets are available. The CMAP data can be obtained via anonymous ftp: ftp.ncep.noaa.gov cdpub/precip/ cmap.

cipitation zones as the intertropical convergence zone, the South Pacific convergence zone, the South Atlantic convergence zone, and storm tracks generally coincide with each other. The seasonal variations from both datasets have similar phase and amplitude. The correlation coefficients of monthly precipitation are mostly greater than 0.90 except in the 4060S oceans. Anomaly correlations are lower, especially over oceans where they range from 0.63 to 0.82. Nevertheless the spatial pattern of cold and warm ENSO episodes agree quite well. However there exist differences in magnitude in mean values, global distribution, and zonal average time series. The mean values over land from GPCP are greater than from CMAP and which is larger over other areas depends on the seasons of the year. The magnitude of the difference of global distribution tends to be larger for more intensive precipitation. The variation of magnitude differences between monthly precipitation of the two datasets causes the lower correlation between the monthly anomaly. The major difference between the two datasets over land is caused by the different treatment of the gauge measurements. Over tropical oceans the incorporation of the tropical atoll data by CMAP in their analysis is the main cause of the differences between CMAP and GPCP, which does not use the atoll data. Over higher-latitude oceans the differences may be caused in part by the different satellite input data, as previously discussed. Clearly, over the oceans, where there are essentially no gauges to anchor the magnitudes, the estimates are very sensitive to the satellite input data or even scattered gauge data such as available from atolls.
2643

Bulletin of the American Meteorological Society

A filtering technique that retains the large-scale characteristics of the precipitation fields was shown as a good way to compensate for the differences of the two analyses especially for use in validating precipitation produced by AGCMs.
Acknowledgments. The authors wish to thank Bob Adler, George Huffman, Pingping Xie, Robert Livezey, and John Janowiak for many fruitful discussions and other help. Also, we wish to acknowledge the NOAA Office of Global Programs, which provided support to two of the authors, AG and XS.

References
Arkin, P. A., and B. N. Meisner, 1987: The relationship between large-scale convective rainfall and cold cloud over the western hemisphere during 198284. Mon. Wea. Rev., 115, 5174. Ferraro, R., and G. Marks, 1995: The development of SSM/I rainrate retrieval algorithms using ground-based radar measurements. J. Atmos. Oceanic Technol., 12, 755770. , F. Weng, N. C. Grody, and A. Basist, 1996: An eight-year (19871994) time series of rainfall, clouds, water vapor, snow cover, and sea ice derived from SSM/I measurements. Bull. Amer. Meteor. Soc., 77, 891905. 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, 28102823. , and Coauthors, 1997: The Global Precipitation Climatology Project (GPCP) combined precipitation dataset. Bull. Amer. Meteor. Soc., 78, 520. Janowiak, J. E., and P. A. Arkin, 1991: Rainfall variations in the tropics during 19861989 as estimated from observations of cloud top temperature. J. Geophy. Res., 96 (Suppl.), 3359 3373. , A. Gruber, C. R. Kondragunta, R. E. Livezey, and G. J. Huffman, 1998: A comparison of the NCEPNCAR reanalysis precipitation and the GPCP rain gaugesatellite combined dataset with observational error considerations. J. Climate, 11, 29602979. Kalnay, E., and Coauthors, 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc., 77, 437471.

Kondragunta, C., and A. Gruber, 1997: Intercomparison of spatial and temporal variability of various precipitation estimates. Adv. Space. Sci., 19, 457460. McCollum, J., A. Gruber, and M. Ba, 2000: Discrepancy between gauges and satellite estimates of rainfall in equatorial Africa. J. Appl. Meteor., 39, 666679. Morrissey, M. L., and J. S. Greene, 1991: The Pacific atoll rain gage data set. Tech. Rep. 648, University of Hawaii at Manoa, Honolulu, Hawaii, 9 pp. [Available from University of Hawaii at Manoa, 2444 Dole St., Honolulu, HI 96822.] Reynolds, R. W., 1988: A real time global sea surface temperature analysis. J. Climate, 1, 7586. Rudolf, B., H. Hauschild, W. Ruth, and U. Schneider, 1996: Comparison of raingauge analysis, satellite-based precipitation estimates, and forecast model results. Adv. Space Res., 18, 5362. Shukla, J., and M. J. Fennessy, 1988: Prediction of time-mean atmospheric circulation and rainfall: Influence of Pacific sea surface temperature anomaly. J. Atmos. Sci., 45, 928. Spencer, R. W., 1993: Global oceanic precipitation from the MSU during 197991 and comparison to climatologies. J. Climate, 6, 13011326. Sperber, K. R., and T. N. Palmer, 1995: Interannual tropical rainfall variability in general circulation model simulations associated wth the atmospheric model intercomparison project. PCMDI Rep. 28, Lawrence Livermore National Laboratory, Livermore, CA, 79 pp. [Available from Lawrence Livermore National Laboratory, 7000 East Ave., Livermore, CA 945509234.] Wilheit, T. J., A. T. C. Chang, and L. S. Chiu, 1991: Retrieval of monthly rainfall indices from microwave radiometric measurements using probability distribution functions. J. Atmos. Oceanic Technol., 8, 118136. Xie, P. P., and P. A. Arkin, 1995: An intercomparison of gauge observations and satellite estimates of monthly precipitation. J. Appl. Meteor., 34, 11431160. , and , 1996: Analyses of global monthly precipitation using gauge observations, satellite estimates, and numerical model predictions. J. Climate, 9, 840858. , 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. , and , 1998: Global monthly precipitation estimates from satellite-observed outgoing longwave radiation. J. Climate, 11, 137164.

2644

Vol. 81, No. 11, November 2000

Das könnte Ihnen auch gefallen