Spatial and temporal variabilities of rainfall in tropical South America as derived from Climate Prediction Center merged analysis of precipitation

INTERNATIONAL JOURNAL OF CLIMATOLOGY Int. J. Climatol. 22: 175–195 (2002) Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/joc.724 SPATIAL AND TEMPORAL VARIABILITIES OF RAINFALL IN TROPICAL SOUTH AMERICA AS DERIVED FROM CLIMATE PREDICTION CENTER MERGED ANALYSIS OF PRECIPITATION H. MATSUYAMA,* J.A. MARENGO, G.O. OBREGON and C.A. NOBRE Centro de Previsão de Tempo e Estudos Climáticos–Instituto Nacional de Pesquisas Espaciais Rod. Pres. Dutra, km 40, CEP 12630-000, Cachoeira Paulista, SP, Brazil Received 10 November 2000 Revised 3 August 2001 Accepted 10 August 2001 ABSTRACT We investigated the spatial and temporal variabilities of Climate Prediction Center merged analysis of precipitation (CMAP) in tropical South America from 1979 to 1998. First, we validated CMAP using other hydrometeorological data. In comparison with the high-density precipitation data of the Global Historical Climatology Network (GHCN) Ver. 2, CMAP reproduces the spatial pattern well, although it underestimates (overestimates) heavy (light) precipitation. CMAP also reproduces the interannual variability well, compared with the discharge data of the River Amazon. Next, we applied the rotated empirical orthogonal function (REOF) to CMAP after subtracting the annual cycle. Simultaneous and lag correlations were calculated among the scores of REOFs 1 to 4, the southern oscillation index, and the dipole index of the Atlantic. REOF 1 (15%) represents the north–south pattern that exhibits the maximum precipitation in the summer hemisphere. REOF 2 (12%) indicates the gradual decrease of precipitation in the northern part of tropical South America, reflecting the effect of the Atlantic. REOF 3 (11%) exhibits an east–west pattern related to El Niño. In REOF 4 (7%), the centre of the factor loading is located in Colombia, and the score jumps abruptly around 1985–86. The Lepage test detected the abrupt increase of CMAP in 1985–86 around Colombia. Since such a jump is not found in GHCN Ver. 2, the discontinuous changes of CMAP and REOF 4 around 1985–86 are artificial and peculiar to CMAP. In this region, CMAP should be applied with caution when evaluating recent trends and the interannual variability. The importance of the abrupt increase of precipitation around Colombia is also addressed. Copyright  2002 Royal Meteorological Society. KEY WORDS: South America; Climate Prediction Center merged analysis of precipitation (CMAP); rotated empirical orthogonal function (REOF); Lepage test 1. INTRODUCTION Precipitation data are essential for studying the hydrological cycle at various spatio-temporal scales. In the past, precipitation measurements were conducted using rain gauges only where humans were living. Recently, satellites have enabled us to make various estimates of precipitation. These estimates are then merged to produce global monthly mean fields of precipitation. Climate Prediction Center (CPC) merged analysis of precipitation (CMAP; Xie and Arkin, 1997) is one such data set. Several individual sources are merged, including gauge observations, satellite estimates, and National Centers for Environmental Prediction (NCEP)–National Center for Atmospheric Research (NCAR) re-analyses. CMAP has been verified by various methods in Xie and Arkin (1997). Although different data sources were used for different subperiods, no discontinuity occurs in the latitudinal mean precipitation. Also, the * Correspondence to: H. Matsuyama, Department of Geography, Graduate School of Science, Tokyo Metropolitan University, 1-1. Minami-Ohsawa, Hachiouji-shi, Tokyo 192-0397, Japan; e-mail: matuyama@geog.metro-u.ac.jp Copyright  2002 Royal Meteorological Society 176 H. MATSUYAMA ET AL. long-term mean global fields and the interannual variability of precipitation in the Pacific seem reasonable. CMAP is, therefore, regarded as one version of ‘truth’ and is used for the diagnostic study of the global hydrological cycle (Trenberth and Guillemot, 1998). Since CMAP is primarily validated on a global scale, it is worthwhile performing continental-scale validation, such as in South America. This is because precipitation over the land is determined mainly by gauge observations in CMAP, and gauge networks are relatively sparse in South America (Xie and Arkin, 1996). Another dense gauge observation network, called the Global Historical Climatology Network (GHCN) Ver. 2 (referred to as GHCN2), compiled by the Carbon Dioxide Information Analysis Center (Peterson and Vose, 1997) is also valuable for this evaluation, since GHCN2 is not incorporated in CMAP. CMAP is used to validate general circulation model (GCM) experiments, because the merged analysis with the observations alone is prepared as well. In tropical South America, the 30 year climatology of precipitation by Figueroa and Nobre (1990) has been used to evaluate the performance of a GCM before conducting the deforestation experiments in the Amazon basin (e.g. Nobre et al., 1991; Lean and Rowntree, 1997). Centro de Previsão de Tempo e Estudos Climaticos–Instituto Nacional de Pesquisas Espaciais, Brazil (CPTEC–INPE) reproduced the climate of South America by GCM experiments for 1982 to 1991 and compared it with observations including CMAP (Cavalcanti et al., 2000; Marengo et al., 2000). It is, therefore, worthwhile validating CMAP using other hydrometeorological data, such as GHCN2 and river discharge data. In relation to this, Costa and Foley (1998) compared the six different estimates of precipitation data in the Amazon basin (three based on rain-gauge measurements alone, one based on a combination of raingauge measurements and satellite data, and two based on the reanalysis data). Although these data exhibited differences in the long-term mean field and the interannual variability, Costa and Foley (1998) did not conclude which data set was the most reliable. We will demonstrate the reliability of CMAP by comparing the river discharge data in the Amazon basin, which is an indicator of basin-scale wet and dry conditions. In addition, CMAP depicts interannual variability, since it is available for 20 years. In this respect, Obregon and Nobre (1990) applied the rotated empirical orthogonal function (REOF) to the monthly precipitation data (1951–85) with and without the annual cycle for 28 stations in the Amazon basin. They found that the second and fourth patterns of the REOF (REOFs 2 and 4) without the annual cycle seem to be related to the El Niño-southern oscillation (ENSO). Since only 28 stations were used, and these are distributed irregularly throughout the Amazon basin, it is worthwhile applying the REOF to CMAP, the gridded data, to confirm the conclusions of Obregon and Nobre (1990). By applying the REOF, we found that REOF 4 is peculiar to CMAP. The importance of this mode will also be addressed. We plan to investigate the interannual variability of the hydrological cycle in the Amazon basin using CMAP and other hydrometeorological data. The study area is, therefore, limited to tropical South America. In this study, CMAP will be validated first, then characteristics of CMAP will be explored using some statistical methods including the REOF. 2. DATA AND METHODS 2.1. CMAP CMAP is monthly precipitation data with 2.5° × 2.5° spatial resolution (Xie and Arkin, 1997). CMAP collection began in January 1979, and data up to December 1998 are used in this study. CMAP is composed of two kinds of data: standard precipitation (STD) and enhanced precipitation (ENH). STD consists of gauge observations (Rudolf et al., 1994; Xie et al., 1996) and five kinds of satellite estimate, namely outgoing longwave radiation (OLR)-based precipitation index (Xie and Arkin, 1998), infrared-based Geostationary Operational Environmental Satellite (GOES) precipitation index (GPI; Arkin and Meisner, 1987), microwave sounding unit (Spencer, 1993), microwave scattering from Special Sensor Microwave/Imager (SSM/I; Grody, 1991; Ferraro et al., 1994), and microwave emission from SSM/I (Wilheit et al., 1991). Since STD has some missing data at high latitudes, NCEP–NCAR reanalysis (Kalnay et al., 1996) is merged in ENH to supply missing data points. The merging algorithm (Xie and Arkin, 1996) and the blending algorithm (Reynolds, 1988) are fully documented in Xie and Arkin (1997). Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) TROPICAL SOUTH AMERICA RAINFALL VARIABILITIES 177 The relative errors in STD and ENH for each grid and for each month have been estimated (Xie and Arkin, 1996). In the preliminary analysis of the water budget in the Amazon basin, the annual precipitation for STD and ENH is almost identical, and the monthly difference is small (about ±10 mm/month). Since relative errors of ENH are generally smaller than those of STD, ENH is employed in this study. CMAP covers the whole globe with a 144 × 72 array. First, this array is converted to 144 × 73 to match the basin template of Oki et al. (1995) — some modification is made to the area of the Amazon basin. The data of the nearest four points are averaged. This does not qualitatively affect the following results. Considering the length of the analysis period (240 months), data for 240 grid points (11.25° N–18.75° S, and 33.75–83.75° W) are used. 2.2. GHCN2 GHCN2 (Peterson and Vose, 1997) is rain-gauge data compiled from more than 20 000 stations using 30 different sources. It contains monthly precipitation for each year, and the periods differ from station to station. GHCN2 has a larger number of stations, better quality control, homogeneity adjustments, and a wider selection of metadata than GHCN Ver. 1 (Vose et al., 1992). Although a beta release of GHCN2 is used in this study, the embedded document states that there will not be that many changes made prior to the final release. GHCN2 is used to evaluate the distribution of CMAP, since it is not incorporated in producing CMAP. Figure 1(a) depicts the number of GHCN2 stations by year in tropical South America after January 1979. The total number of stations gradually decreases, and suddenly drops in January 1992. This is common to the stations in Brazil, which account for a large part of the total. From these facts, GHCN2 can be used for the analysis until 1991. Since stations are fairly well distributed in Brazil (Figure 1(b)), the distribution of CMAP will be evaluated by GHCN2. Stations of GHCN2 (a) 3000 All Brazil 2500 2000 1500 1000 500 0 1980 1984 1988 1992 1996 2000 Year (b) 10°N 0° 10°S 80°W 70°W 60°W 50°W 40°W Figure 1. (a) Number of GHCN2 stations by year after January 1979. The stations located in the area 33.75–83.75 ° W and 11.25 ° N to 18.75 ° S are selected. Solid line indicates all stations in the region above, and the dashed line corresponds to stations in Brazil. (b) Observational network of GHCN2 from December 1981 to February 1991. Stations where data are available for periods of 8 years or more are plotted. SENAMHI stations in Peru are included Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) 178 H. MATSUYAMA ET AL. 2.3. Precipitation data in Peru GHCN2 has many stations in Brazil, but as much precipitation data should be collected in the other countries as possible. To this end, precipitation data for Peru, compiled by El Servicio Nacional de Meteorologia e Hidrologia (SENAMHI), are used. One of the authors (HM) archived these data, which mainly cover the period from 1978 to the middle of 1991 and which are collected from 27 stations. Matsuyama (1992a) confirmed the accuracy of the SENAMHI precipitation data by comparing it with the World Monthly Surface Station Climatology (WMSSC; Spangler and Jenne, 1990). In the following, the term ‘GHCN2’ will be used to mean the combination of the original GHCN2 data and the SENAMHI precipitation data. 2.4. River discharge data at Obidos River discharge data at Obidos, located in the lower reaches of the River Amazon (Figure 2), are obtained from Agéncia Nacional de Energia Elétrica, Brazil (ANEEL). They are daily data from December 1927 to May 1998. The data from September 1979 to August 1997 are used because there are no missing points. Discharge is expressed as runoff divided by the drainage area at Obidos (4 618 750 km2 ), which is used to evaluate the interannual variability of CMAP. The decreasing number of stations of GHCN2 in recent years makes it difficult to evaluate the interannual variability of the recent decade (Figure 1(a)). Independent data other than precipitation are therefore necessary for this validation. This study thus uses river discharge data, such as in Marengo (1995) and Marengo et al. (1998), as an indicator of the basin-scale wet and dry conditions. 2.5. Southern oscillation index (SOI) and dipole index (DI) The SOI and the DI (Servain, 1991), characteristic of the tropical Pacific and Atlantic respectively, are adopted to investigate the relationship between the REOFs and surrounding environments. The former is the normalized pressure difference between Tahiti and Darwin, and the latter is the difference of the normalized sea-surface temperature (SST) anomaly between the regions N and S in Figure 2 (N minus S). The DI points out the high correlation between tropical oceanic and atmospheric circulations (Servain, 1991). The WMSSC is used from March 1978 to February 1999 to calculate the SOI. The 2° × 2° monthly-mean SST of Reynolds (1988) for the same period is used for the DI. The regions for calculating DI are the same as employed in Wagner (1996) and differ slightly from the original definition of Servain (1991) to avoid the effect of the upwelling around the equator and the African coast. 30°N 20°N N 10°N 0° O 10°S S 20°S 80°W 70°W 60°W 50°W 40°W 30°W 20°W 10°W 0° Figure 2. Orientation map. The grey zone indicates the Amazon basin. River networks and position of Obidos (O) are also displayed. Thick lines illustrate the location of regions N (10–30 ° N, 10–50 ° W) and S (0–20 ° S, 0–40 ° W) Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) TROPICAL SOUTH AMERICA RAINFALL VARIABILITIES 179 3. VALIDATION OF CMAP 3.1. Distribution of precipitation Figure 3 depicts the December–February (DJF) total precipitation. We derived 10 year averages from December 1981 to February 1991 from GHCN2 and CMAP (Figure 3(a) and (b)); the differences are shown in Figure 3(c). In GHCN2, stations that have no missing data are selected for each year. These irregularly distributed station data are converted to 2.5° × 2.5° gridded data for each year, using adjustable-tension, continuous-curvature surface gridding (Smith and Wessel, 1990); the 10 year average is then calculated. Also, we plotted stations where GHCN2 data are available for eight or more years in Figure 3(a). Brazilian (a) 10°N 90 0 600 DJF (81/82-90/91) 0° 90 10°S 80°W (b) 10°N 70°W 60°W 70°W 60°W 0 50°W 40°W 50°W 40°W 300 600 0° 10°S 80°W (c) 10°N 1% 5% -6 00 0 0° 10°S 150 80°W 70°W 60°W 50°W 40°W Figure 3. The 3 month total precipitation (millimetres) from December to February. The 10 year average from December 1981 to February 1991 is displayed. (a) Derived from GHCN2. Contours are drawn at 30, 60, 150, 300, 450, 600, 750, 900, 1050 and 1200 mm to match Figueroa and Nobre (1990). Precipitation exceeding 900 mm is shaded. (b) Derived from CMAP. (c) (b)–(a). Contours are drawn every 150 mm. Negative values are shaded and drawn by dashed lines. Significant differences at the 5% (1%) level depicted by the Lepage test are displayed by small (large) circles Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) 180 H. MATSUYAMA ET AL. Amazonia has a fairly large number of stations, but many stations are concentrated in the northeastern part of Brazil (Nordeste). A DJF precipitation maximum appears in the southeastern part of the Amazon basin, with good agreement between GHCN2 and CMAP. Though some contours of 900 mm appear in GHCN2 (Figure 3(a)), such as in Figueroa and Nobre (1990), CMAP lacks this feature (Figure 3(b)). CMAP thus underestimates the heavy precipitation around the South Atlantic convergence zone (SACZ; Figure 3(c)). In contrast, CMAP overestimates the precipitation on the western side of the Andes, where the absolute values are small. We confirmed the difference statistically by the Lepage test (Lepage, 1971) — see Appendix A. Many grid points around the Andes display significant differences at the 1% level, which may come from the interpolation scheme used in this study. The root-mean-square error (RMSE) in Figure 3(c) is 132 mm (1.5 mm/day). Figure 4 is the same as Figure 3, but for June–August (JJA) 1982 to 1991. In this season, the intertropical convergence zone (ITCZ) shifts northward, which causes maximum precipitation above 900 mm north of the (a) 10°N 600 JJA (1982-1991) 00 12 90 0 0° 30 0 60 10°S 80°W 70°W 60°W 50°W 40°W 50°W 40°W (b) 10°N 600 0° 10°S 80°W 70°W 60°W (c) 50 0 0 -300 -4 0° -15 10°N 1% 5% 0 -150 10°S 80°W 70°W 60°W 50°W 40°W Figure 4. As Figure 3, but from June 1982 to August 1991 Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) TROPICAL SOUTH AMERICA RAINFALL VARIABILITIES 181 equator. In contrast, minimum precipitation is found around inner Nordeste. Both GHCN2 and CMAP exhibit these features, although minor differences exist between them. In JJA, CMAP underestimates heavy precipitation in the northern part of tropical South America (Figure 4(c)). The underestimation is common for DJF (Figure 3(c)). The only exception is that CMAP overestimates the precipitation of mountainous regions around Colombia. The reason for this will be discussed later. The significant differences at the 1% level are distributed everywhere in Figure 4(c). They are generally found where the absolute differences are large. In addition, some grid points reach the 1% level where the absolute differences are small, due to the non-linearity of the Lepage test. In the southern part of Peru and Nordeste, where there is less precipitation, the overestimation of CMAP is remarkable, and it also occurs in DJF. In JJA, the RMSE in Figure 4(c) is estimated as 125 mm (1.4 mm/day). This is almost the same as for DJF. The pattern correlation between CMAP and GHCN2 is calculated and displayed in Figure 5. The monthly anomaly data from September 1981 to August 1991 are used. Most grid points reach the 5% level except for those around the Andes, where significant differences are found both in Figures 3(c) and 4(c). We should keep this in mind for the following analyses. From these results, CMAP reproduces the distribution fairly well, although it underestimates (overestimates) heavy (light) precipitation. The reasons for the differences are (1) the relatively sparse gauge networks in South America in CMAP (Xie and Arkin, 1996), and (2) the merged analysis itself. For (1), it is worthwhile comparing GHCN2 and the gauge observations used for CMAP (Rudolf et al., 1994; Xie et al., 1996), whereas comparing the latter and CMAP itself will eliminate (2). Unfortunately, the gauge observations used for CMAP are not open to the public so far (C. Smith of CPC, personal communication). 3.2. Interannual variability of the water budget in the Amazon basin Figure 6 displays the interannual variability of the runoff at Obidos and the precipitation in the Amazon basin as evaluated by CMAP. Considering the number of GHCN2 stations (Figure 1(a)), the precipitation derived from GHCN2 is also plotted until the water year 1991. Note that the water year 1991 starts in September 1990 and ends in August 1991 (Marengo, 1992; Costa and Foley, 1998). In general, the runoff and precipitation have similar variations. The correlation coefficient between CMAP and runoff from 1980 to 1997 is calculated as 0.50 (significant at the 5% level). As mentioned above, the absolute values of GHCN2 from 1982 to 1991 are systematically larger than those of CMAP. In order to analyse this feature further, we tabulate the year-to-year variation of the total precipitation in the Amazon basin (Table I). The tabulation reveals that DJF total precipitation is systematically larger in GHCN2 than in CMAP. In contrast, CMAP overestimates JJA total precipitation from 1980 to 0.2 0.6 10°N 0. 6 0° 0.4 10°S 0.6 80°W 70°W 60°W 50°W 40°W Figure 5. Pattern correlation between CMAP and GHCN2 from September 1981 to August 1991. Monthly anomaly data are used for this calculation. Contours are drawn every 0.2. The shaded area is significant at the 5% level Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) 182 H. MATSUYAMA ET AL. P, Ro (mm/y) 3000 GHCN2 (2153 mm/y, 80-91) 2500 2000 CMAP (2014 mm/y, 80-97) 1500 1000 Runoff at Obidos (1139 mm/y, 80-97) 500 r = 0.50* (CMAP & Ro, 80-97) 0 1980 1984 1988 1992 1996 2000 Water Year (from previous September to August) Figure 6. Interannual variability of precipitation in the Amazon basin as evaluated by CMAP (thick solid line), GHCN2 (thin solid line), and runoff at Obidos (dashed line), integrated according to the water year. Owing to the rapid decrease of the number of GHCN2 stations (Figure 1(a)), GHCN2 is calculated until 1991. The correlation coefficient (significant at the 5% level) between CMAP and runoff for 1980–97 is also shown Table I. Year-to-year variation of the total precipitation in the Amazon basin as evaluated from CMAP and GHCN2 (unit: mm)a Water year DJF total JJA total Annual total CMAP GHCN2 CMAP GHCN2 CMAP GHCN2 Ro 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 712.5 660.3 754.0 581.0 661.8 648.6 655.9 606.8 692.4 725.4 718.7 663.9 557.9 668.7 672.3 614.4 728.5 658.0 541.0 730.4 707.1 803.3 612.1 760.1 708.9 742.7 690.6 748.6 788.0 729.8 711.3 — — — — — — — 292.0 328.9 362.3 232.5 304.0 321.0 297.8 264.3 264.0 317.8 320.0 286.1 284.2 277.6 329.8 273.3 300.2 232.5 284.7 305.8 309.0 319.7 257.6 328.2 349.5 362.1 295.6 318.7 375.8 346.9 328.8 — — — — — — — 2103.8 2092.3 2252.6 1774.9 2022.8 1980.8 2019.4 1942.3 2010.2 2159.0 2044.4 1965.6 1722.9 1904.3 2112.7 2034.2 2143.8 1973.3 1690.5 2103.7 2119.2 2280.6 1910.0 2233.7 2104.2 2216.7 2089.2 2107.4 2343.8 2135.3 2193.7 — — — — — — — 968.9 1035.8 1200.5 1014.8 1120.7 1088.6 1204.1 1186.9 1112.7 1293.5 1193.6 1156.9 946.3 1171.2 1314.7 1123.6 1147.1 1221.8 — Average 1980–91 1980–97 673.4 665.6 727.7 — 299.2 293.8 324.8 — 2030.7 2014.4 2153.1 — 1131.4 1139.0 a Note that the water year 1980 (DJF 1980) corresponds to the period from September 1979 to August 1980 (from December 1979 to February 1980). The column ‘Ro’ shows the annual runoff at Obidos for that water year. 1982. An almost identical annual precipitation from 1980 to 1982 (Figure 6) will be derived from this compensation. In the Amazon basin, precipitation and runoff decrease during ENSO periods (e.g. Richey et al., 1989; Marengo, 1992). In Figure 6, ENSO episodes of 1982–83 and 1991–93 are captured well, although the Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) TROPICAL SOUTH AMERICA RAINFALL VARIABILITIES 183 1987–88 event is unclear. From 1980 to 1997, the annual runoff of Obidos (CMAP precipitation) is calculated as 1139 mm/year (2014 mm/year). In the Amazon basin, about half of the annual precipitation drains to the Atlantic, and the remaining half is recycled over the basin (e.g. Salati, 1987; Matsuyama, 1992b). In the light of this observation, the ratio obtained, 0.57 (1139/2014), agrees well with these previous studies. The interannual variability of precipitation in the Amazon basin was also investigated by Costa and Foley (1998). In comparison with their study, we concluded that CMAP is quantitatively reliable for both the long-term mean and the interannual variability. This advantage is derived from the analysis of the water budget in the Amazon basin, using the river discharge data. Since CMAP was found reliable, spatio-temporal characteristics of CMAP will be explored by REOF next. 4. PREVAILING PATTERNS OF CMAP AS REVEALED BY REOF 4.1. Spatial and temporal characteristics of REOFs This section describes the spatio-temporal characteristics of REOFs 1 to 4. The physical interpretation of each REOF will be presented in Section 4.3. The REOF starts from a correlation matrix of 240 × 240 grid points. To this end, we prepared a time series without the annual cycle for each grid point by subtracting the 20 year mean precipitation (from January 1979 to December 1998) of each month from the respective monthly value. The first six components are rotated by the varimax method, such as in Obregon and Nobre (1990). In this study, we discuss the characteristics of REOFs 1 to 4 to compare them with those derived by Obregon and Nobre (1990), then we investigate the trend of the scores of REOFs 1 to 4 by the two-sided version of the Mann–Kendall rank statistic (Press et al., 1989). The scores are prewhitened by the ‘AR(1) filter’ of Kulkarni and von Storch (1995) to be regarded as independent samples before applying the Mann–Kendall rank statistic. Figure 7(a) shows the spatial pattern of the factor loading of REOF 1. The negative (positive) sign prevails south (north) of central Amazonia. Apparently, REOF 1 exhibits a north–south pattern, which explains 15% of the total variance. The score has no significant trend at the 5% level (Figure 7(b)). The spatial patterns of REOF 2 and REOF 1 look similar (Figures 8(a) and 7(a)), in that the negative sign covers Nordeste and the eastern part of the Amazon basin. Also, the positive sign prevails over the northern part of Amazonia, although the center is located in the Atlantic. This pattern explains 12% of the total variance, and the score has a decreasing trend, significant at the 5% level (Figure 8(b)). For REOF 3 (Figure 9(a)), the positive sign covers most of the studied area, while a strong negative sign is found in the Pacific. REOF 3 thus displays an east–west pattern, which explains 11% of the total variance. The score (Figure 9(b)) has no significant trend and looks like that of the SOI (figure not shown); negative values prevail in 1982–83, 1987–88, and 1991–93 during ENSO warm events. Figure 10 is the same as Figure 7 but for REOF 4. The negative sign prevails from northwest to southeast in the tropical area, and the centre is found in Colombia. This northwest–southeast pattern explains 7% of the total variance, and the score has some conspicuous features (Figure 10(b)). It shows an abrupt jump around 1985–86, but this decreasing trend is not significant at the 5% level. 4.2. Relationships among the scores of REOFs, SOI, and DI Next, we investigate the relationships among REOFs, SOI and DI because of the interannual variability of precipitation in accordance with ENSO in this region (e.g. Ropelewski and Halpert, 1987; Richey et al., 1989). Also, the effect of the Atlantic is important (e.g. Nobre and Shukla, 1996; Wagner, 1996; Uvo et al., 1998). To explore which components are related to SOI and which to DI, we calculated simultaneous and lag correlations with the scores of REOFs 1 to 4 and illustrated them as in Zeng (1999). Figure 11(a) is the result for SOI. Here, positive lags indicate SOI leading. The maximum correlation, which is significant at the 5% level, appears when SOI leads by 3 months. This feature is also found in Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) 184 H. MATSUYAMA ET AL. (a) 10°N 0.2 0 0° 0.2 -0 .8 10°S 80°W 70°W 60°W 50°W 40°W Score (b) 4 3 2 1 0 −1 −2 −3 −4 −5 NS 1980 1984 1988 1992 1996 2000 Year Figure 7. (a) Distribution of the factor loading of REOF 1. Contours are drawn every 0.1. The hatched area indicates the Amazon basin. (b) Score of REOF 1. Raw value (thin line), 13 month running mean (thick line), and linear trend (thick dotted line) are displayed. Significant level of the Mann–Kendall rank statistic, applied to the prewhitened time series (Kulkarni and von Storch, 1995) is also indicated. NS means that it is not significant at the 5% level the relationship between SOI and CMAP itself (Zeng, 1999). In contrast, scores of the other REOFs do not exhibit any significance. For DI, none of the correlation coefficients reach the significance level (Figure 11(b)). This may be explained as follows. DI decreases from 1979 to 1984 (Figure 11(c)), reflecting the decadal-scale variation (Servain, 1991). After that, it increases steadily except for the sudden drop and recovery from 1993 to 1994. The differences of the variation for the recent decades thus lead to the low correlation for 1979–98 among DI and REOFs 1 to 4. Although the absolute values of the correlation coefficients are low, the 1 month lag correlation with REOF 2 and the simultaneous correlation with REOF 3, have local extrema (Figure 11(b)). This will be discussed in Section 4.3, in relation to the interpretation of REOFs. 4.3. Physical interpretation of each REOF Based on the above results, we will present the physical interpretation of each REOF here, with the aid of important findings of previous studies. The north–south pattern of REOF 1 is also depicted as REOF 1 by Obregon and Nobre (1990). They stated that their REOF 1 represents the maximum precipitation in the summer hemisphere, derived by applying REOF to the precipitation series with the annual cycle. They found that the factor loading and scores of their REOF 1 with and without the annual cycle are quite similar, which is also confirmed by this study (results not shown). Obregon and Nobre (1990) also pointed out that some seasonality still remains in the precipitation series without the annual cycle. Specifically, the variance is larger in the rainy season than in the dry season. From these facts, REOF 1 of this study also represents the maximum precipitation in the summer hemisphere. Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) TROPICAL SOUTH AMERICA RAINFALL VARIABILITIES 185 (a) 10°N 0.2 0° 10°S -0. 0 80°W 70°W 2 60°W 50°W 40°W Score (b) 4 3 2 1 0 −1 −2 −3 −4 −5 5% 1980 1984 1988 1992 1996 2000 Year Figure 8. As Figure 7, but for REOF 2. The decreasing trend of the prewhitened time series is significant at the 5% level (a) 10°N 2 0.4 0. 0.6 0° 10°S 0 80°W 70°W 60°W 50°W 40°W Score (b) 4 3 2 1 0 −1 −2 −3 −4 −5 NS 1980 1984 1988 1992 1996 2000 Year Figure 9. As Figure 7, but for REOF 3 Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) 186 H. MATSUYAMA ET AL. (a) 10°N -0.6 .4 -0 0° -0.2 0 10°S 80°W 70°W 60°W 50°W 40°W Score (b) 4 3 2 1 0 −1 −2 −3 −4 −5 NS 1980 1984 1988 1992 1996 2000 Year Figure 10. As Figure 7, but for REOF 4 As mentioned later, abrupt changes do not occur in the score of REOF 1. Also, the score has no significant trend (Figure 7(b)). The low correlation among the scores of REOF 1, SOI, and DI (Figure 11(a) and (b)) also indicates that REOF 1 is not related to the long-term variability or periodic phenomena. The score of REOF 2 has a significant decreasing trend (Figure 8(b)). This does not contradict the decreasing (increasing) trend of precipitation in the northern (southern) part of the Amazon basin after the 1960s (Marengo and Nobre, 2001). Also, an increase of precipitation is reported in Nordeste after 1980 (Marengo et al., 1998). The southward shift of the ITCZ in the Atlantic after the 1950s (Wagner, 1996) is responsible for these phenomena. REOF 2 thus represents the effect of the Atlantic. In order to support this idea, the 1 month lag correlation between the score of REOF 2 and SST is calculated (SST leads 1 month), based on the results of Figure 11(b). The dipole pattern appears in Figure 12, where the equatorial northern (southern) Atlantic shows positive (negative) correlations. It is found that a few grid points reach the 5% significant level. Tanimoto and Xie (1999) illustrated that the dipole pattern, represented by DI, appears not only on a decadal time scale (8 to 16 years) but also on an interannual time scale (<5 years). In this case, the 1 month lag correlation in Figure 11(b) reflects the interannual time scale phenomenon. The cooling (warming) of SST in region N (S) in Figure 2 will thus decrease (increase) the moisture flux convergence in the northern (southern) part of the Amazon basin through evaporation, decreasing (increasing) the precipitation in these parts. REOF 3 is related to ENSO, as confirmed statistically by the lag correlation with SOI (Figure 11(a)). This pattern is also depicted as REOF 4 by Obregon and Nobre (1990). It is reported that precipitation in the Amazon basin and Nordeste decreases during ENSO periods (e.g. Ropelewski and Halpert, 1987; Richey et al., 1989; Marengo, 1992), when SOI changes 3 or 4 months before precipitation decreases (Zeng, 1999). This feature is also captured by Figure 11(a). ENSO affects the formation of the SST anomaly in the northern Atlantic through the atmospheric teleconnection (e.g. Nobre and Shukla, 1996; Uvo et al., 1998). A significant lag correlation, therefore, Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) TROPICAL SOUTH AMERICA RAINFALL VARIABILITIES 187 Correlation Coefficients (a) 0.6 1% REOF3 5% 0.4 REOF2 0.2 REOF4 0.0 REOF1 −0.2 0 −2 2 4 6 8 10 Lag (months of leading SOI) Correlation Coefficients (b) 5% 0.4 0.2 REOF2 0.0 REOF1 REOF4 −0.2 REOF3 0 −2 2 4 6 8 10 Lag (months of leading Dipole Index) Normalized SSTA (N-S) (c) 4 3 2 1 0 −1 −2 −3 −4 NS 1980 1984 1988 1992 1996 2000 Year Figure 11. (a) Simultaneous and lag correlations between SOI and the scores of REOFs 1 to 4, calculated from January 1979 to December 1998. Positive lag indicates SOI leading. Significant levels (5%, 1%) of the correlation coefficients are indicated by the dashed lines. (b) Same as (a), but for DI from January 1979 to December 1998. Only the 5% level is displayed that differs from (a), considering the autocorrelation of SOI, DI and the scores of REOFs (Leith, 1973). (c) As Figure 7(b), but for DI without the linear trend exists between SOI and DI, in which the changes of SOI lead those of DI by a few months (Servain, 1991). In Figure 11(b), the simultaneous correlation between DI and the score of REOF 3 has local extrema, which reflects the lag relation between SOI and DI. The spatial pattern of the factor loading of REOF 4 looks like that of REOF 2 obtained by Obregon and Nobre (1990), although an ordered difference is found between them. Since the rotation procedure is the same, this difference arises from the analysis years and different data sets. Obregon and Nobre (1990) pointed out that their REOF 2 is related to ENSO. However, the score of REOF 4 in this study has little relation to SOI or to DI (Figure 11(a) and (b)). The physical interpretation of REOFs 1 to 3 is thus successful, but does not apply to REOF 4. We will discuss what REOF 4 represents in Section 5. Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) 188 H. MATSUYAMA ET AL. 30°N 0.1 0.2 20°N 0.2 10°N 0.3 2 0.1 0 0° 3 0. 0. 0 10°S 20°S 80°W 70°W 60°W 50°W 40°W 30°W 20°W 10°W 0° Figure 12. The 1 month lag correlation between monthly anomaly of SST and the score of REOF2 (SST leads 1 month). Contours are drawn every 0.05. Negative values are drawn by dashed lines. The grid points that reach the 5% significance level are shaded 5. DISCUSSION 5.1. Detection of the abrupt jump in the score of REOF 4 around 1985–86 REOF 4 is characterized by an abrupt jump of the score around 1985–86. Here, we will show objectively that this jump is the largest of all scores. The Lepage test is again used for this evaluation. In this case, we searched the sample size m in multiples of 12 months. As is found in the interpretation of REOF 1, some seasonality still remains in the precipitation series without the annual cycle. Therefore, the selection of m = 12 is not suitable in this case. Also, when m is set to 36 or larger, the analysis period becomes too short. As a compromise, m is selected as 24. In the following, all Lepage tests are conducted with m = 24. Before this test, all time series are prewhitened by the ‘AR(1) filter’ of Kulkarni and von Storch (1995), such as in the case of the Mann–Kendall rank statistic. The results are displayed in Figure 13. None of the HK values (the Lepage statistics) reach the 5% level in REOF 1 (Figure 13(a)), although some HK values pass the 5% level in REOF 2 (Figure 13(b)). The situation changes in REOF 3. Statistically significant peaks appear every 2 or 3 years in Figure 13(c). El Niño and La Niña events are responsible for these sharp peaks, since the Lepage test can detect cyclic variations (Yonetani, 1993). Some El Niño events persist for 2 years (Tomita and Yasunari, 1993); hence, the Lepage test with m = 24 clearly detects the El Niño–La Niña cycle. A peak is also apparent in REOF 4 (Figure 13(d)). Although the HK value in 1996–97 almost reaches the 1% level, it is found that the peak in 1985–86 is the largest of all the HK values in Figure 13. The abrupt jump of the score of REOF 4 (Figure 10(b)) is thus objectively detected by the Lepage test with m = 24. Next, we seek to determine the distribution of the abrupt change around 1985–86. 5.2. Distribution of the abrupt jump around 1985–86 in CMAP and GHCN2 Here, the Lepage test is applied to CMAP without the annual cycle for each grid point. These anomaly data are also prewhitened by the ‘AR(1) filter’. Figure 14(a) shows the distribution of the abrupt change of CMAP from 1985 to 1986, where HK values exceed the 5% and 1% levels. The abrupt increases are concentrated in the mountainous regions of Panama, Colombia, Ecuador and Peru, as well as in central Amazonia and Nordeste. In contrast, the abrupt decreases are mainly found in the equatorial Atlantic and the South Pacific. Over land, this figure is similar to the distribution of the factor loading of REOF 4 (Figure 10(a)), in that the strong signals appear in Colombia and the adjacent regions. Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) TROPICAL SOUTH AMERICA RAINFALL VARIABILITIES 189 HK (dimensionless) (a) 20 REOF 1 (m=24) 15 10 1% 5% 5 0 1980 1984 1988 1992 1996 2000 HK (dimensionless) (b) 20 REOF 2 (m=24) 15 10 1% 5% 5 0 1980 1984 1988 1992 1996 2000 HK (dimensionless) (c) 20 REOF 3 (m=24) 15 EN 10 EN EN 1% 5% 5 0 1980 1984 1988 1992 1996 2000 HK (dimensionless) (d) 20 REOF 4 (m=24) 15 10 1% 5% 5 0 1980 1984 1988 1992 1996 2000 Year Figure 13. Time series of the Lepage statistic (HK, dimensionless) applied to the prewhitened scores of REOFs 1 to 4, when the averaging period m is set to 24 months. Significance levels (5% and 1%) are indicated by the dashed lines. (a) REOF 1. (b) REOF 2. (c) REOF 3; EN roughly indicates the warm periods of ENSO. (d) REOF 4 To validate Figure 14(a), we performed the same analysis with GHCN2. We selected 775 stations without missing data from January 1983 to December 1988. The stations where HK values exceed the 5% and 1% levels during 1985–86 are displayed in Figure 14(b). Although some stations in Colombia show the abrupt increase at the 5% level, many stations in Colombia show abrupt decreases during 1985–86. Note that many Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) 190 H. MATSUYAMA ET AL. (a) 10°N 0° 10°S 80°W 70°W 60°W 50°W 40°W 80°W 70°W 60°W 50°W 40°W (b) 10°N 0° 10°S Figure 14. Distribution of the abrupt change of precipitation in 1985–86, detected by applying the Lepage test to the prewhitened time series of precipitation anomaly. The averaging period m is set to 24 months. Black (white) circles indicate the increase (decrease) in 1985–86. Significant changes at the 5% (1%) level are displayed as small (large) circles. (a) The results of CMAP. (b) The results of GHCN2 other stations have no significant changes. This feature apparently differs from Figure 14(a), where the abrupt increases prevail in and around Colombia. We calculated the time series of the precipitation anomaly in Colombia and Panama by CMAP and GHCN2 respectively. An average of six grid points of CMAP (Figure 15(a)), where the precipitation increase rate from January 1979 to December 1998 exceeds 1 mm/year, is shown in Figure 15(b). The average of the monthly precipitation anomaly as evaluated by GHCN2 is depicted in Figure 15(c). Since the number of stations of GHCN2 in Colombia and Panama suddenly drops in January 1991, the anomaly is calculated until the end of 1990. An abrupt increase is found at the beginning of 1986 in CMAP (Figure 15(b)), whereas no such discontinuous change is found in GHCN2 (Figure 15(c)). Around 1985–86, no abrupt changes of OLR (Chu et al., 1994) or river discharge (Poveda and Mesa, 1997; Restrepo and Kjerfve, 2000) are reported in Colombia. Furthermore, the precipitation in the Amazon basin shows no such discontinuous increase in 1985–86 (Marengo and Nobre, 2001). From these facts, we concluded that the abrupt change of CMAP around 1985–86 (Figure 15(b)) is artificial. Judging from the description of Xie and Arkin (1997), causes of the abrupt change of CMAP are (1) the introduction of GPI in 1986, and (2) the change of the gauge observations from Xie et al. (1996) to Rudolf et al. (1994) in 1986. Cause (1) does not contradict Figure 15(b), since GPI overestimates the precipitation over tropical land (Arkin and Meisner, 1987; Janowiak, 1992), and cumulus clouds frequently cover the mountainous region around Colombia. The distribution of the factor loading of REOF 4 (Figure 10(a)) does not contradict (1) either, since the strong signals mainly appear over tropical land. In order to evaluate the effect of introducing GPI, it would be better to prepare CMAP without the GPI version, and to compare it with the original CMAP. Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) 191 TROPICAL SOUTH AMERICA RAINFALL VARIABILITIES (a) 10°N 0 0° 0 -1 0 10°S 80°W 70°W 60°W 50°W 40°W Anomaly of P (mm/mon) (b) 150 n=6 100 50 0 −50 −100 CMAP −150 1980 1984 1988 1992 1996 2000 200 200 150 150 100 100 50 50 0 0 −50 GHCN2 Number of Stations Anomaly of P (mm/mon) (c) −100 1980 1984 1988 1992 1996 2000 Year Figure 15. (a) Distribution of the linear trend of precipitation (mm/year) from January 1979 to December 1998. Contour interval is 1.0 mm/year. Solid (dashed) lines denote positive (negative) trends. Grid points with small (large) circles indicate that the prewhitened time series reach the 5% (1%) significance level. (b) Monthly precipitation anomaly of CMAP, averaged for the six grid points in Colombia and Panama, where the linear trend is larger than 1.0 mm/year. The linear trend is depicted as the dashed line. (c) Same as (b) but for GHCN2 from 1979 to 1990 in Colombia and Panama (solid line), calculated with all stations available at that month (dashed line). Annual cycle is removed by subtracting the average (January 1979–December 1990) from the respective monthly value For (2), Xie et al. (1996) compared their precipitation data with Rudolf et al. (1994) from 1986 to 1992. They found the difference to be small, but the effect of changing the gauge data in 1986 has not yet been investigated (P. Xie of NCEP, personal communication). To solve this problem, we must apply the Lepage test to the time series of Xie et al. (1996) and Rudolf et al. (1994). As mentioned in Section 3, however, the data set of Xie et al. (1996) is not open to the public. 5.3. Comparison with related studies We will next discuss the problem of applying CMAP to evaluate precipitation trends in tropical South America, based on the findings of Section 5.2. Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) 192 H. MATSUYAMA ET AL. Chu et al. (1994) investigated the recent trend of OLR in the Amazon basin from 1974 to 1990. They found that the convective activity in the Amazon basin as a whole has an increasing trend. They also found that the Mann–Kendall rank statistic of OLR reaches the 5% level in some grid points of Colombia. Dias de Paiva and Clarke (1995) validated the result of Chu et al. (1994) using the precipitation data of 48 stations in Brazilian Amazonia that had collected data for 15 years or more up to the early 1990s. They spatially extrapolated the results for Brazilian Amazonia to the adjacent countries, including Colombia. Although the decreasing trend covers Colombia, the procedure of Dias de Paiva and Clarke (1995) is not a true validation. CMAP can be utilized to evaluate the recent trend of precipitation, considering the length of the data collection period used by Chu et al. (1994), and by Dias de Paiva and Clarke (1995). Figure 15(a) shows the distribution of the linear trend of precipitation from January 1979 to December 1998. The increasing trend of the six grid points in Colombia and Panama exceeds 1 mm/year. We would like to use this figure to validate Figure 4 of Chu et al. (1994). Here, we have to remember that the significant increasing trends of Figure 15(a) result from the abrupt jump around 1985–86 (Figure 15(b)). Since this discontinuous change is artificial, we must be careful when applying CMAP to evaluate precipitation trends in this region and the interannual variability. 6. CONCLUSIONS We validated CMAP using other hydrometeorological data, then investigated the spatial and temporal variabilities of CMAP in tropical South America. The important and new findings of the present study are summarized below. (1) CMAP reproduces a spatial pattern in agreement with GHCN2, although it underestimates (overestimates) heavy (light) precipitation. CMAP also reproduces the interannual variability, in agreement with the river discharge near the mouth of the Amazon river. (2) We applied the REOF to CMAP after subtracting the annual cycle and calculated simultaneous and lag correlations among SOI, DI, and the scores of REOFs 1 to 4. REOF 1 (15%) represents the north–south pattern and indicates the maximum precipitation in the summer hemisphere. The score of REOF 2 (12%) indicates a gradual decrease of precipitation in the northern part, which reflects the effect of the Atlantic. REOF 3 (11%) exhibits an east–west pattern related to ENSO. In REOF 4 (7%), the centre of the factor loading is located in Colombia, and the score displays an abrupt jump around 1985–86 that cannot be explained well on a physical basis. (3) CMAP rainfall shows an abrupt increase in 1985–86 around Colombia. Since such a jump is not found in GHCN2, the abrupt increase of CMAP is artificial. REOF 4 above is thus also artificial and peculiar to CMAP. In this region, CMAP should be applied with caution when evaluating recent trends and the interannual variability. ACKNOWLEDGEMENTS This study was conducted when the corresponding author (HM) was a visiting scientist at CPTEC/INPE as a Japan Society for the Promotion of Science Postdoctoral Fellow for Research Abroad (1998 to 2000). Comments and English correction by Dr P. Satyamurty of CPTEC/INPE are sincerely appreciated. Constructive comments by an anonymous reviewer have been valuable for revising the original manuscript. APPENDIX A A.1. The Lepage test, modified after Yonetani (1992) The Lepage (1971) test is a non-parametric test that investigates significant differences between two samples, even if the distributions of the parent populations are unknown. When the size of each sample m is equal to or Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) TROPICAL SOUTH AMERICA RAINFALL VARIABILITIES 193 greater than ten, the Lepage statistic (HK) follows the χ 2 distribution with two degrees of freedom. Hirakawa (1974) demonstrated that the Lepage test is more statistically powerful than other similar non-parametric tests. The Lepage statistic (HK) is a sum of the squares of the standardized Wilcoxon’s and Ansari–Bradley’s statistics (Equation(1)). When HK exceeds 5.991 (9.210), the difference between two samples is judged at the 5% (1%) significance level. HK = [W − E(W )]2 /V (W ) + [A − E(A)]2 /V (A) (1) HK is calculated as follows. Let x = (x1 , x2 , . . . , xn1 ) and y = (y1 , y2 , . . . , yn2 ) be two independent samples of size n1 and n2 respectively. Assume that ui = 1 if the ith smallest observation in a combined sample of the size (n1 + n2 ) belongs to x, and ui = 0 if it belongs to y. The terms of Equation (1) are expressed as follows. W = n
1 +n2 iui (2) i=1 1 n1 (n1 + n2 + 1) 2 1 n1 n2 (n1 + n2 + 1) V (W ) = 12 n
n1 1 +n2
(n1 + n2 − i + 1)ui iui + A= E(W ) = i=1 (3) (4) (5) i=n1 +1 When n1 + n2 is even, E(A) and V (A) are estimated as follows: n1 (n1 + n2 + 2) 4 n1 n2 (n1 + n2 − 2)(n1 + n2 + 2) V (A) = 48(n1 + n2 − 1) (6) E(A) = (7) When n1 + n2 is odd, these equations become E(A) = n1 (n1 + n2 + 1)2 4(n1 + n2 ) (8) and V (A) = n1 n2 (n1 + n2 + 1)[(n1 + n2 )2 + 3] 48(n1 + n2 )2 (9) REFERENCES Arkin PA, Meisner BN. 1987. The relationship between large scale convective rainfall and cold cloud over the western hemisphere during 1982–1984. Monthly Weather Review 115: 51–74. Cavalcanti IFA, Satyamurty P, Marengo JA, Trosnikov I, D’Almeida C. 2000. Climatological features represented by the CPTEC/COLA global circulation model. In Sixth International Conference on Southern Hemisphere Meteorology and Oceanography, Santiago, Chile. Chu PS, Yu ZP, Hastenrath S. 1994. Detecting climate change concurrent with deforestation in the Amazon basin: which way has it gone? Bulletin of the American Meteorological Society 75: 579–583. Costa MH, Foley JA. 1998. A comparison of precipitation datasets for the Amazon basin. Geophysical Research Letters 25: 155–158. Dias de Paiva EMC, Clarke RT. 1995. Time trends in rainfall records in Amazonia. Bulletin of the American Meteorological Society 76: 2203–2209. Ferraro RR, Grody NC, Marks GF. 1994. Effects of surface conditions on rain identification using SSM/I. Remote Sensing Reviews 11: 195–209. Figueroa SN, Nobre CA. 1990. Precipitation distribution over central western tropical south America. Climanálise 5(6): 36–45. Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) 194 H. MATSUYAMA ET AL. Grody NC. 1991. Classification of snow cover and precipitation using the Special Sensor Microwave Imager. Journal of Geophysical Research 96: 7423–7435. Hirakawa K. 1974. The comparison of powers of distribution-free sample test. TRU Mathematics 10: 65–82. Janowiak JE. 1992. Tropical rainfall: a comparison of satellite-derived rainfall estimates with model precipitation forecasts, climatologies, and observations. Monthly Weather Review 120: 448–462. Kalnay E, Kanamitsu M, Kistler R, Collins W, Deaven D, Gandin L, Iredell M, Saha S, White G, Woollen J, Zhu Y, Chelliah M, Ebisuzaki W, Higgins W, Janowiak J, Mo KC, Ropelewski C, Wang J, Leetmaa A, Reynolds R, Jenne R, Joseph D. 1996. The NCEP/NCAR 40-year reanalysis project. Bulletin of the American Meteorological Society 77: 437–471. Kulkarni A, von Storch H. 1995. Monte Carlo experiments on the effect of serial correlation on the Mann-Kendall test of trend. Meteorologische Zeitschrift 4: 82–85. Lean J, Rowntree PR. 1997. Understanding the sensitivity of a GCM simulation of Amazonian deforestation to the specification of vegetation and soil characteristics. Journal of Climate 10: 1216–1235. Leith CE. 1973. The standard error of time-average estimates of climatic means. Journal of Applied Meteorology 12: 1066–1069. Lepage Y. 1971. A combination of Wilcoxon’s and Ansari–Bradley’s statistics. Biometrika 58: 213–217. Marengo JA. 1992. Interannual variability of surface climate in the Amazon basin. International Journal of Climatology 12: 853–863. Marengo JA. 1995. Variations and changes in South American streamflow. Climatic Change 31: 99–117. Marengo JA, Nobre CA. 2001. General characteristics and variability of climate in the Amazon basin and its links to the global climate system. In The Biogeochemistry of the Amazon Basin, McClain M, Victoria R, Richey J (eds). Oxford University Press: New York; 17–41. Marengo JA, Tomasella J, Uvo CR. 1998. Trends in streamflow and rainfall in tropical South America: Amazonia, eastern Brazil, and northwestern Peru. Journal of Geophysical Research 103: 1775–1783. Marengo JA, Cavalcanti IFA, Satyamurty P, Bonatti J, Nobre C, Pezzi L, Sampaio G. 2000. Ensemble simulation of interannual climate variability using the CPTEC/COLA global circulation model for the period 1982–1991. In Sixth International Conference on Southern Hemisphere Meteorology and Oceanography, Santiago, Chile. Matsuyama H. 1992a. Are precipitation data in Peru really reliable? — The comparison between World Monthly Surface Station Climatology and original data obtained from hydrometeorological agency of Peru. Journal of Japan Society of Hydrology & Water Resources 5(2): 30–33 (in Japanese with English abstract). Matsuyama H. 1992b. The water budget in the Amazon river basin during the FGGE period. Journal of the Meteorological Society of Japan 70: 1071–1084. Nobre P, Shukla J. 1996. Variations of sea surface temperature, wind stress, and rainfall over the tropical Atlantic and South America. Journal of Climate 9: 2464–2479. Nobre CA, Sellers PJ, Shukla J. 1991. Amazonian deforestation and regional climate change. Journal of Climate 4: 957–988. Obregon GO, Nobre CA. 1990. Principal component analysis of precipitation fields over the Amazon river basin. Climanálise 5(7): 35–46. Oki T, Musiake K, Matsuyama H, Masuda K. 1995. Global atmospheric water balance and runoff from large river basins. Hydrological Processes 9: 655–678. Peterson TC, Vose RS. 1997. An overview of the Global Historical Climatology Network temperature data base. Bulletin of the American Meteorological Society 78: 2837–2849. Poveda G, Mesa OJ. 1997. Feedbacks between hydrological processes in tropical South America and large-scale ocean–atmosphere phenomena. Journal of Climate 10: 2690–2702. Press WH, Flannery BP, Teukolsky SA, Vetterling WT. 1989. Numerical Recipes, The Arts of Scientific Computing Fortran Version. Cambridge University Press: New York; 702 pp. Restrepo JD, Kjerfve B. 2000. Magdalena river: interannual variability (1975–1995) and revised water discharge and sediment load estimates. Journal of Hydrology 235: 137–149. Reynolds RW. 1988. A real-time global sea surface temperature analysis. Journal of Climate 1: 75–86. Richey JE, Nobre C, Deser C. 1989. Amazon river discharge and climate variability: 1903 to 1985. Science 246: 101–103. Ropelewski CF, Halpert MS. 1987. Global and regional scale precipitation patterns associated with the El Niño/southern oscillation. Monthly Weather Review 115: 1606–1626. Rudolf B, Hauschild H, Rueth W, Schneider U. 1994. Terrestrial precipitation analysis: operational method and required density of point measurements. In NATO ASI Series, vol. 126, Global Precipitation and Climate Change, Desboism M, Désalmand F (eds). Springer–Verlag: Berlin; 173–186. Salati E. 1987. The forest and the hydrological cycle. In The Geophysiology of Amazonia, Dickinson RE (ed.). Wiley: New York; 91–101. Servain J. 1991. Simple climatic indices for the tropical Atlantic Ocean and some applications. Journal of Geophysical Research 96: 15 137–15 146. Smith WHF, Wessel P. 1990. Gridding with continuous curvature splines in tension. Geophysics 55: 293–305. Spangler WML, Jenne RL. 1990. World Monthly Surface Station Climatology and associated data sets. National Center for Atmospheric Research Dataset DS570.0: Boulder; 20 pp. Spencer RW. 1993. Global oceanic precipitation from the MSU during 1979–1991 and comparisons to other climatologies. Journal of Climate 6: 1301–1326. Tanimoto Y, Xie SP. 1999. Ocean–atmosphere variability over pan-Atlantic basin. Journal of the Meteorological Society of Japan 77: 31–46. Tomita T, Yasunari T. 1993. On the two types of ENSO. Journal of the Meteorological Society of Japan 71: 273–284. Trenberth KE, Guillemot C. 1998. Evaluation of the atmospheric moisture and hydrological cycle in the NCEP/NCAR reanalysis. Climate Dynamics 14: 213–231. Uvo CB, Repelli CA, Zebiak SE, Kushnir Y. 1998. The relationships between tropical Pacific and Atlantic SST and northeast Brazil monthly precipitation. Journal of Climate 11: 551–562. Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002) TROPICAL SOUTH AMERICA RAINFALL VARIABILITIES 195 Vose RS, Heim R, Schmoyer RL, Karl TR, Steurer PM, Eischeid JK, Peterson TC. 1992. The Global Historical Climatology Network: long-term monthly temperature, precipitation, sea level pressure, and station pressure data. ORNL/CDIAC-53, NDP-041. Carbon Dioxide Information Analysis Center, Oak Ridge National Laboratory: Oak Ridge; 99 pp + appendices. Wagner RG. 1996. Decadal-scale trends in mechanisms controlling meridional sea surface temperature gradients in the tropical Atlantic. Journal of Geophysical Research 101: 16 683–16 694. Wilheit TJ, Chang ATC, Chiu LS. 1991. Retrieval of the monthly rainfall indices from microwave radiometric measurements using probability distribution functions. Journal of Atmospheric and Oceanic Technology 8: 118–136. Xie P, Arkin PA. 1996. Analyses of global monthly precipitation using gauge observations, satellite estimates, and numerical model predictions. Journal of Climate 9: 840–858. Xie P, Arkin PA. 1997. Global precipitation: a 17-year monthly analysis based on gauge observations, satellite estimates, and numerical model outputs. Bulletin of the American Meteorological Society 78: 2539–2558. Xie P, Arkin PA. 1998. Global monthly precipitation estimates from satellite-observed outgoing longwave radiation. Journal of Climate 11: 137–164. Xie P, Rudolf B, Schneider U, Arkin PA. 1996. Gauge-based monthly analysis of global land precipitation from 1971 to 1994. Journal of Geophysical Research 101: 19 023–19 034. Yonetani T. 1992. Discontinuous changes of precipitation in Japan after 1900 detected by the Lepage test. Journal of the Meteorological Society of Japan 70: 95–104. Yonetani T. 1993. Detection of long term trend, cyclic variation and step-like change by the Lepage test. Journal of the Meteorological Society of Japan 71: 415–418. Zeng N. 1999. Seasonal cycle and interannual variability in the Amazon hydrologic cycle. Journal of Geophysical Research 104: 9097–9106. Copyright  2002 Royal Meteorological Society Int. J. Climatol. 22: 175–195 (2002)