The Version-2 Global Precipitation Climatology Project (GPCP) Monthly Precipitation Analysis (1979–Present)

a. Datasets utilized in Versions 1 and 2

b. Additional datasets used in Version 2

c. Error estimates

The error estimates for the input fields and combination products are based on Huffman (1997) and Huffman et al. (1997). The estimates are for random error (bias error is assumed to be removed in the analysis procedure) and include both algorithm and sampling random errors. The work is based on seeking to parameterize σ², the (random) error variance that remains when we compute r, the average precipitation rate, in a grid box from a finite set of observations (N samples). When the samples are sparse in space (gauges) or time (satellites), σ² may be approximated simply from the accumulated histogram of individual precipitation rates that contribute to r, including zeroes. The key step is that the summation over the histogram on which σ² depends can be nondimensionalized and labeled H, which turns out to have a very narrow range of values for typical precipitation histograms. After several approximations, the final form of the relation for σ² is

View Expanded

where S is introduced to account for σ²(r = 0) ≠ 0, I is a multiplicative factor approximately relating N to the number of independent samples in the histogram, and the term in the square brackets is an empirical parameterization of the probability of precipitation in terms of r, based on examination of satellite estimates and gauge analyses; H/I and S are taken to be constant and then set for each of the input single-source datasets by comparison against gauge data. Note that this calibration is most applicable when the local sampling and climatology are similar to those in the calibration area. Once the error estimates are calculated, they can provide weights for the combination procedure, and the users of both the input products and combined analyses have information with which to judge the utility of the dataset for their particular application.

a. Combination method for 1987–present

For the 1987–present period, during which SSM/I data are available, the combination method is very similar to that described in Huffman et al. (1997). It is designed to use the strengths of each input dataset to produce merged global, monthly precipitation fields that are superior to any of the individual datasets. The technique is also designed to reduce bias in each step by using the input original or intermediate product with the presumed smallest or zero bias to adjust the bias of other products. For example, microwave-based estimates are presumed to have lower biases than IR-based estimates, and gauges are presumed to be unbiased relative to the combination of satellite estimates over land.

A key feature of the GPCP merge technique has centered on combining the superior physical basis of the microwave-based observations from a low-orbit satellite and the frequent time sampling of the geosynchronous IR observations. Adler et al. (1991, 1993) described a technique for using precipitation estimates from low-orbit microwave data to “adjust” GPI precipitation estimates made from geosynchronous IR data. The resulting “microwave-adjusted IR” estimates provide an objective means of correcting known biases in the geosynchronous IR estimates, while retaining the high sampling rate of the geosynchronous satellite. This approach was then adapted to be applied on a global, tropical basis, as described by Adler et al. (1994), and by the GPCP (Huffman et al. 1997). First, for each month the microwave estimates are approximately time and space matched with geosynchronous IR GPI estimates to derive additive and multiplicative microwave/IR calibration factors (see Fig. 1). In regions lacking geosynchronous IR data, rainfall estimates from the NOAA polar-orbiting satellites are adjusted using a smoothly varying interpolation of the microwave/geosynchronous IR adjustment ratio. The spatially varying arrays of adjustment coefficients are then applied to the full month of GPI estimates, producing the adjusted GPI (AGPI) precipitation field, which has the desirable, 3-h sampling of the geosynchronous data and the typically small bias of the instantaneous microwave estimates. Verification against rain gauge analyses over water and land, and subjective examination of the resulting maps and zonally averaged fields indicate that known biases in the GPI in the subtropics and over land are reduced using this adjustment approach (Adler et al. 1993, 1994).

The GPCP Version-2 merging procedure introduces TOVS-based estimates to fill in the high latitudes, where SSM/I-based estimates are unavailable or unreliable. Equatorward of seasonally varying latitude boundaries (about 40°–50°N and 40°–50°S, chosen by inspection of monthly climatological zonal profiles) the microwave data is used by itself. Poleward of 70°N and 70°S the TOVS is adjusted based on a comparison of TOVS and climatological gauge values. In the midlatitudes an average of TOVS and SSM/I is used, with a smooth transition over 60°–70°N and 60°–70°S to the gauge adjustment at 70°N and 70°S. Areas lacking SSM/I equatorward of 60°N and 60°S (generally over land) are filled with TOVS, adjusted by the ratio of zonal average SSM/I (or SSM/I–TOVS average) and TOVS.

A multisatellite precipitation product is formed of AGPI, between 40°N and 40°S, and SSM/I and TOVS values elsewhere. In areas of the Tropics where geosynchronous data are unavailable (e.g., the Indian Ocean) microwave-adjusted low-orbit IR is combined with SSM/I estimates.

Next, the large-scale (5 × 5 grid box) average of the multisatellite analysis is adjusted to agree with the large-scale average of the gauges (over land and where available). This keeps the bias of the satellite and gauge combination close to the (presumably small) bias of the gauge analysis on a regional scale. Finally, the gauge-adjusted, multisatellite estimate and the gauge analysis are combined with inverse-error–variance weighting to produce the final, merged analysis. This gauge–satellite combination approach allows the multisatellite estimate to provide important local variations in gauge-sparse areas, while still retaining the overall gauge bias. The efficacy of this procedure was demonstrated by McCollum et al. (2000) in their study of biases in satellite estimates relative to gauges over equatorial Africa.

Random error fields are computed for each of the underlying fields and the final product. As outlined in section 2c for underlying fields, the nominal error depends on the particular field (e.g., microwave), and fluctuations depend on sampling and rain amount. The theoretically based functional form currently used was derived in Huffman (1997) and validated by Krajewski et al. (2000). Random error estimates for the satellite–gauge combination are computed in the combination step.

b. Combination method for 1979–87

The GPCP analysis is extended back to 1979 using the OPL (Xie and Arkin 1998) trained against a concurrent period of the GPCP-merged dataset. The OPI was developed based on the observation that, while the mean annual cycle of OLR (calculated using the NOAA AVHRR data stream) reflects both the surface temperature and cloudiness, OLR anomalies have a clear negative correlation with precipitation anomalies over most of the globe. This information was used to develop regression coefficients that relate OLR anomalies to precipitation. While these coefficients are spatially inhomogeneous and seasonally varying, they can be expressed rather accurately as a globally uniform linear function of the local mean precipitation anomalies. The precipitation anomalies are computed from the OLR anomaly relationship and then added to base period mean values to generate monthly precipitation accumulations. In this application the OPI is trained against the GPCP-merged analysis for the 1988–97 period. The OPI relations developed with the GPCP dataset are then applied to the OLR anomalies for the 1979–85 period to produce the monthly OPI-based precipitation estimates. The OPI estimates take the place of the SSM/I and geosynchronous estimates for this early period, and are then merged with the gauge analysis over land as in the later period.

a. Example month

Figure 2 (top panel) shows an example of the GPCP monthly merged final product, in this case for February 2001. The general rainfall patterns for this month are typical of Northern Hemisphere winter months. Over southeast Africa a maximum of precipitation (20 mm day⁻¹) is associated with the devastating floods that occurred there at this time. The middle panel of Fig. 2 has the random error field associated with the merged monthly estimate. The error estimate includes both error due to the algorithm and sampling error. The pattern closely follows the rain field in the top panel due to the close relation of the error to the estimated rain as indicated in Eq. (1). This relation can be seen in Fig. 3, which shows the relation between estimated error and rain rate for this month, with each point representing the result for a 2.5° square. The general increase of error with rain rate can be seen, but there is fairly large scatter. However, there are also identifiable curves formed by collections of points that relate to particular regions that contain the same combination of input data. The collection of points extending out to the letter A in the diagram is for ocean areas in the 40°N–40°S area where both SSM/I and geosynchronous IR data are used. The sampling provided by the IR data lowers the random error, as compared to the ocean areas outside latitude 40°, which have only the sampling of a single SSM/I, and are indicated by B in Fig. 3. The letter C denotes the near end point of a collection of points over land, which have a generally lower error due to the gauge input (in addition to satellite), but have significant error variability due to the variation in number of gauges in each box. Overall the errors range between 10%–30% at rain amounts above 100 mm month⁻¹ (3 mm day⁻¹). This percentage error is for a 2.5° latitude × 2.5° longitude square, and time and space averaging will significantly reduce the estimated random error.

To better see variations in the quality of the final estimate, a quality index (QI) has been defined (Huffman et al. 1997) that is approximately equivalent to the number of gauges required to produce the error in question (for the given amount of rain). The map of QI for the sample month is shown in the bottom panel of Fig. 2. In general, the land areas have a higher QI because the land areas have the gauge information in addition to that from the merged satellite estimates. Variations in QI over land are related mainly to variations in gauge density, with Europe and southeastern Australia having some of the areas of highest QI. Over the ocean the QI is higher (2–4) in the 40°N–40°S region relative to a QI of 1–2 at higher latitudes. The higher QI (lower error for the same rainfall) is due to the added sampling available with the geosynchronous satellite observations. The QI and error fields should be helpful for evaluation of research results utilizing this dataset. However, it should be remembered that the error estimates are for random error only, but include estimates of both retrieval error and sampling error.

b. Climatology

Figure 4 displays the 1979–2001 precipitation climatology map based on the GPCP Version-2 final merged product. The GPCP climatology shows the expected main features, with maxima in the Tropics in the intertropical convergence zone (ITCZ) in the Atlantic, Pacific, and Indian Oceans; in the South Pacific convergence zone (SPCZ); and over tropical Africa, South America, and the Maritime Continent between the Pacific and Indian Oceans. Dry zones in the eastern parts of the subtropical oceans are evident, as are the desert areas over land. The Pacific ITCZ is a narrow band on the Northern Hemisphere (NH) side of the equator, with peaks in both the western and eastern parts of the ocean with nearly equal values of 8.8 and 9.1 mm day⁻¹, respectively. The Atlantic Ocean ITCZ maximum is weaker and the Indian Ocean feature extends westward from Sumatra and narrows along the equator.

In the midlatitudes, the storm tracks in the Northern Hemisphere oceans are very distinct with peak values in the Atlantic and Pacific Oceans greater than in the Southern Hemisphere (SH) circumpolar storm track. A secondary maximum is evident along the northwest coast of North America from Alaska to California at the eastern end of the Pacific Ocean storm track. The Southern Hemisphere does have weak maxima southeast of Africa and South America, and a poleward extension of the SPCZ in the South Pacific Ocean. The narrow, circumpolar maximum at 65°S indicates the Southern Hemisphere storm track, but also is the latitude zone where the combination technique is transitioning from SSM/I-based precipitation estimates to TOVS-based estimates. Therefore, although the feature certainly exists, the magnitude and exact placement are more questionable than many of the other features in the climatology.

A comparison of the new Version 2 with the GPCP Version-1 climatology indicates slightly lower rainfall amounts over the tropical oceans due to a modification of the passive microwave algorithm over the ocean (Chang et al. 1995) and an increase at high latitudes (>50°) over the ocean due to the merger of the TOVS-based estimates with the SSM/I-based estimates at these latitudes.

The zonal-averaged latitudinal profile of precipitation is given in Fig. 5 for the GPCP climatology and for two conventional climatologies, from Jaeger (1976, hereafter JAE) and Legates and Willmott (1990, hereafter LW). The GPCP peak of 5.5 mm day⁻¹ at 6°N is nearly the same as that from JAE. The LW peak is related to a very strong maximum shown in their climatology in the central Pacific Ocean, which is not supported by satellite observations. The GPCP curve shows a secondary tropical peak at 5°S, which appears in the other climatologies as changes in slope of the decrease from the main peak. It is tempting to interpret this secondary peak as a manifestation of a second ITCZ on the south side of the equator, but the peak is actually a composite of various ocean and land features that do not connect over any large longitudinal range (see the map in Fig. 4). The overall tropical peak in the GPCP climatology (Fig. 5) is somewhat narrower than in either of the conventional climatologies, and the subtropical minima of just below 2 mm day⁻¹ are located a little equatorward of the climatologies, especially in the Southern Hemisphere. The land and ocean breakdown of the latitudinal profile is shown in Fig. 6, where it can be seen the double peak exists in the ocean by itself, but that the land profile has its overall peak just south of the equator. The land peak is closer to the equator than the ocean peak.

The GPCP NH midlatitude peak is similar in magnitude and shape to that of the LW profile (Fig. 5) with higher values than the JAE profile. The main peak, at about 40°N is an oceanic peak, not showing up over land (see Fig. 6). It is mainly composed of the two oceanic storm tracks east of Asia and North America. Poleward of that peak, the GPCP curve in Fig. 5 has a weak secondary maximum at 55°N, which is clearly based over land (see Fig. 6). The LW climatology has a similar feature. The ocean profile in Fig. 6 does show a leveling off before a steep drop off to the pole. This high-latitude activity perhaps reflects a separate cyclone track at this latitude, and also the southwest-to-northeast (ocean to land) orientation of the main midlatitude rain features in the NH. From 55°N the mean precipitation decreases rapidly toward the pole dropping below 1 mm day⁻¹ at 70°N. In this zone the GPCP agrees fairly well with the conventional climatologies.

In the SH the broad GPCP midlatitude maximum between 40°–60°S has two peaks, the larger being at 57°S. The exact latitude of the peaks in this zone varies among the climatologies. The distinct peak at 57°S in the GPCP field is also evident in the map in Fig. 4. As mentioned before, this zone is also where the GPCP merger technique transitions from SSM/I-based estimates to those from TOVS and because that transition is done as a function of latitude this peak, or at least the magnitude of the peak, is open to question and requires further analysis. From the midlatitude maximum the mean precipitation drops off rapidly toward the pole, similar to that of JAE. The similarities among the climatologies are encouraging, considering that they cover different periods of years. The differences, especially in the tropical oceans, are likely due to biases in the data sources.

In comparing the precipitation of the two hemispheres in Fig. 6 it is evident that the NH has more mean rainfall in the 0°–25° zone, probably due to the more extensive land mass in the NH, which in turn draws the ITCZ into the NH and also helps produce the stronger Asian monsoon in the NH. However, in the 25°–65° zone the SH has somewhat greater precipitation overall, but the NH has a greater total in this zone, if only oceans are considered. Thus, it seems that the standing waves in the NH help produce larger precipitation values over the oceans there, but the drier landmasses in this latitude zone reduce the total precipitation to slightly below the SH ocean and total values, which are nearly the same because of the relative lack of land in that zone.

c. Global totals

Table 1 contains the global precipitation totals for the GPCP Version 2, along with those for the two conventional climatologies. For the entire globe (90°N–90°S) the GPCP total is 2.61 mm day⁻¹, comparable to the value of JAE, but lower than that of LW. However, the LW value is affected by a questionable rainfall maximum in the central Pacific Ocean (see Huffman et al. 1997 and Janowiak et al. 1995 for discussion) so that its global total value is questionable. The breakdown between ocean and land indicates higher values over ocean and because the ocean also occupies the larger area, this means that 76% of global precipitation falls over the world's oceans according to the GPCP climatology.

When the latitude band is restricted to the Tropics, that is, 30°N–30°S, the mean goes up to 2.93 mm day⁻¹ and the difference between land and ocean is much smaller. The GPCP values in the Tropics are also in close agreement with those of JAE, but very different than those of LW for the aforementioned reason. As indicated in Fig. 5, the LW values appear very reasonable outside the Tropics, and even within the Tropics climate maps are similar to those of GPCP, except for the central Pacific Ocean. The GPCP Version-2 values for the Tropics are about 5% lower than the GPCP Version-1 means over the oceans, while over land they are nearly identical because of the influence of the rain gauges. The ocean difference is due to a modification of the microwave algorithm as mentioned earlier.

The lower part of Table 1 shows the totals for various latitude bands. When the globe is divided into the NH and SH Table 1 indicates that the NH has slightly more precipitation. When only considering oceans, the NH is still larger, but over land the SH has a higher mean precipitation rate. When the hemispheres are divided into Tropics and extratropics (bottom part of Table 1) the NH has more precipitation in the Tropics and the SH has more precipitation in the extratropics.

A time series of monthly, global anomalies from the 23-yr mean value in Table 1 is shown in Fig. 7. A 12-month running mean is also shown. The curve for land and ocean combined in the top panel indicates no noticeable trend over the observation period. This lack of a positive trend might be surprising considering predictions based on climate models of a precipitation increase associated with global warming. However, the size of the predicted increase is small and may not be detectable over a short period, such as 20 yr.

The monthly anomalies occasionally reach near a magnitude of 0.2 mm day⁻¹, but typically stay below 0.1 mm day⁻¹, a 4% variation when compared to the mean of 2.6 mm day⁻¹. The ocean and land separately have larger anomalies because of the smaller area, and the land variations are most obvious and are generally related to El Niño–Southern Oscillation (ENSO) variations. Figure 8 shows a similar diagram limited to 30°N–30°S and includes a plot of the Niño-3.4 (5°N–5°S, 120°–170°W) sea surface temperature (SST) index along with the indicators of the occurrence of major volcanic eruptions. The variations over land in the bottom panel are clearly connected to ENSO. Two major land areas in this latitude range, South America and the Maritime Continent, have dry periods during El Niño, and also the Indian summer monsoon rainfall is weaker, which results in the negative correlation between Niño-3.4 and land precipitation. A positive correlation with ocean rainfall is present, especially pronounced during the 1986/87 and 1997/98 El Niño events and the subsequent switches to La Niña conditions, but not so obvious with the 1982/83 and 1991/92 events. The major volcanic events also may play a role and might help to explain some decreases in precipitation and the lack of a positive ocean anomaly during the 1982 and 1992 El Niños. Research is ongoing in exploring these relations and possible causes.

d. Seasonal cycle

The evolution of the annual cycle of precipitation is seen in Fig. 9 in terms of mean maps for selected months. Figure 10 shows the zonally averaged annual cycle, separately for land and ocean and together. In tropical regions the major precipitation peaks in the Indian and Pacific Oceans, South America and the Maritime Continent are below the equator in January. In midlatitudes the NH ocean peaks are the strongest and are located further south in January. In the SH storm track the precipitation maximum is weakest. In April the western Pacific Ocean maximum is already larger north of the equator, but a double ITCZ pattern exists in the eastern Pacific Ocean. By July the precipitation in the Tropics has followed the sun northward and both the land and ocean peaks are farthest north. The NH storm tracks are weak, but the circumpolar track in the SH is at its strongest. The impact of land is in evidence in Fig. 10, which shows the mean zonally averaged annual cycle (two cycles are shown). The range of latitudes covered by the precipitation maximum during the annual cycle is clearly larger for the land areas than for the oceans, indicating the effect of the land heating at amplifying the annual cycle. In northern midlatitudes the ocean peak is clearly defined in winter at 40°N, while over land the peak is in summer, at 50°–60°N.

The regional characteristics of the seasonal cycle are given by the magnitude and phase of the first annual harmonic of precipitation seen in Fig. 11. The areas of large amplitude are associated with monsoonal regimes and shifts of the ITCZ. Shifts in phase of 180° are noted between South Asia and northern Australia, the Amazon and north coast of South America, and along the equator across the Pacific Ocean. The amplitude of the strong feature in the eastern Pacific is partly due to the impact of tropical cyclones there in July–September. The NH midlatitude storm tracks have broad maxima with approximately a January peak. A June peak exists in the SH storm track at 50°S from south of Africa to south of Australia.

Other interesting features include a variation in the subtropical dry areas. The SH oceanic dry areas west of South America and Africa stay dry year-round, but their NH counterparts vary seasonally to a much greater degree with a relatively wet winter. The NH polar region also shows a seasonal variation with peak precipitation in the summer.

e. Example of ENSO variations

One of the most important prospective uses for any analysis of global precipitation is to examine and understand interannual changes in the distribution of large-scale precipitation and their relationship to variations in the general circulation. It has long been clear that a significant portion of the interannual variability in the large-scale atmospheric circulation is associated with the ENSO phenomenon (Bjerknes 1969; Arkin 1982; Schneider and Fleer 1989). ENSO-related changes in global precipitation, as inferred from rain gauge observations, have been presented by Ropelewski and Halpert (1987, 1989, 1996), and Janowiak and Arkin (1991); Arkin et al. (1994), and Dai and Wigley (2000) have described the ENSO signal in precipitation estimated from satellite observations. GPCP Version 2 has already been used by Curtis and Adler (2000) to create indices of ENSO based on patterns of precipitation anomalies. Their standard rainfall index, the ENSO precipitation index (ESPI), is plotted with Niño-3.4 and precipitation anomalies divided by 4.0 over the Niño-3.4 domain (Fig. 12). During the 1997/98 El Niño, ESPI led Niño-3.4, which in turn led the collocated precipitation index.

The 23-yr record contains six El Niños, including the powerful events in 1982/83 and 1997/98 (Fig. 12). An important application of GPCP is to characterize the evolution of global precipitation during ENSO over the 23-yr period, for example, in Curtis and Adler (2003), where precipitation anomaly patterns as a function of state of evolution and intensity are presented. Here we examine rain rates, anomalies, and normalized anomalies for the seasons of July–August–September (JAS) 1997, January–February–March (JFM) 1998, JAS 1998, and JFM 1999 (Fig. 13). The boreal summer of 1997 (Fig. 13a) was characterized by a reversal of the Walker circulation, as defined by the east–west precipitation gradient in the equatorial Pacific (Curtis and Adler 2000). The strongest precipitation anomalies (Fig. 13b) were confined to the equator, with increases in rainfall over much of the Pacific and decreases over the Maritime Continent, the Amazon, and Central Africa. However, normalized anomalies, ranked percentiles of rain rate based on a gamma distribution (Ropelewski and Halpert 1987), show significant changes in seasonal precipitation in the mid- to high-latitudes (Fig. 13c). A weak La Niña emerged in the boreal summer of 1998 (Fig. 13g). A strong gradient of precipitation anomalies is seen across the Maritime Continent, with positive values southwest of Sumatra and negative values north of New Guinea (Fig. 13h). However, the normalized anomalies (Fig. 13i) do not show the global extremes observed during the El Niño. The La Niña strengthened into the beginning of 1999 (Fig. 13j). Above-normal rainfall was observed for the Maritime Continent and Amazon, but central Africa was as dry as JFM 1998 (Fig. 13k). Normalized anomalies show significantly dry conditions over the southwestern United States and southern Indian Ocean (Fig. 13l). For a detailed description of GPCP Version-2 observations of precipitation during the 1997–99 ENSO cycle, see Curtis et al. (2001).