Subregional precipitation climate of the Caribbean and relationships with ENSO and NAO

Authors


Abstract

[1] Thirty-five meteorological stations encompassing the Caribbean region (Cuba, Bahamas, Jamaica, Dominican Republic, Puerto Rico, US Virgin Islands, St. Maarten, and Barbados) were analyzed over the time interval 1951–1981 to assess regional precipitation patterns and their relationships with the North Atlantic Oscillation (NAO) and El Niño-Southern Oscillation (ENSO). Application of factor analysis to these series revealed the existence of four geographically distinct precipitation regions, (C1) western Cuba and northwestern Bahamas, (C2) Jamaica, eastern Cuba, and southeastern Bahamas, (C3) Dominican Republic and northwestern Puerto Rico, and (C4) eastern Puerto Rico, US Virgin Islands, St. Maarten, and Barbados. This regionalization is related to different annual cycles and interannual fluctuations of rainfall. The annual cycle is more unimodal and largest in the northwest Caribbean (C1) and becomes increasingly bimodal toward lower latitudes (C4) as expected. Year-to-year variations of precipitation are compared with two well-known climatic indices. The ENSO relationship, represented by Niño 3.4 sea surface temperatures (SST), is positive and stable at all lags, but tends to reverse over the SE Caribbean (C4) in late summer. The NAO influence is weak and seasonally dependent. Early summer rainfall in the northwest Caribbean (C1) increases under El Niño conditions. Clusters 2 and 3 are less influenced by the global predictors and more regional in character.

1. Introduction

[2] The Caribbean Islands form an archipelago stretching southeastward in an arc between Florida and eastern Venezuela. They are divided into the Greater Antilles (Jamaica, Cuba, Hispañola, and Puerto Rico) and the Lesser Antilles (an island chain extending from the Virgin Islands in the north to Trinidad and Tobago in the south). In most parts of the Caribbean, precipitation is bimodal with an initial maximum around May, a relative minimum in June–August, and a second peak in September–October [Rudloff, 1981; Giannini et al., 2000; Chen and Taylor, 2002]. This general tendency differs in the southernmost Caribbean (off the Venezuelan coast), which receives less rainfall and has a primary rain season in the winter (November–January; Martis et al. [2002]).

[3] Caribbean islands are especially vulnerable to natural phenomena such as hurricanes, droughts, tsunamis, and earthquakes. Development pressure in low-lying coastal areas is especially high. Therefore knowledge of precipitation patterns and variability in the Caribbean is essential in policy and planning decisions.

[4] A number of studies have analyzed Caribbean precipitation patterns and recognized the atmospheric and oceanic influence from the North Atlantic and eastern Pacific. The El Niño-Southern Oscillation (ENSO) is a dominant mode of coupled atmosphere-ocean variability on interannual timescales [Trenberth and Stepaniak, 2001]. The frequency and severity of storms, droughts, and floods in different parts of the world have been shown to be related to ENSO through atmospheric and oceanic circulation responses known as teleconnections [Dai et al., 1998]. The relationship between ENSO events and Caribbean rainfall anomalies has been discussed in numerous papers [Hastenrath, 1976; Ropelewski and Halpert, 1986; Rogers, 1988; Enfield, 1996; Malmgren et al., 1998; Enfield and Alfaro, 1999; Giannini et al., 2000, 2001a, 2001b; Waylen and Quesada, 2001; Chen and Taylor, 2002; Taylor et al., 2002; Martis et al., 2002; Spence et al., 2004]. For example, Jamaica experiences an increase in rainfall during May–July of the year following a warm ENSO event [Chen et al., 1997], whereas Cuba receives more rain earlier in the season (January–March) [Giannini et al., 2000].

[5] The North Atlantic Oscillation (NAO) is another large-scale mode of variability in the climate system known to influence precipitation patterns across the North Atlantic winter (December through March) over northwestern Europe [Hurrell, 1995; Uvo, 2003], the Mediterranean region [Hurrell, 1995; Rodó et al., 1997], North Africa [Lamb et al., 1997], and as far away as southeastern Africa [McHugh and Rogers, 2001]. NAO teleconnections via the North Atlantic anticyclone and its trade winds are thought to be responsible for NAO-influenced precipitation anomalies with the Caribbean [Malmgren et al., 1998; George and Saunders, 2001; Giannini et al., 2001a, 2001b].

[6] In the present study, precipitation time series from 35 stations in the Antilles (Cuba, Bahamas, Jamaica, Dominican Republic, Puerto Rico, US Virgin Islands, St. Maarten, and Barbados) spanning 1951–1981 have been analyzed using factor analysis. One objective is to identify distinct subregions marked by common patterns of annual and interannual fluctuations, the geographical coherency of these regions, and differences in the precipitation variability among them. We offer some explanations as to why the clusters occur as they do. The second objective is to investigate the response of these subregions to ENSO and NAO, both seasonally and continuously, as a step toward long-term prediction.

2. Data and Methods

[7] Precipitation data from 143 climate stations in the Caribbean were initially available for the study. Many of these stations are of a limited temporal scope (less than 10 years), are not complete over longer time periods, or comprise vastly different periods of time. A total of 35 stations covering the interval 1951–1981 (31 years) best achieve a balance between geographical coverage and temporal completeness. (A similar period was used in the study by Giannini et al. [2000]). The locations of these stations are mapped in Figure 1, and Table 1 provides details on their locations. Five stations are from Cuba, two are from Jamaica, six are from the Dominican Republic (D.R.), twelve are from the Bahamas, and seven are from Puerto Rico (P.R.). The islands of the US Virgin Islands (USV.I.), St. Maarten, and Barbados each contributed one station.

Figure 1.

Locations of the 35 stations included in the analysis (red dots); details are provided in Table 1. Geographic extent of the four precipitation clusters identified by factor analysis. Annual cycles left to right for clusters 1 to 4, based on averages over the 31-year period.

Table 1. Stations Included in the Studya
StationLatitude (°N)Longitude (°W)Station NameIsland% Missing Observations
  • a

    DR, Dominican Republic; USVI, US Virgin Islands; PR, Puerto Rico.

123.1782.35CasablancaCuba0.0
221.5777.85CamagueyCuba0.8
322.5583.30Paso Real de San DiegoCuba0.3
421.9079.50Sancti SpiritusCuba1.6
519.9075.15Guantanamo BayCuba5.1
617.9076.80KingstonJamaica0.0
718.4077.60Clarks Town (Trelawny)Jamaica6.2
818.4069.90Santo DomingoDR1.6
919.7371.43Villa VazquezDR4.0
1019.8070.60Puerto PlataDR5.9
1119.5070.70SantiagoDR4.0
1219.6369.90CabreraDR0.5
1318.7569.02SeyboDR0.3
1425.1077.40Nassau IABahamas0.8
1526.7078.97West EndBahamas12.1
1626.7577.30Green Turtle CayBahamas10.8
1725.7279.30Alice TownBahamas11.6
1825.5076.63Dunmore TownBahamas0.0
1924.3075.52The BightBahamas5.1
2024.0574.52Cockburn TownBahamas5.4
2123.5275.78GeorgetownBahamas0.8
2223.1074.98Clarence TownBahamas0.0
2322.1875.72Duncan TownBahamas3.8
2422.3772.97Abraham BayBahamas1.9
2520.9373.68Matthew TownBahamas0.5
2617.7064.80Alex HamiltonUSV.I.0.0
2718.0563.12Juliana AeroSt. Maarten1.3
2813.1059.50Grantley/Seawell APBarbados1.3
2918.1766.15CidraPR8.3
3018.3867.15ColosoPR1.3
3117.9766.12GuayamaPR3.8
3218.4767.07IsabelaPR5.1
3318.4266.43ManatiPR14.5
3418.2567.15MayagüezPR5.9
3518.5066.00San JuanPR1.1

[8] Each time series consists of 372 data points (31 years × 12 months). The percentages of missing observations are shown for each station in Table 1. More than half of the stations (19) have complete or nearly complete series of observations. The median percentage of missing observations is 1.6 (the mean value is 3.6%). Five stations (Casablanca, Kingston, Dunmore Town, Clarence Town, and Alex Hamilton) have no missing observations. Fourteen stations are nearly complete (missing less than 2% of the observations; 7 data points). Among the stations with more than 4% of missing data, eight series lack data for an entire calendar year (Villa Vazquez, Puerto Plata, and Santiago, D.R.; The Bight, Cockburn Town, Duncan Town, Bahamas, Isabela and Mayagüez). The four stations with more than 10% missing observations (West End, Green Turtle, Alice Town, and Manati) lack data for three calendar years, but none of these series has more than 1 year of missing consecutive monthly observations throughout the time interval considered. In case of missing observations, the governing principle was to replace the missing observation by either temporally adjacent values or the climatological mean for the respective month.

[9] We employed factor analysis (FA) in order to analyze regional patterns of fluctuations among the 35 station time series over the 31-year time interval. In contrast to the standard multivariate ordination method of principal component analysis (PCA), which represents maximum variance of the data in a minimum number of new variables (the principal components), the factor-analytical model is instead directed toward the reproduction of the interrelations (in this case the correlation coefficients) between pairs of variables in a minimum number of factors.

[10] We used the LISREL 8.52 software [Jöreskog and Sörbom, 1996] to perform the FA. The LISREL factor model used was estimated using the two-stage least squares (TSLS) method [Jöreskog et al., 2000]. TSLS-based FA is a noniterative and very fast mathematical procedure [Jöreskog et al., 2000]. A further advantage of TSLS-based FA is the possibility to compute standard errors for the factor loadings; standard errors can be obtained for the loadings of all variables, except for the so-called “reference variable” (the variable with the largest loading for a factor). These standard errors allow computations of a t value (loading/standard error) that can be used to assess the significance of the contribution of a variable to a particular factor [Jöreskog et al., 2000]. The t values higher than 2 indicate a significant loading.

[11] In order to establish a common scale of variability, we used the correlation matrix as the basis for the LISREL factor model. In order to enhance the interpretability of the factor loadings, the initial orientations of the factor axes were subjected to an oblique rotation. In LISREL, the factor scores are computed by an extension of a formula given by Anderson and Rubin [1956]. These factor scores are unbiased estimates of the factors, and their sample covariance matrix is equal to the estimated covariance or correlation matrix of the reference variable factors [Jöreskog et al., 2000]. The result of factor analysis was the creation of spatial clusters, with orthogonal (uncorrelated) time scores. To intercompare adjacent regions and their climatic drivers, we sought to reconstruct time series to reflect the true (possibly correlated) climate. For this, we correlated the resultant factor scores with the individual rainfall time series falling into each cluster. We then reconstructed each cluster's time series by averaging the rainfall data of each station with a weighting derived from the correlation value corresponding with the factor score.

[12] We analyzed the annual cycle by averaging by month over 31 years and then removed this to study the residual fluctuations. To reduce noise in the subsequent analyses, we smoothed the residual series with a 5-month running mean. We used the NAO index described by Jones et al. [1997] and an ENSO index based on Pacific sea surface temperature (SST) in the 3.4 region; (5°S–5°N, 170°–120°W; Trenberth and Stepaniak [2001]) to find associations with the Caribbean rainfall time series. Data were collected from http://www.cgd.ucar.edu/cas/indices, and a 5-month running mean smoothing was applied to the residual Niño 3.4 and NAO data. Before analyzing the residuals, the time series were linearly detrended to remove “drift.” In order to avoid erroneous inferences, statistical significance at the 90% confidence limit was considered to be attained if the correlation value exceeded 0.18 for a degrees of freedom of 74, for example, the record length minus 1 divided by the smoothing value of 5 [Benjamini and Hochberg, 1995; Katz, 2002; Ventura et al., 2004]. This helps account for natural or sampling-induced autocorrelation in the record. We also considered the seasons independently, using the preceding November to March values of Niño 3.4 and NAO and the following March to July (labeled May) and August to December (October) averaged values of the weighted rainfall for each cluster. These seasons were chosen on the basis of the seasonal peaks of rainfall that tend to occur across all clusters.

[13] Cross-correlation was employed to understand the degree of association. For the seasonal data with 30 degrees of freedom, statistical significance was achieved when the absolute value exceeded 0.30. We employed MATLAB software to generate the results.

[14] Once the relationship between time series and climatological indices was established, spatial correlations were mapped in the same time period using gridded National Centers for Environmental Prediction (NCEP) precipitable water and the two reference variables, Nino 3.4 SST and the NAO. In this way, an independent analysis of the influence of these two climatic signals was performed. This was done for different seasons at various lags, using the Website http://www.cdc.noaa.gov/Correlation.

3. Results

3.1. Precipitation Patterns From Factor Analysis

[15] The number of clusters retained was computed by determining the number of residuals with an absolute value >0.05 for n number of factors (Figure 2). This resulted in a 4-factor solution with 85% of residual correlation coefficients (the difference between the coefficients reproduced by the factor model and the actual correlations) being less than ∣0.05∣. An additional 3 factors would only have decreased the number of residuals >∣0.05∣ by 15% (Figure 2). The 4-factor solution also provides a geographically coherent pattern. Table 2 shows the estimated rotated factor loadings using TSLS and the t values for each factor loading. In the interpretation of the factor structure, a station is considered to be contributing to the factor for which the loading is highest. The corresponding t values for the stations contributing to other factors are also listed in Table 2. The t values >2 are statistically significant.

Figure 2.

Decrease in percentage of residuals (portion of correlations not reproduced by a particular factor model) >∣0.05∣ for increasing number of factors, as a justification for clustering the rainfall data into four groups.

Table 2. Factor Loadings Estimate Using TSLS for the 35 Stations Along the First Four Promax-Rotated Factors (F)a
StationF1F2F3F4
  • a

    A station is considered to contribute to the factor for which the loading is highest (marked in bold). The t values for these loadings (they should be >2 to be significant) are in parentheses. The highest loading in each factor is not associated with a t value, since it is the reference variable in the computation of the LISREL factor model.

10.54 (7.6)0.20−0.18−0.03
20.61 (9.6)0.23−0.020.02
30.58 (8.4)0.06−0.110.08
40.840.000.000.00
50.090.60 (7.4)0.19−0.09
60.180.45 (6.2)0.230.05
7−0.010.41 (5.9)0.320.04
80.40 (6.4)−0.010.290.34
90.010.100.43 (5.5)0.05
10−0.410.100.53 (6.9)−0.14
110.000.000.770.00
12−0.09−0.070.53 (6.7)0.24
130.140.070.48 (7.2)0.31
140.62 (9.7)0.09−0.200.19
150.58 (8.0)−0.05−0.230.09
160.56 (7.8)0.14−0.200.08
170.72 (10.1)0.03−0.120.02
180.56 (8.4)0.40−0.23−0.05
190.450.51 (6.7)−0.11−0.06
200.270.63 (8.0)−0.05−0.07
210.48 (7.8)0.40−0.040.06
220.240.68 (9.3)0.05−0.06
230.000.820.000.00
24−0.030.58 (7.3)0.130.01
25−0.100.60 (7.7)0.200.06
26−0.160.050.240.61 (8.8)
27−0.18−0.020.220.62 (8.4)
280.09−0.050.050.42 (5.5)
290.00−0.250.170.74 (10.0)
300.64 (9.2)−0.240.230.26
310.000.000.000.89
320.23−0.090.49 (6.5)0.20
33−0.22−0.250.380.49 (5.6)
340.48 (7.4)−0.180.050.46 (6.7)
350.09−0.280.400.57 (7.6)

[16] Twelve stations (Table 1), 1–4, 8, 14–18, 21, 30, and 34, contribute to cluster 1. These stations occur primarily in western Cuba and the northwestern Bahama Islands (Figure 1) although there are some outliers. Cluster 2 is composed of 9 stations from Jamaica, eastern Cuba, and the southeastern Bahamas. The stations from the northern Dominican Republic and northwestern Puerto Rico make up cluster 3, while cluster 4 is made up of stations from eastern Puerto Rico, US Virgin Islands, St. Maarten, and Barbados.

[17] Since the factors are formed from the raw data, they are likely to reflect differences in both the annual cycle and interannual fluctuations. Annual mean patterns for clusters 1–4 in the period 1951–1981 are given in Figure 1. The clusters can be distinguished from each other by their seasonal precipitation cycle and the timing of the summer precipitation maximum (as discussed below). For cluster 1 in the northwest Caribbean, there is a significant rainfall maximum in the early summer around May. This slowly declines in a southeasterly direction through clusters 2 and 3 to be replaced by a late-summer maximum around October for cluster 4. The following statistics were computed in order to better elucidate differences in cluster characteristics:

[18] 1. Total annual precipitation.

[19] 2. Evenness of monthly precipitation. This variable is expressed as the standard deviation of the difference between the monthly contributions to the total annual precipitation and the mean monthly contribution (100/12 = 8.33).

[20] 3. Contribution of various seasonal rainfalls to total annual precipitation (measured in percent).

[21] Table 3 provides means, ranges, and 95% confidence intervals (CI) for these variables in each cluster.

Table 3. Statistics Showing Some Fundamental Differences Among the Precipitation Clustersa
 Cluster 1Cluster 2Cluster 3Cluster 4
  • a

    Evenness of monthly precipitation is determined as the standard deviation of monthly contributions (%) to the mean (M, mean; R, ranges; CI, 95% confidence intervals).

(1) Total Annual Precipitation, mmM, 1403M, 872M, 1411M, 1308
R, 1018–2060R, 608–1392R, 687–1942R, 999–1585
CI, 1225–1581CI, 684–1061CI, 886–1937CI, 1095–1521
(2) Evenness of Monthly PrecipitationM, 4.80M, 4.87M, 3.08M, 3.16
R, 3.62–6.05R, 2.70–6.28R, 2.27–4.29R, 2.07–4.67
CI, 4.37–5.24CI, 4.03–5.72CI, 2.32–3.84CI, 2.35–3.98
(3) Contribution of Jan–Apr Precipitation to Annual Precipitation, %M, 15M, 17M, 26M, 20
R, 11–21R, 10–22R, 21–36R, 13–27
CI, 14–17CI, 14.1–20.0CI, 20.4–31.6CI, 16–24
(4) Contribution of May–July Precipitation to Annual Precipitation, %M, 37M, 27M, 26M, 26
R, 32–44R, 23–31R, 16–30R, 23–30
CI, 34–39CI, 24–29CI, 20–32CI, 23–29
(5) Contribution of Aug–Oct Precipitation to Annual Precipitation, %M, 39M, 40M, 26M, 35
R, 36–42R, 29–48R, 16–35R, 25–43
CI, 38–40CI, 36–44CI, 20–33CI, 30–40
(6) Contribution of Nov–Dec Precipitation to Total Annual Precipitation, %M, 9M, 16M, 22M, 19
R, 5–11R, 10–23R, 15–32R, 14–22
CI, 8–10CI, 13–19CI, 15–28CI, 17–22
(7) Contribution of May–Dec Precipitation to Total Annual Precipitation, %M, 64M, 62M, 55M, 80
R, 59–66R, 58–68R, 48–59R, 73–87
CI, 62–65CI, 60–64CI, 51–59CI, 76–84

[22] The stations constituting cluster 1 are characterized by relatively higher contributions of May–July rainfall (variable 4; mean of 37%) and relatively lower contributions of November–December rainfall (variable 6; 9%) than the stations from clusters 2–4 (26–27 and 16–22%, respectively; Figure 1). As seen in Table 3, the confidence intervals for cluster 1 (for variables 4 and 6) do not overlap those of the other clusters. Being further north and on the edge of the tropical rainfall belt, this cluster exhibits a larger and more unimodal annual cycle. Clusters further to the southeast are situated at lower latitude, where the rainy season tends to spread across the year.

[23] The mean total annual rainfall in cluster 2 (872 mm) is statistically significantly lower than in clusters 1 and 4 (1403 and 1308 mm, respectively). The total rainfall in cluster 3 is highly variable, and the mean (1411 mm) is indistinct from the means of the other clusters. Cluster 2 is furthermore distinguished from cluster 3 by the tendency for a more even distribution of rainfall throughout the year as evidenced by the evenness statistic (3.08 and 4.87, respectively; Table 3).

[24] Cluster 3 is marked by higher contributions of January–April rainfall (26%), and lower contributions of August–October rainfall (26%) than clusters 1 and 2 (15–17% and 39–40%, respectively). Cluster 3 exhibits more rainfall throughout the year than clusters 1 and 2 (3.08 for cluster 3 relative to 4.80–4.87 for clusters 1 and 2). This cluster can be distinguished from cluster 4 on the basis of the lower contribution of May–December rainfall (55% relative to 80%).

[25] The seasonal variation in cluster 4 is marked by low precipitation amounts through the winter and early spring (January–April), followed by increasing rainfall amounts from May onward with a maximum peak in October. The mean contribution of precipitation over the 8-month period May–December to the total annual is 80% in this cluster, which is significantly higher than in the other clusters (55–64%, Table 3). The patterns of rainfall in the northwest Caribbean (clusters 1 and 2) is more variable than in the southeast (clusters 3 and 4). This evenness is also manifested by a greater contribution of January–April rainfall in cluster 3 to the annual total than in clusters 1 and 2.

3.2. Interannual Fluctuations

[26] The reconstructed time series for the four clusters are shown in Figure 3. These are based on station data averaged with a weighting dependent on the correlation with the factor-loading time series. The residuals from the annual mean are shown for each cluster in Figure 4 based on the reconstructed series, together with their respective wavelet spectrum. The ENSO and NAO variables were cross-correlated against each cluster, and those with the highest values (for example, Niño 3.4 SST and cluster 1 residuals; NAO and cluster 4 residuals) were plotted. Each of the residual series has a unique cycle of fluctuations as identified by wavelet analysis in the right panels of Figure 4. In general, the variability is most coherent in clusters 1 and 3. This is shown by the concentration of spectral energy in specific periods over a length of years. Clusters 2 and 4 are less coherent, and their wavelet spectra are more chaotic. Cluster 1 exhibits cycles around 5 years early in the record and around 3 years from 1961–1975 that are largely in phase with ENSO. Cluster 2 has weaker cycles drifting between 3 to 4 years in the 1960s and 1970s. Cluster 3 also exhibits 2- to 4-year cycles in the 1950s and 5-year cycles in the 1970s. Cluster 4 has 2-year cycles throughout the record, and a 4- to 5-year cycle in 1970s. The drift of spectral power across various frequencies reflects a chaotic interaction of global and regional climate signals, some of which are multidecadal in nature (for example, NAO) and may not appear in our relatively short detrended record.

Figure 3.

Reconstructed time series for the four clusters (C1–C4) based on station data averaged with a weighting dependent on the correlation with the factor loading time series. Note the larger annual cycle in cluster 1, situated at the higher latitude.

Figure 4.

Fluctuations in weighted precipitation for each cluster after removal of the annual cycle, cluster 1 (top left) to cluster 4 (bottom left). Wavelet spectrum for each cluster in right panels, with enclosed areas significant with respect to a red noise background. Nino 3.4 SST anomalies are plotted with the cluster 1 time series (upper left), and NAO anomalies with cluster 4 (lower left).

3.3. Relationships to NAO and ENSO

[27] The cross-correlation functions are plotted in Figure 5 with respect to the continuous smoothed monthly residuals. The NAO exerts some influence on the southeastern Caribbean (cluster 4) with a weak positive value that almost achieves statistical significance at zero lag. On the other hand, ENSO mostly influences the northwest Caribbean (cluster 1) with stronger positive values that achieve statistical significance across all lags from −6 to +6 months.

Figure 5.

Lag correlation function for ENSO (Nino 3.4 SST) and C1 residuals (light shading) and NAO and C4 residuals (dark shading) based on continuous data. Absolute values >18% are significant.

[28] If the different rainfall seasons are individually considered with respect to preceding winter season (December to February) values of the climatic indices, the ENSO signal continues to show a significant positive association in the NW Caribbean (Figure 6) with a value as high as 43% for early season rainfall (May–March to July) in cluster 1 (a 3-month lead time). Associations with other clusters are also positive. For the summer/fall seasonal rainfall (October–August to December; 9-month lead), the associations are still positive in the NW and become weakly negative in the SE Caribbean. The NAO, on the other hand, is negatively related in both early and late rainy season across all clusters. This is most significant for clusters 2 (−41%), 3, and 4 in the May season and less so for clusters 1 (−31%), 2, and 3 in the October season. The negative association between NAO and Caribbean rainfall has been found in earlier studies [Malmgren et al., 1998; George and Saunders, 2001; Charlery et al., 2006] using seasonal data and suggests that there is a preferred season (winter) of influence. However, its lack of persistence and the change of association in different seasons suggest instability.

Figure 6.

Preceding winter ENSO (upper) and NAO (lower) influence on seasonal (“May” = AMJ and “October” = SON) rainfall for the different clusters. Absolute values >30% are significant. The lead time for May is 3 months, for October is 9 months.

[29] When the ENSO and NAO index correlations are mapped against NCEP precipitable water in the 1951–1981 period, the ENSO signal reaches a maximum (positive) influence in the March to June season (Figure 7) with a center of action that spreads from central America, indicating that warm phase increases moisture at the beginning of the rainy season.

Figure 7.

Correlation between March to June NCEP precipitable water and Nino 3.4 SST (1951–1981). Unshaded (white) r > 0.7, dark shaded r < 0.6. Values are just as strong with ENSO leading but weaker at other lags and for other seasons.

[30] The signal is just as strong at 3 months lead time as at zero lag but is considerably weaker at 3 months lag. Hence the ENSO signal may be a useful predictor for early summer rainfall. The NAO signal does not achieve a significant value and changes sign. It is negative most of the time and positive in the October season (not shown) with a center of action spreading from the north Atlantic toward the eastern Caribbean.

4. Discussion

[31] Previous attempts have been made to establish a subregional classification of precipitation variability for the Caribbean. Using PCA, Giannini et al. [2000] divided the Caribbean into three subregions which identified the bimodal nature of Caribbean precipitation, an initial maximum precipitation peak in May–June, a relative minimum in July–August, and a second peak in September–October. Other papers [Charlery et al., 2006] also discuss the bimodality of the seasonal rainfall pattern. This general tendency differs in the southernmost parts of the Caribbean (islands along the Venezuelan coast), which have overall less precipitation and a primary rain season in the winter (October–January; Martis et al. [2002]). Some of Giannini’s clusters, PC2 and PC3, cover our cluster 4. Chen and Taylor [2002] used empirical orthogonal function (EOF) and spectral analyses to divide the Caribbean into 12 subregions. Their northern Caribbean subregion corresponds to our cluster 1, and their main Caribbean basin subregion corresponds to clusters 3 and 4.

[32] This study has found that the precipitation along the Antilles island chain in the Caribbean can be divided spatially into regional patterns composed of four distinct clusters in the time interval 1951–1981. Precipitation in the Caribbean is modulated by SST anomalies in the tropical Pacific [Hastenrath, 1978, 1984; Ropelewski and Halpert, 1986, 1989, 1996; Rogers, 1988; Enfield, 1996] and tropical North Atlantic [Enfield, 1996; Chen et al., 1997; Giannini et al., 2000, 2001a, 2001b; Chen and Taylor, 2002; Taylor et al., 2002]. Enfield and Alfaro [1999] were the first to show that rainfall in the Caribbean depends on the combined SST anomaly effects in both these tropical oceans (corroborated by Giannini et al. [2000, 2001a, 2001b] and Spence et al. [2004]). Tropical North Atlantic SSTs lag those of the equatorial Pacific by 4–5 months [Enfield and Mayer, 1997; Chen et al., 1997; Mélice and Servain, 2003]. Our study is unique to those done earlier, as we include the Bahamas and exclude Central and South American rainfall. This helps to focus our results on areas with similar annual totals and similar seasonal characteristics. Earlier work has found the Caribbean climate to be somewhat uniform, particularly in the wet season. We offer a more dynamic explanation of subregional variability that contrasts continuous and seasonal responses.

[33] The distribution of rainfall in our climatological analysis is in general agreement with the NCEP reanalysis model; however the model overestimates the mean annual rainfall amount (>1800 mm compared with ∼1400 mm here). On the other hand, the Climate Prediction Center Merged Analysis of Precipitation (CMAP) rainfall climatology, which is derived from satellite data, reveals a westward increase from low values (∼1000 mm) over the Lesser Antilles to higher values over Hispañiola. The eastern Caribbean is influenced by cooler SSTs near Africa and associated anticyclonic subsidence, and the CMAP satellite rainfall climatology is more marine in character, compared with our use of island data. Assuming a uniform rate of subsidence above the trade winds, the gradual westward increase in SST would be expected to induce a similar increase in rainfall. However, this is not as well reflected in the station data as in the satellite data, for reasons yet unknown.

[34] A significant feature in Table 3 is that cluster 2 over eastern Cuba and the Bahamas receives less rainfall than the rest of the Caribbean. This appears to result from the large-scale divergence that occurs as the anticyclonic flow splits: a Caribbean axis south of Cuba continues into Mexico, while an Atlantic axis north of Cuba recurves toward Florida (evident in NCEP reanalysis climatological wind fields). A secondary effect is the convergence that occurs over the mountains of Hispañola, creating a downstream leeward effect over eastern Cuba.

[35] To better understand the regional forcing of wet spells, we extracted the five wettest years for May rainfall in cluster 1 and October rainfall in cluster 4 and analyzed the NCEP meteorological fields. The composite average upper-level meridional wind anomaly is informative. We found increased southerly flow poleward of the area of increased rainfall, irrespective of season (not shown). This suggests that standing waves in the subtropical jet act to draw equatorial moisture across the Caribbean, a subject for further study.

[36] We have demonstrated that precipitation in some of the clusters identified here is related to ENSO and, to less extent, NAO. Clusters 1 and 2 are closest to the axis of the subtropical jet streams and affected mostly by ENSO, whereas clusters 3 and 4 are located in the oceanic trade wind zone and are influenced more by NAO. These findings are more consistent with those of Malmgren et al. [1998], George and Saunders [2001], and Giannini et al. [2001a, 2001b]. George and Saunders proposed a NAO linkage through the tropical North Atlantic trade winds that strengthen following a positive phase and are accompanied by subsidence and reduced rainfall. However, the additional heat flux from ocean to atmosphere at tropical latitudes enhances moisture uptake causing some fluctuations in NAO influence as noted here and in the studies of Giannini et al.

[37] The ENSO influence comes through the equatorward shift of the subtropical jet stream in warm phase that helps draw convective outflows northward from the Intertropical Convergence Zone (ITCZ) and South American monsoon. Subtropical jet-stream troughs enhance early- and late-summer wet spells. However, it is known that the same jet-stream activity acts to suppress Caribbean rainfall during the hurricane season [Hastenrath, 1976; Ropelewski and Halpert, 1986, 1987, 1989, 1996; Rogers, 1988]. Our research has determined that El Niño suppression of Caribbean rainfall during the hurricane season is outweighed by the positive effect over the remainder of the year (compare Figures 5 and 6).

5. Summary

[38] We have analyzed regional precipitations in the Caribbean region on the basis of consecutive monthly precipitation series from 35 meteorological stations spanning the period 1951–1981 (Figure 1). This time interval was found to provide the optimum balance between temporal and geographical coverage. Our primary results can be summarized as follows:

[39] 1. Four precipitation clusters from NW to SE (clusters 1 to 4) were identified from time series of 35 stations within the Caribbean.

[40] 2. The clusters have distinct annual cycles and residual fluctuations. A unimodal structure is apparent in cluster 1, while the others exhibit bimodal tendencies. Clusters 1 and 3 have more distinct interannual cycles, with periods around 3 years evident.

[41] 3. NAO is negatively associated with Caribbean rainfall when the preceding winter season is considered in respect to seasonal rainfall. However, a continuous monthly analysis suggests an unstable and generally positive simultaneous association with the SE Caribbean rainfall, particularly in late summer.

[42] 4. Warm ENSO events represented by Niño 3.4 SST have a generally beneficial effect on Caribbean climate, enhancing rainfall except over the SE region in late summer. An equatorward shift of the subtropical jet stream assists in the development of wet spells in early summer through enhanced upper-level cyclonic vorticity, particularly over the NW Caribbean.

[43] Further work could focus on how the station rainfall compares with more recent satellite data in terms of clustering, annual cycle, residuals, and relationships with global predictors. Specialized predictors that reflect the regional uptake of global climatic signals need to be found and tested, cluster-by-cluster and season-by-season.

Ancillary