Joint modes of climate variability across the inter-Americas


  • Mark R. Jury,

    Corresponding author
    1. Department of Physics, University of Puerto Rico at Mayagüez, PR 00681-9016, Puerto Rico
    2. Univ. Zululand, KwaDlangezwa 3886, South Africa
    • Department of Physics, University of Puerto Rico at Mayagüez, PR 00681-9016, Puerto Rico.
    Search for more papers by this author
  • Björn A. Malmgren

    1. Dept Earth Sciences, Uppsala University Villavägen 16, SE-752 36 Uppsala, Sweden
    Search for more papers by this author


Surface temperature (Ts), sea-level pressure (SLP), and zonal wind (U) fields from NCEP/NCAR reanalysis spanning the period 1949–2006 have been jointly analysed by means of principal component analysis to assess the dominant modes of climate variability in the inter-American region, including the eastern Pacific and Atlantic Oceans. Two aspects of these time-series were analysed: the annual cycle and its residual. Seasonal modes 1 and 2 take the form of north-south dipoles of Ts over oceanic and terrestrial environments. Seasonal mode 3 derives from SLP changes over the Amazon with a Caribbean wind response. Temporal fluctuations in the residual (ANOM) fields are resolved by six modes. ANOM1–3 is dominated by SLP and Ts changes across the equator in the Atlantic and east Pacific. The ANOM4 pattern is an isolated wind mode in the NE Atlantic. ANOM5 and 6 are dominated by sea temperature patterns in the South Atlantic and in the path of African easterly waves crossing the tropical Atlantic. The residual time scores are marked by significant warming (Ts) trends and co-located lower pressure. Cycles in the anomaly time scores tend to occur at 2- to 5-year periods. Modes dominated by atmospheric variables exhibit greater high-frequency ‘noise’ than oceanic modes. Predictability is assessed through development of multivariate regression models trained on Caribbean rainfall and related target time-series. The seasonal dipole between North and South America is found to modulate Pacific El Niño-Southern Oscillation and tropical Atlantic influence on Caribbean rainfall. Copyright © 2011 Royal Meteorological Society

1. Introduction

The inter-America region is a focal point for atmospheric exchanges in the cross-equatorial axis between North and South American monsoons and in the along-equatorial axis between the tropical Pacific and Atlantic Oceans. In the Caribbean Sea and Gulf of Mexico, surface temperatures rise above 28 °C from July to October each year, instilling high convective energy and attracting hurricanes (Wang and Enfield, 2001, 2003; Wang et al., 2006; Wang and Lee, 2007). But along the northern coast of Venezuela and Columbia, there is an axis of strong trade winds that drive ocean upwelling, inhibiting convection across the Antilles Islands particularly in mid-summer (Granger, 1985; Amador, 1998; Magaña et al., 1999; Amador et al., 2000; Jury et al., 2007; Muñoz et al., 2008). The Atlantic and east Pacific inter-tropical convergence zones (ITCZ) oscillate north-south following the sun and surface heating of North and South America land masses (Horel et al., 1989; Higgins and Shi, 2001; Liebmann and Marengo, 2001; Nogués-Paegle et al., 2002; Vera et al., 2005). A zone of subsidence extends across the inter-Americas during boreal winter, preventing northward outflows from the South American monsoon. This breaks down in May and October as surface temperatures rise in the west Atlantic and trade winds weaken. Here, we will attempt to decompose these important features through statistical analysis of historical data.

This study examines climatic processes of the inter-Americas using principal component analysis (PCA) to determine regional to hemispheric scale co-variability in monthly surface temperature, air pressure and zonal wind (U). The seasonality is analysed, then filtered, to study year-to-year fluctuations, using methods and making interpretations similar to Enfield and Alfaro (1999) among others. We evaluate the amplitude, frequency and spatial coherence of the leading modes across the inter-Americas and their significance in the forcing of Caribbean climate. The following scientific questions are addressed: (1) Does the annual cycle decompose into terrestrial and marine modes represented as north-south (N-S) dipoles? (2) Do the leading inter-annual modes derive from Pacific El Niño-Southern Oscillation (ENSO) and N-S features across the Atlantic? (3) How important are the west Atlantic warm pool and Caribbean low-level jet (CLLJ)? and (4) How useful are the PCA time scores in predicting Caribbean Antilles rainfall, hurricanes, and related climate sensitive resources? These questions are intended to provide a framework to our research, to elucidate the ‘cross-roads problem’ that affects Caribbean climate, with potentially conflicting influences from terrestrial monsoons and oceans with east-west and north-south gradients that oscillate at different frequencies, as outlined by Enfield and Alfaro (1999).

2. Data and methods

Our analysis represents covariability of surface temperature (Ts), sea-level pressure (SLP), and U zonal wind across the tropical eastern Pacific and Atlantic Oceans, and sub-tropical parts of North and South America and West Africa, to understand the hydrostatic adjustment of pressure to temperature and the consequent kinematic response. A domain that considers both Atlantic and Pacific Oceans is necessary to gauge the covariability of environmental conditions that influence Caribbean climate (Enfield and Mayer, 1997; Enfield and Alfaro, 1999; Wang et al., 2006). We chose the variables Ts, SLP, and U because they are underpinned by an extensive database of surface ship and station observations (Woodruff et al., 1987; Kalnay et al., 1996) and have the advantage of being continuous in space and so enable comparisons of changes over land and sea. A parallel study to evaluate a number of different data sets (Jury, 2009) reveals that reanalysis of surface ship and station data is necessary to remove poor quality observations and provide a data set that is fit for the purpose of long-term PCA.

Monthly surface temperature, sea-level pressure, and 10-m U data are derived from the NCEP/NCAR reanalysis data set provided by NOAA/CDC from their website at The NCEP/NCAR reanalysis data used here consist of time-series spanning the period January 1949 through May 2006 (689 months) in the latitudes 20°S and 35°N and longitudes 140°W and 20°E. The original 2.5° gridded data were smoothed into 5° latitude by 5° longitude. The data matrix consisting of 689 consecutive maps each containing 352 grid values for three variables (Ts, SLP, and U) was subjected to PCA to explore the essential relationships among the monthly time-series, using methods consistent with earlier studies (Wallace and Dickinson 1972; Weare and Nasstrom, 1982; Horel, 1984; Barnston and Livezey, 1987; Kim et al., 1996a; Kim and Wu, 1999). We made separate analyses for the raw time-series with seasonal cycle retained (SEAS) and with the seasonal cycle removed (anomaly, ANOM) by subtracting the monthly averages at each grid point over the study period.

Some important goals of PCA in the context of climate analysis is to identify and extract from a geophysical data set independent patterns that help us understand the time and space scales of environmental variability (Schnur et al., 1993; Xu, 1993; Montroy, 1997), to distinguish the annual cycle driven by north-south movement of the sun, interannual fluctuations that reveal regional circulations and ocean-atmosphere interaction (e.g. ENSO), trends related to multi-decadal climate change, and determinations of the veracity of repetitious cycles and the geographic origin of climate signals (Shen et al., 1994; Kim et al., 1996b; Kim, 1997; Schwing, 2002). Interpretations depend on whether the modes rise above some measure of background ‘noise’, their spatial extent and temporal frequency, and which environmental field tends to dominate (Graham et al., 1987; Xu and von Storch, 1990; Blumenthal, 1991); modes carrying a higher signal-to-noise ratio are more useful in explaining natural variability (Hasselmann, 1988, 1993; Santer et al., 1994; North et al., 1995; North and Kim, 1995; Hegerl et al., 1996).

We used the correlation matrix as the basis for our PCA to establish a common scale of variability for the three field variables, due to their different numerical values, which is particularly evident for the SEAS series. The use of the correlation matrix rather than the covariance matrix implies a standardization of the original time-series to zero means and unit standard deviations prior to the extraction of the PCs. To enhance the interpretability of the factor loadings, the principal components were subjected to orthogonal (varimax) rotation (Richman, 1986; Cheng et al., 1995). This regionalization of climatic variability helps separate ENSO from other modes. Because PCA is designed to describe global variance in the data, individual modes tend to yield a relatively small fraction of variance in any given region (Plaut and Vautard, 1994; North et al., 1995).

The scree test was applied to determine the number of modes for interpretation (Cattell, 1966). For the seasonal cycle, the first three modes were found to have declining and thus significant variance (Figure 1(a)), while for the anomalies, the first six modes were significant (Figure 1(b)). A North et al. (1982) test showed that the PCs are independent in both analyses. For each mode, the loading patterns were considered, and it was found that one meteorological variable tended to dominate. Thus, the combined time score reflects a certain feature, and we label the modes accordingly. The loading maps represent amplitude with a potential range from − 100 to + 100. Lower values refer to PCs with lower variance or reduced covariability between the three variables. Thus, inter-comparisons were made possible.

Figure 1.

(a, b) Display the eigenvector loadings in descending order of variance explained, for the seasonal and anomaly time-series, respectively. The vertical line divides significant modes from ‘noise’. This figure is available in colour online at

We analysed temporal variability of the repetitious seasonal cycle by creating a monthly climatology over 57 years; computing the means and standard deviation by month. To reduce noise somewhat, all time-series were smoothed with a 3-month running mean prior to wavelet spectral and trend analysis. Trends in the original data were retained in the PCA to determine the level of climate change embedded within the inter-annual variability. Trends in the time scores were tested using second-order polynomial regression analysis. The R2 fit of a second-order polynomial trend was computed on the 3-month smoothed anomaly time scores, as a measure of the ratio of climate change to variability. Cyclicity in the anomaly time scores was analysed using wavelet spectral analysis, from the website based on Torrence and Compo (1998). Cross-correlations between the various PC time scores were assessed at lags − 12, − 6, 0, 6, and 12 months. For the seasonal data, significance at the 99% confidence limit is reached with equation image for ∼50 degrees of freedom. For the anomaly series, significance is reached with equation image for > 100 degrees of freedom.

Predictability was assessed via statistical hindcast fit of the leading PCA time scores to various indicators of Caribbean climate at 6-month lead time. For this, seasonal (October–December, OND-1 and March–May, MAM) time score values were extracted from the first three seasonal modes and the first six anomaly modes. Combining variables into linear multivariate regression models, we permitted < 4 uncorrelated predictors per algorithm.

The target data include station rainfall averaged into four Caribbean Antilles Island areas, based on the cluster analysis of Jury et al. (2007): (1) western Cuba (annual mean and standard deviation: 103, 53 mm/month), (2) eastern Cuba (82, 40), (3) Dominican Republic and Puerto Rico (92, 42), and (4) the Lesser Antilles of the east Caribbean (92, 45). The station data were extracted from NCDC/GHCN, and gaps were filled using gridded rainfall estimates from NCEP, based on the CMAP reconstruction technique (Xie and Arkin, 1997; Chen et al., 2002; Yin et al., 2004), available at the website In the study by Jury et al. (2007), rainfall data were truncated at 1981, but here, we employ a record extending from 1951 to 2005. The rainfall data are divided into early and late summer, MAM and August to October (ASO), respectively. This tends to distinguish wet spells driven by westerly troughs (MAM) from those contributed by easterly waves and tropical cyclones. To check whether area averages of rainfall fit the PCA mode scores better than island-scale data, we employ station data for the Dominican Republic and Barbados.

Hurricane indices were formulated from the National Hurricane Center reanalysis HURDAT data set for the period 1951–2005 in two ways: (1) by counting the number of days of hurricanes with Vmax > 33 m/s in the region southeast of Key West, FL, and (2) by counting the number of tropical cyclones (Vmax > 16 m/s) passing through the same region (e.g. 12–25°N, 30–82°W).

3. Results

3.1. Dominant modes of climate variability

Figure 1 illustrates the percentage of variance explained by the leading principal components for the seasonal (SEAS) and anomaly (ANOM) time-series. Three modes explain 64% of the total variance, whereas for the anomaly data, six modes explain about half (48%) of the variability in the time-series. We consider these modes to be significant for interpreting seasonal climate variability across the inter-Americas. This application of PCA enables a reduction of dimensionality from 1056 to three and six variables for the seasonal and anomaly time-series, respectively. For the seasonal modes, the patterns can be labelled as SEAS1 marine dipole (37.5% of variance explained), SEAS2 continental monsoon (17.7%), and SEAS3 Amazonian (8.8%). For the anomaly modes, the patterns may be labelled as ANOM1 equatorial Atlantic air pressure (14.4%), ANOM2 east Pacific sea temperature (11.9%), and ANOM3 east Pacific air pressure (8.1%), the last two linked to ENSO variability. The lesser anomaly modes include ANOM4 the northeast Atlantic wind (6.1%), ANOM5 the South Atlantic sea temperature (4.3%), and ANOM6 tropical North Atlantic sea temperature (3.4%). Although we refer to the anomaly modes by the field with strongest loading, the result naturally involves the other two fields and contributions of loadings from other areas.

The mean time scores for the seasonal modes are dominated by 12-month cycles that follow the sun for modes 2 and 3 and are shifted by ∼2 months for marine mode 1 (Figure 2(a)). The amplitude of mode 1 is naturally highest, while modes 2 and 3 have very similar mean structure and amplitude, only deviating in DJF months. The mean time score for mode 2 peaks in June–July. Mean standard deviations (Figure 2(b)) are lowest in April–May for modes 2 and 3, and remain low for continental mode 2, rising slowly to a peak in January. Marine mode 1 deviations remain high through the year and peak in northern summer.

Figure 2.

Mean seasonal climatology of PCA time scores (a) and its standard deviation (b) averaged each month over 57 years. The marine mode (1) lags the terrestrial modes (2, 3) as expected. (c) Mean Caribbean rainfall climatology averaged each month over a similar period and predicted SEAS values (dashed) based on a multivariate fit using the scores in (a). This figure is available in colour online at

The mean annual cycle of Caribbean rainfall may be reconstructed from seasonal mode scores by linear multivariate regression (Figure 2(c)). Variance (93%) is explained with a weak positive contribution from SEAS1 (0.68 coefficient), a positive contribution from SEAS2 (0.84), and a strong negative contribution from SEAS3 (−1.02). Although a reasonable third-order polynomial fits to the observed rainfall, the model fails to predict a number of features: March and August are too wet and May and June are too dry. Hence, the July ‘dip’ is lost, and instead, the model rainfall rises monotonically from a minimum in February to a maximum in September following regional trends in ocean temperatures.

3.2. The seasonal modes

The colour shaded contour maps shown in Figure 3(a)–(c) illustrate the geographical variability in the PC loadings for Ts, SLP, and U for the three seasonal modes considered here to be significant (SEAS1–3).

SEAS1 represents the oceanic annual cycle with similar contributions for each field variable (Fig. 3a). There is a N-S dipole structure as expected from changes in solar insolation. It has a greater overall time score amplitude due to the larger surface area of ocean, the more homogeneous Ts signal, and the stronger U. Naturally, the centres of action of SLP oppose Ts, while bands of U locate at the interface between SLP centres of action. The loading pattern for SLP reaches a minimum in the path of Atlantic hurricanes. The U provides horizontal shear along the suspected axis of the ITCZ along 10°N that becomes discontinuous near the Andes mountains. The strength of loading for the dominant seasonal mode is nearly as strong in the Pacific as the Atlantic.

Figure 3.

Loading maps of field variability for the combined PCA of surface temperature (Ts), air pressure (SLP), and zonal wind (U) for seasonal (SEAS) modes 1 (a), 2 (b), and 3 (c): oceanic, terrestrial, and Amazon centres of action, respectively, with N-S dipoles in first two modes. The colour bars are stretched over the maximum and minimum values per mode; thus, the influence of PCs and the contribution of different variables can be inter-compared. Neutral loadings are white and maximum and minimum values have darkest colour. This figure is available in colour online at

SEAS2 is the continental monsoon annual cycle with similar contributions for each variable (Figure 3(b)). There is an N-S dipole with positive loadings south of about 10°N and negative loadings north of this latitude. The loading structure is essentially confined to land surfaces but contains positive loadings also over the ocean northeast of South America (west of about 30°W) and negative loadings over the ocean southeast of the North American landmass and west of Africa. Its amplitude is lower than SEAS1 due to the small area affected and the heterogeneous response of the underlying land surface to seasonal inputs of heat. The centres of action in the SLP field form a triangle connecting the three continents: (1) North America, (2) Southern Amazon, and (3) West Africa. Zones of onshore flow are found west of the continental areas, in response to strong heating gradients as outlined by Barnett (1983, 1985).

SEAS3 has strongest expression over northeastern South America in the SLP field (Figure 3(c)). Ts and SLP loadings extend from a centre of action over North America into the equatorial Pacific. This is distinct from SEAS2 mode, where it is confined to the land. There is a notable loading for U in the Caribbean that represents the low-level trade wind jet axis along 15°N (CLLJ) peaking in June–July. It appears related to a centre of action in SLP over the Amazon and Atlantic ITCZ which lies to its east. It is surprising that the CLLJ does not associate with the North Atlantic anticyclone in our analysis but to the aforementioned east to west gradient in SLP. There is a U loading directed towards the West African Sahel on 15°N, as a result of SLP polarity between the Amazon and Sahara. These features exert control over the annual cycle of climate around the inter-Americas and are important in the context of predictability as outlined above. For Caribbean rainfall, terrestrial and marine annual cycles drive an increase in convection in early and late summer, respectively, while SEAS3 mode restricts moist inflows from South America in mid-summer via the CLLJ. Overall, the seasonal loading maps separate the annual cycle strength into marine and terrestrial modes 1 and 2, each playing a distinct role in the regional climate. Mode 3 appears important in linking the Amazon and Caribbean. The Atlantic Warm Pool fails to emerge as a key feature of seasonal variability here.

3.3. The inter-annual modes

Contour maps for the loadings of Ts, SLP, and U for the six significant anomaly modes (ANOM1–6) are shown in Figures 4–9. ANOM1 is represented as the equatorial Atlantic SLP mode (Figure 4). We are surprised to find an Atlantic mode that is larger than the Pacific ENSO. The centre of action is within the oceanic ITCZ, and it reflects a pressure gradient near 25°N that leads to changes in wind and SST that influence the convective environment along the path of Atlantic easterly waves and tropical cyclones (Aiyyer and Thorncroft, 2006). The ANOM1 time score is noisy and has a step near 1976; hence, SLP is higher over the Atlantic south of 15°N in the second half of the record. This is thought to be due to a low-frequency (LF) shift of climate (Jury and Enfield, 2010), not necessarily changing observations. The trend suggests a weakening of easterly winds over the North Atlantic around 30°N from the 1950s to the recent decade. Superimposed on this trend is an approximate 5-year cycle though parts of the record (mid-1970s to mid-1980s) and mid-1990s, which is most likely related to ENSO variability.

Figure 4.

Left: Geographical variability in the PC loadings for T, SLP, and U for ANOM1: equatorial Atlantic centre of action, with rising pressure trend and HF spectral energy. Upper right: Variation in time scores, with second-order polynomial trend and its R2 ‘fit’. Lower right: Wavelet analysis; dark grey indicates significant spectral energy. Hatched areas are outside of cone of significance. This figure is available in colour online at

Figure 5.

Geographical variability in the PC loadings for T, SLP, and U for ANOM2: east Pacific SST centre of action, with warming trend and 3- to 5-year spectral energy, layout as in Fig. 4. This figure is available in colour online at

Figure 6.

Geographical variability in the PC loadings for T, SLP, and U for ANOM3: east Pacific SLP centre of action, with rising trend and both 3 year and HF spectral energy, layout as in Fig. 4. This figure is available in colour online at

Figure 7.

Geographical variability in the PC loadings for T, SLP, and U for ANOM4: northeast Atlantic SLP centre of action, with weak trend and HF spectral energy, layout as in Figure 4. This figure is available in colour online at

Figure 8.

Geographical variability in the PC loadings for T, SLP, and U for ANOM5: south Atlantic SST centre of action, with warming trend and both 2–3 year and HF spectral energy, layout as in Figure 4. This figure is available in colour online at

Figure 9.

Geographical variability in the PC loadings for T, SLP, and U for ANOM6: tropical North Atlantic SST centre of action, with polynomial trend and 2- to 3-year spectral energy, layout as in Figure 4. This figure is available in colour online at

ANOM2 is the east Pacific ENSO mode expressed through tropical SST in the longitudes 80–120°W (Figure 5). r2 for the relationship between the monthly time-series of Pacific Niño3.4 SST and ANOM2 is 0.67 (p < 0.001), confirming the linkage to ENSO. The wind loading is 40° west of the SST loading. The SLP loading is weak here. There is an anti-phase response in the Atlantic that is similar to the Enfield and Alfaro (1999) pattern. The time score for ANOM2 is relatively cyclical and smooth, with a transition around 1976 to El Niño prevalence in agreement with earlier studies (Graham et al., 1987; Latif and Flügel, 1991; Latif et al., 1993). There is a warming trend and the wavelet analysis reveals that significant spectral energy is maintained in the 3- to 5-year band. ANOM3 is another expression of Pacific ENSO through the SLP field (Figure 6). This is confirmed by the r2 for the ENSO 3.4 and ANOM3 series, which is 0.21 (<0.001). It is surprising that this separates from the SST mode 2. Again there is a weak anti-phase east Atlantic response. The time score is rather noisy and potentially engaged with ANOM2 features, despite the statistical condition of orthogonality imposed by PCA. The wavelet analysis of the time score reveals 3-year spectral energy and significant high-frequency (HF) noise. The trend for ANOM3 is downward, so we find that coupled with the loading map, the air pressure has gradually increased over the eastern Pacific.

ANOM4 is an isolated mode in the NE Atlantic dominated by U and SLP (Figure 7). The time score is very noisy and non-cyclical with little trend. This mode holds significant variance, but the area of influence is quite small (15° latitude × 20° longitude). Changes in pressure and wind in the northeast Atlantic (off the coast of Morocco) are important to the upwelling process that influences downstream SST and the development of convection around African easterly waves. ANOM4 displays a significant negative cross-correlation at zero lag with NAO (r0 = − 0.53; p < 0.001). Hence, a negative NAO is associated with strong U and low SLP in the northeastern Atlantic (Figure 7). Cyclicity in the ANOM4 is weak and restricted to 2–3 cycles in the 1950s and mid-1990s.

ANOM5 is dominated by SST in the South Atlantic, with a wind loading 20° to the west (Figure 8). There is an anti-phase response in the SE Pacific. This time score has the strongest (warming) trend and some decadal cyclicity. This mode has been found, with different statistical techniques (Déqué and Servain, 1989), as an N-S Atlantic gradient or dipole mode, but here it is seen as two distinct features. The SLP loading is coincident with the SST loading and not displaced to the southwest as might be expected due to downstream hydrostatic adjustment. The wavelet analysis reveals 2- to 3-year cycles and HF noise. ANOM6 is the tropical North Atlantic SST mode [(labelled major development region (MDR)], with corresponding SLP and U responses that are located in the path of African easterly waves that form into tropical cyclones (Avila and Pasch, 1995; Figure 9). SLP loading is 20° west of SST, indicative of a fetch effect and downstream delay induced by locally strong trade winds. The time score has a U-shaped polynomial trend suggesting a single ± 50-year cycle possibly linked to the Atlantic multi-decadal oscillation (AMO; Gray et al., 2004). A relationship between the ANOM6 and AMO series is supported by correlation analysis (r2 = 0.54; p < 0.001). The wavelet analysis reveals that 2- to 5-year cyclicity is maintained through the period.

All trends are computed as second-order polynomials in Figures 4–9 (Table I). ANOM modes 1, 3, and 5 have particularly large trends relative to the variability, while ANOM modes 4 and 6 have weaker warming trends. In all but the ANOM6 case, trends are relatively linear or linear with steps. Cross correlations between the various PC time scores are presented in Table II. Seasonal modes 1 (marine) and 2 (terrestrial) are positively related at lags − 12, 0, and + 12 and inversely related at − 6 and + 6 lags, with correlations of order 0.58. Seasonal modes 1 and 3 are insignificantly related, while modes 2 and 3 exhibit similar sign changes to 1 and 2, with greater values (0.71 to − 0.83). Terrestrial mode 2 links with and geographically overlaps the Amazon mode 3 in terms of annual cycle. The degree of freedom is one per year (∼50) for the seasonal PC scores, whereas for the anomaly scores it is > 100. Anomaly time scores 1 and 3 are positively and significantly related at all lags (rmax = 0.62), suggesting that the central Atlantic and eastern Pacific pressure fields covary. ANOM1 and 4 are only related at zero lag, whereas ANOM1 and 5 are positively related at 0- to 12-month lag. Hence, there is a delayed connection between central Atlantic SLP and MDR-area SST. ANOM2 is related to ANOM6 at 0- and 6-month lag (rmax = 0.48), so the east Pacific SST signal covaries with North Atlantic SST. ANOM3 and 5 are positively and significantly related at all lags (rmax = 0.58), such that the east Pacific pressure signal interacts with MDR-area SST. ANOM3 and 6 are negatively related at − 6 to + 6 lags, while ANOM4 and 6 reach a significant correlation only at zero lag. ANOM5 and 6 are significantly negatively related at lags − 12 to + 6 months (rmax = − 0.50) suggesting that MDR-area and South Atlantic SST interact in dipole fashion.

Table I. Details of second-order polynomial trends in the ANOM1–6 time scores
 LabelTrend R2Trend equation
  1. Explanation of climate symbols: EqAtl, equatorial Atlantic; Epac, eastern Pacific; NEAtl, NE Atlantic; Satl, S Atlantic; mdrAtl, major development region of tropical cyclones in the North Atlantic.

ANOM1EqAtl-SLP0.300.0005x2 − 0.56x + 118
ANOM2EPac-SST0.21− 4E-05x2 + 0.23x − 72
ANOM3EPac-SLP0.310.0003x2 − 0.45x + 112
ANOM4NEAtl-U0.130.0001x2 − 0.14x + 27
ANOM5SAtl-SST0.36− 6E-05x2 − 0.11x + 48
ANOM6mdrAtl-SST0.160.0006x2 − 0.35x + 28
Table II. Cross-correlations between all combinations of PC scores for seasonal and anomaly data at lags of − 12, − 6, 0, + 6, and + 12 months. Bold values are significant at 99% confidence limit
− 120.570.320.72
− 60.58− 0.290.82
+ 60.58− 0.280.83
+ 120.580.320.71
LagANOM 1–2ANOM 1–3ANOM 1–4ANOM 1–5ANOM 1–6
− 120.110.370.160.22− 0.04
− 6− 0.010.400.170.21− 0.09
00.060.620.320.30− 0.18
+ 6− 0.110.430.060.380.30
+ 12− 0.020.420.050.400.27
−− 0.09
00.150.23− 0.020.48
+ 120.140.11− 0.050.18
LagANOM 3–4ANOM 3–5ANOM 3–6ANOM 4–5ANOM 4–6ANOM 5–6
− 120.210.47−
− 60.170.500.
00.220.580.32− 0.010.440.50
+ 60.170.560.
+ 120.180.56−− 0.11

4. Use of modes to predict Caribbean climate

Hastenrath (1978, 1984), Enfield (1996), and Enfield and Mayer (1997) have shown that rainfall anomalies around the Caribbean are related to both tropical North Atlantic and east Pacific SSTs. Although the two SST indices are positively correlated, their influence on rainfall is strongest when anomalies are of opposite signs. Enfield and Alfaro (1999) show that a cool east Pacific/warm central Atlantic contributes to above normal Caribbean rainfall. Here, we have seen how the regional climate and Pacific-Atlantic interaction is represented through application of PCA to gridded data sets of monthly T, SLP, and U fields. The leading seasonal and inter-annual mode scores were employed in multivariate regression models at two season lead times to test how they ‘fit’ rainfall and hurricane.

Table III lists the results of multivariate linear regression. The models tend to employ 2–3 predictors, often one seasonal mode and two anomaly modes. The results suggest a limited predictability that is better for the southeastern Caribbean islands in MAM season (r2 = 0.31) and western Cuba in ASO season (r2 = 0.29). The Lesser Antilles (southeastern Caribbean) early summer rainfall model uses positive SEAS2 and negative SEAS3 mode contributions, without input from the anomaly modes. The model indicates that Caribbean rainfall is enhanced following winters with warmer temperatures over Mexico, southern USA, and the Sahara region. The SLP enhances early summer Caribbean rainfall when it is above normal over South America but below normal over the Lesser Antilles in the preceding winter. The central Caribbean (Dominican Republic and Puerto Rico) early summer rainfall is predicted using only a positive SEAS2. The western Cuba late summer rainfall model uses negative ANOM2 and positive ANOM3, both representing Pacific ENSO. So with cool east Pacific SST and higher SLP consistent with La Nina conditions, we speculate that easterly wave- and tropical cyclone-driven wet spells tend to strengthen and propagate further west to reach Cuba in late summer. For the other rainfall areas and seasons, the explained variance is limited, as illustrated in Figure 10.

Figure 10.

R2 fit of multivariate regression equations to early and late summer season (MAM and ASO) rainfall based on PCA predictors at two-season lead. Horizontal dashed line is 95% significance level. This figure is available in colour online at

Table III. Multivariate algorithms fitting Caribbean climate indices at two season lead for the March–May and August–October seasons. Refer to Table I for explanation of climate symbols
  1. aR2 is hindcast fit deflated for predictors, with df > 50, R2 > 0.12 is significant at 95% confidence limit.

0.21W Cuba− 0.54 × EqAtl-SLP (ANOM1)− 0.62 × EPac-SST (ANOM2)+ 0.67 × EPac-SLP (ANOM3)
0.12E Cuba+ 0.20 × SEAS1+ 0.26 × SEAS2+ 0.22 × EPac-SLP (ANOM3)
0.17Dominican Republic and Puerto Rico+ 0.43 × SEAS2  
0.31SE Caribbean+ 0.44 × SEAS2− 0.29 × SEAS3 
0.29W Cuba− 0.38 × EPac-SST (ANOM2)+ 0.44 × EPac-SLP (ANOM3) 
0.16E Cuba+ 0.29 × EqAtl-SLP (ANOM1)− 0.38 × SAtl-SST (ANOM5) 
0.07Dominican Republic and Puerto Rico− 0.30 × SEAS2  
0.16SE Caribbean− 0.23 × SEAS3− 0.54 × EPac-SST (ANOM2)+ 0.20 × mdrAtl-SST (ANOM6)
0.17No. of Caribbean hurricane days− 0.54 × EPac-SST (ANOM2)− 0.31 × SAtl-SST (ANOM5)+ 0.53 × mdrAtl-SST (ANOM6)
0.23No. of Caribbean tropical cyclones− 0.65 × EPac-SST (ANOM2)− 0.37 × SAtl-SST (ANOM5)+ 0.59 × mdrAtl-SST (ANOM6)
0.16Dominican Republic station data− 0.41 × EPac-SST (ANOM2)+ 0.61 × mdrAtl-SST (ANOM6) 
0.16Barbados station data+ 0.26 × SEAS1+ 0.40 × EqAtl-SLP (ANOM1)+ 0.17 × mdrAtl-SST (ANOM6)

Summarizing the contributions from all predictors (adding coefficients from the best fit algorithms) we find that SEAS1 is selected once with a cumulative influence of 0.20. SEAS2 is selected four times, its early summer influence is + 1.13, in late summer, its influence is weaker and changes sign (−0.30). SEAS3 is selected twice and it influence is − 0.52 spread over both seasons. The cumulative regression coefficient for the Atlantic SLP (ANOM1) is − 0.54 in early summer, but + 0.29 in late summer, being selected twice. Pacific SST (ANOM2) is selected four times, and its influence is consistently strong and negative (−1.54). Pacific SLP mode (ANOM3) is selected three times and is consistently positive with cumulative influence of + 1.33. ANOM4 is never selected, while ANOM5 and 6, tropical South and North Atlantic SST, are selected once each (−0.38 and + 0.20, respectively). The result is not significantly different when island-scale rainfall data (i.e. Dominican Republic, Barbados) are used, suggesting that the idea of downscaling is plausible—yet little variance is accounted for (16%). About one quarter of tropical cyclone frequency is predicted (r2 = 0.23) using a negative contribution from eastern Pacific SST and South Atlantic SST and positive contribution from tropical North Atlantic SST. Hence, a dipole in not represented, instead the entire tropical Atlantic should be warmer than usual for enhanced tropical cyclones and hurricanes, features consistent with Enfield and Alfaro (1999).

5. Discussion

We have jointly analysed surface temperature, sea-level pressure, and U fields in the period 1949–2006 across the Western Hemisphere to study the joint modes of climatic variability. Data composed of 352 grid points at 5° resolution in the area 20°S–35°N, 140°W–20°E derived from NCEP/NCAR reanalysis were subjected to PCAs for raw and residual data. We found that three seasonal (SEAS) and six anomaly (ANOM) modes reproduce the most essential information regarding climate of the inter-Americas.

The SEAS1 pattern represents the oceanic annual cycle with an N-S dipole of Ts loading evident over the central oceans. SEAS2 is the terrestrial annual cycle represented by an N-S dipole of Ts loading mainly over land areas. SEAS3 is also a terrestrial mode but with a SLP centre of action over the Amazon that affects Central America through the Caribbean.

ANOM1 is the equatorial Atlantic SLP mode, which exceeds the Pacific ENSO mode in our analysis—given a domain that includes most of the Atlantic, but only half the Pacific Ocean. ANOM2 is represented by the Pacific ENSO mode with Niño1–3 SSTs dominant and an anti-phase response in the Atlantic. The Pacific ENSO is again represented in ANOM3 through the SLP loading in the eastern Pacific and a weak anti-phase Atlantic response. Because of the spatial domain (most of the Atlantic, but only half of the Pacific Ocean), greater variance is explained by the equatorial Atlantic SLP mode compared with ENSO. The ANOM4 pattern is an isolated mode in the NE Atlantic dominated by U and SLP. It is never selected to predict Caribbean climate. ANOM5 is the South Atlantic SST mode, with a wind loading 20° to the west, and an anti-phase response in the SE Pacific. ANOM6 reflects the tropical North Atlantic SST mode (MDR), with corresponding SLP and U responses that fall in the path of African easterly waves that develop into Caribbean hurricanes. Its SLP loading lies 20° west of the SST loading.

While the time scores for the SEAS modes exhibit little trend, the time scores for all six ANOM modes display significant trends that suggest a climate drift related to warmer surface temperatures and concomitant lowering of surface air pressure. These trends are often removed in studies on climate variability and predictability, but here they have been retained, together with trends in the target data consisting of Caribbean rainfall and tropical cyclones. In general, there is a drying trend in the Caribbean climate over the period, embedded within which there is variability in space and time. The early summer rainfall is spatially rather heterogeneous, with a mean correlation (r) across four Caribbean sub-regions of + 0.10. On the other hand, the late summer rainfall is more consistent, r = + 0.40. This is likely due to the westward sweep of rain producing weather systems in late summer affecting the Antilles Islands coherently, whereas in early summer, sub-tropical westerly troughs spawn meridional cloud bands that affect certain longitudes. The downward trend of Caribbean Antilles rainfall averaged for all areas and seasons has a R2 fit of 0.35, consistent with the trends of ANOM1, 3, and 5. Rainfall has decreased from a mean of 108 mm/month in 1951 to 76 mm/month in 2005. Jury and Winter (2009) show that subsidence from the Hadley circulation is an underlying factor in the drying trend. The frequency of Caribbean hurricanes corresponds with the second-order polynomial trend in tropical North Atlantic Ts (ANOM6), a single 50-year cycle. The work is new in that it identifies the covariability in thermodynamic and kinematic fields embedded within a global warming trend.

Our use of joint PCA scores to develop predictive algorithms at two season lead time yielded 20% of variance on average. There was little difference between seasons and no specific pattern in the predictability, other than for Pacific ENSO and tropical Atlantic predictors to select with opposing sign. New techniques we explored included (1) the use of seasonal modes to supplement the anomaly contributions and (2) the retention of trends in both predictor and target time-series. The performance of our models informs us that the interaction of ocean and atmosphere, and Atlantic and Pacific, is a complex process that is difficult to meaningfully reduce even through joint PCA techniques. A significant result was the value of seasonal modes in predicting inter-annual rainfall departures at 6-month lead time. Seasonal mode 2 has positive influence on early summer rainfall but negative influence on late summer rainfall. This mode reflects a north-south gradient, such that lower pressure and warmer temperature over North America in contrast with higher pressure and cooler temperature over South America anticipates above normal early summer rainfall in the Caribbean and below normal late summer rainfall. This seasonal signal modulates the well-known Pacific ENSO and Atlantic dipole influence on Caribbean rainfall.

Referring to the scientific questions posed earlier, we uncovered many expected results and a few surprises. With regard to whether the annual cycle is decomposed into terrestrial and marine modes represented as N-S dipoles (scientific question 1), we found this to be the case, with the marine mode lagging by 2 months. Although the marine mode Ts field had a broad N-S dipole structure, a SLP axis was found in sub-tropical zones frequented by hurricanes (20°N). The terrestrial mode exhibited strong Ts gradients along the SW monsoon regions of West Africa and Central America. The third mode of annual cycle variability contained the CLLJ linked with a centre of action in SLP over the northeastern Amazon and equatorial west Atlantic. With reference to the importance of the west Atlantic Warm Pool and CLLJ (scientific question 3), we did not find a strong linkage of the CLLJ to the North Atlantic anticyclone, despite our domain covering its ridge axis. Instead, the trade winds associated with the SLP gradient between the North Atlantic anticyclone and equatorial trough emerged as the leading inter-annual mode. The Pacific ENSO was relegated to ANOM modes 2 and 3, while features in the tropical Atlantic slipped to ANOM modes 5 and 6 (south leading the north in variance). The West Atlantic warm pool appeared as a Ts ‘side lobe’ extending from the eastern Pacific in ANOM3 and also made its presence felt in the SLP field of SEAS1.

Although our PCA work has uncovered a number of useful inter-relationships, further work could be done to include the meridional component of wind and precipitable water, to round out the suite of climate variables to account for seasonal and anomaly variability contributed by LF troughs that bring meridional cloud bands to the Caribbean. Another recommendation for further research is to modify the way the joint analysis is presented, by mapping regression coefficients between the three variables. This would yield amplitudes for each mode that might improve our understanding of how the different influences evolve in time and space.