Trends and Variability of Sea Surface Temperature in the Northwest Atlantic from Three Historical Gridded Datasets

ABSTRACT

Historical variability in sea surface temperature (SST) in the North Atlantic (NA) is examined using trend and Empirical Orthogonal Function (EOF) analyses of annual and summer means from three interpolated monthly datasets: Hadley Centre Sea Ice and Sea Surface Temperature (HadISST1), Extended Reconstruction of SST (ERSST), and Centennial in situ Observation-Based Estimates (COBE). Comparisons with time series of upper-ocean temperature from four monitoring sites off Atlantic Canada reveal substantial similarity in the interannual to multi-decadal variability but notable differences in the longer-term trends. The magnitude of decadal-scale variability is comparable to, or greater than, the long-term changes in all of the datasets; together with the trend discrepancies, this needs to be considered in climate change applications. Averaged over the NA, the annual means have a long-term increasing trend and a pronounced multi-decadal variation, resembling those in global mean (land-ocean) surface temperature and the Atlantic Multi-decadal Oscillation (AMO). There is remarkable similarity in the spatial and temporal variability of the three leading EOF modes from the different gridded datasets, with the first highly correlated with the AMO, the second modestly correlated with the winter North Atlantic Oscillation, and the third apparently related to ocean circulation variability. Trends since 1981 are generally two to three times larger than those since 1900 and 1950, which is at least partly related to the phase of the AMO. Trends in the summer means are generally larger than in the annual means. Overall, the results provide support for both anthropogenic global warming and decadal-scale natural variations making important contributions to ocean climate variability in the Northwest Atlantic.

RÉSUMÉ

[Traduit par la rédaction] Nous examinons la variabilité passée des températures à la surface de la mer (SST) de l'Atlantique Nord, et ce, en analysant les tendances et les fonctions orthogonales empiriques (FOE) des moyennes annuelles et estivales de trois séries de données mensuelles interpolées : les températures à la surface de la glace marine et de la mer issues du Hadley Centre (HadISST1), la reconstitution élargie des SST (ERSST) et les estimations provenant d'observations in situ centennales (COBE). Les comparaisons avec des séries temporelles de la température de la couche supérieure de l'océan, provenant de quatre sites de mesure au large de la côte atlantique canadienne, révèlent des similitudes substantielles dans les variabilités interannuelles à multidécennales, mais des différences notables dans les tendances à plus long terme. La magnitude de la variabilité à l’échelle décennale est comparable ou supérieure aux changements à long terme dans toutes les séries de données. Ce résultat et les différences dans les tendances devront être pris en considération dans le cadre d’études sur les changements climatiques. En moyenne, sur l'Atlantique Nord, les moyennes annuelles montrent une tendance à long terme à la hausse et une variation multidécennale prononcée, semblables à celles des moyennes mondiales de températures de surface (terre-océan) et de l'oscillation atlantique multidécennale (OAM). Il existe une similitude remarquable dans les variabilités spatiales et temporelles des trois modes principaux des FOE, calculés pour les diverses séries de données réparties sur une grille. Le premier mode étant hautement corrélé à l'OAM; le second, modérément corrélé avec l'oscillation nord-atlantique (ONA) hivernale; et le troisième, apparemment lié à la variabilité de la circulation océanique. Les tendances depuis 1981 sont généralement de deux à trois fois supérieures à celles calculées depuis 1900 et 1950, ce qui est, partiellement du moins, lié à la phase de l'OAM. Les tendances des moyennes estivales s'avèrent généralement plus grandes que celles des moyennes annuelles. En somme, les résultats montrent que le réchauffement anthropique de la planète et les variations naturelles à l’échelle décennale contribuent tous deux considérablement à la variabilité du climat océanique dans l'Atlantique du Nord-Ouest.

Keywords:

• sea surface temperature
• ocean climate trends
• decadal-scale variability
• North Atlantic
• Atlantic Multi-decadal Oscillation
• North Atlantic Oscillation

1 Introduction

Sea surface temperature (SST) is one of the most important variables for ocean climate change and variability studies. With the longest observational records of any ocean variable (e.g., Yasunaka & Hanawa, 2011), it reflects the important coupling between the atmosphere and ocean in the climate system (e.g., Deser, Alexander, Xie, & Phillips, 2010; Yu & Boer, 2006) and is an indicator of the upper-ocean temperature that affects ocean biogeochemistry and marine ecosystems, marine activities, and coastal human populations.

Observations of SST are available from dedicated and opportunistic observations on vessels, ocean monitoring programs over the past 50–100 years, and satellite observations during the past three decades. Because of diurnal variability and the large vertical temperature gradient near the ocean surface, there are substantial challenges in the interpretation of SST variability in relation to possible aliasing or biases from sampling with different methodologies (e.g., Rayner et al., 2010). In this paper, we describe results from an analysis for spatial and sub-annual (i.e., at interannual and longer time scales) variability in SST in the North Atlantic (NA) in three global interpolated gridded monthly SST datasets used in the Fifth Assessment Report (AR5; Hartmann et al., 2013) of the Intergovernmental Panel on Climate Change (IPCC). The datasets are the Hadley Centre Sea Ice and Sea Surface Temperature (HadISST1; Rayner et al., 2003), the Extended Reconstruction of SST (ERSST; Smith, Reynolds, Peterson, & Lawrimore, 2008), and the Centennial in situ Observation-Based Estimates (COBE; Ishii, Shouji, Sugimoto, & Matsumoto, 2005). Most of the results are discussed in more detail in a precursor report (Loder, Wang, van der Baaren, & Pettipas, 2013).

Our focus is the Northwest Atlantic (NWA), in order to provide input to Fisheries and Oceans Canada's (DFO's) Aquatic Climate Change and Adaptation Services Program (ACCASP). Our goal is to describe past SST variability in the NWA off Atlantic Canada using these interpolated gridded datasets in conjunction with available time series of in situ upper-ocean temperature observations from DFO monitoring programs. In particular, our systematic examination of the three interpolated datasets was motivated by our exploratory investigation of readily available time series from two of them, which indicated marked differences between the Labrador Sea (LS) and Scotian Shelf (SS) in the SST trends and variability in HadISST1 and notable differences between HadISST1 and ERSST in the trends on the SS (Loder et al., 2013).

The NWA is one of the most dynamic parts of the world's oceans, with the competing influences of the Labrador Current and Gulf Stream, Arctic Ocean outflows, and atmospheric flows from the Arctic and North American continent, and variability in the Atlantic Meridional Overturning Circulation (AMOC) that has one of its primary densification (sinking) regions in the LS. Analyses of historical hydrographic profiles have revealed cool fresh ocean conditions in the SS region during the 1960s (e.g., Petrie & Drinkwater, 1993) and cool fresh conditions in the LS in the 1980s and early 1990s (e.g., Yashayaev, 2007). In both cases, the anomalous conditions were attributed to persistent winter anomalies of the North Atlantic Oscillation (NAO). In the LS case, a direct influence of positive NAO anomalies on ocean temperature and salinity was suggested, and in the SS case the suggestion was an indirect influence from persistent negative NAO anomalies and resulting changes in the equatorward extent of the NA's subpolar gyre (and Labrador Current). However, it is not clear that the origin of these and other ocean anomalies are solely related to variability in the NAO (e.g., also see Deser et al., 2010, for direct NAO influences on SST in the broader NWA).

On larger scales, there is strong multi-decadal variability in SST in the NA (e.g., Deser et al., 2010; Polyakov, Alexeev, Bhatt, Polyakova, & Zhang, 2010; Terray, 2012). In particular, the Atlantic Multi-decadal Oscillation (AMO), also referred to as the Atlantic Multi-decadal Variability (AMV; Hakkinen, Rhines, & Worthen, 2011; Ou, 2012), has been identified in spatially averaged SST (over the NA) with a predominant variation with a period of approximately 60–70 years (Knight, Allan, Folland, Vellinga, & Mann, 2005; Wang, Dong, Evan, Foltz, & Lee, 2012). There are indications of overall cooling from the late 1950s to the early 1970s and overall warming from the 1980s to the early 2000s. The underlying spatial patterns in both cases point to the largest magnitudes occurring in the subpolar NA (Drinkwater et al., 2014). Analyses of proxy climate variables over the past 8000 years (Knudsen, Seidenkrantz, Jacobsen, & Kuijpers, 2011) and of δ¹⁸O variability in Greenland ice cores, together with atmosphere–ocean model simulations over the past millennium (Chylek et al., 2012), suggest intermittent variability in air temperatures similar to the recent AMV in SST. This indicates that the AMV is at least partly related to natural internal variability of the coupled climate system. On the other hand, there have been numerous discussions of a relation between the AMV and decadal-scale variability in global mean (surface) air temperature (e.g., Booth, Dunstone, Halloran, Andrews, & Bellouin, 2012; Parker et al., 2007; Ting, Kushnir, & Li, 2014; Zhang et al., 2013) and of the need to account for a larger-scale contribution in discussions of the AMO (e.g., Mann, Steinman, & Miller, 2014; Trenberth and Shea, 2006). There is also an increasing recognition that decadal-scale ocean temperature variability is an important factor in “hiatus” decades during which global surface air temperature has not followed the longer-term global warming trend (e.g., Chen & Tung, 2014; Drijfhout et al., 2014; Maher, Sen Gupta, & England, 2014).

There are suggestions in the literature that the AMO may be related to variability in the AMOC (e.g., Drijfhout, van Oldenborgh, & Cimatoribus, 2012; Medhaug and Furevik, 2011; Park and Latif, 2010; Rahmstorf et al., 2015), multi-decadal variability in the NAO (e.g., Robson, Sutton, Lohmann, Smith, & Palmer, 2012), or atmospheric blocking over the northern NA (e.g., Hakkinen et al., 2011). Thus, it may also be appropriate to think of the AMO, NAO, and AMOC variability as different manifestations (with somewhat differing space and time scales) of coupled atmosphere-ice-ocean variability in and over the NA and adjoining regions (e.g., Arctic and North America via the NAO and South Atlantic via AMOC variability) (Grossmann & Klotzbach, 2009; Miles et al., 2014; Walter & Graf, 2002). All of these possibilities make it important to understand SST variability in the NWA in relation to larger-scale climate variability, particularly for attribution studies and climate change projections. In particular, the extent to which regional ocean temperature variability (e.g., in the LS and on the SS as mentioned above) is influenced by, or linked to, the AMO and decadal-scale variability in global air temperature is unclear.

Regionally, Friedland and Hare (2007) using ERSST data and Shearman and Lentz (2010) using ship- and shore-based observations found overall warming and multi-decadal variability in SST along the east coast of the United States since the late 1800s. Han, Ma, and Bao (2013) and Hebert (2013) found strong spatial variability in the trends in the off-shelf and shelf parts, respectively, of the NWA since approximately 1950, and the latter found substantially enhanced trends since 1979. Belkin (2009), using the Hadley SST climatology, found changes of 1.0°, 0.9°, and 0.2°C from 1982 to 2006 on the Newfoundland and Labrador Shelves, SS, and northeastern US Shelf, respectively, and suggested that the recent “rapid” warming in the subpolar NA was likely related to NAO variability. Mills et al. (2013) have pointed to accelerated warming in the Gulf of Maine (GoM) since 1982, but it seems likely that this is at least partly related to decadal-scale variability rather than just a reflection of global warming.

In Section 2 of the paper we describe the data and methods used in the analyses. In Section 3 we present a brief comparison between the gridded datasets and DFO observational time series at four monitoring sites, and in Section 4 we present an initial exploration of long-term (since 1870) variability in two of the gridded datasets. We present the primary results in Sections 5 and 6, with the former focused on trends over various periods in the gridded datasets and the latter on the spatial and temporal structure of the leading modes of variability. We conclude the paper with a summary and discussion in Section 7.

2 Data and methods

a Interpolated Gridded SST Datasets and Analyses

The three interpolated gridded datasets of historical monthly SST considered here are

• HadISST1 (version 1), which is the UK Met Office's Hadley Centre Interpolated SST dataset, with 1° × 1° resolution extending back to 1870 (Rayner et al., 2003);

• ERSST (version 3b), which is the US National Oceanic and Atmospheric Administration's (NOAA) Extended Reconstruction of SST dataset, with 2° × 2° resolution extending back to 1854 (Smith et al., 2008); and

• COBE SST, which is the Japanese Meteorological Agency's Centennial in situ Observation-Based Estimates of variability of SST, with 1° × 1° resolution extending back to 1891 (Ishii et al., 2005).

Data quality, especially during the early part of the records and in remote and/or ice-covered regions, is an issue with these datasets (e.g., Hartmann et al., 2013; Yasunaka and Hanawa, 2011) such that the results in some periods and areas need to be interpreted with caution. These datasets were derived from differing combinations of the International Comprehensive Ocean-Atmosphere Data Set (ICOADS) and other observations, using different methodologies for quality control and interpolation. HadISST1 used reduced-space optimal interpolation; ERSST used Empirical Orthogonal Functions (EOFs), and COBE used an optimal interpolation scheme.

Two derived datasets for interannual and longer-term variability studies of SST in the NA have been computed from each of the source datasets (on their native grids): (i) annual means computed from the monthly means for all months, and (ii) summer means computed from the monthly means for July, August, and September. For this purpose, the NA is defined as the region in the latitude and longitude ranges in Table 1 (with the Mediterranean excluded). Because ERSST and COBE have estimated monthly SST values (often −1.8°C) in ice-covered regions, such as Baffin Bay and the Greenland and Labrador shelves, their values were used for the entire domain. In contrast, HadISST1 has missing or dubious data in some ice-covered areas, and in those cases we omitted those grid squares from our annual mean analyses. We analyzed the summer means to provide increased spatial coverage in these areas and to obtain an indication of the seasonal dependence and the variability in a particular season.

Table 1. Domain and sub-regions for which trends of spatially averaged gridded SSTs are presented (Table 4), approximated by latitude–longitude boxes with the indicated ranges (also see Fig. 1a). The Mid-latitude WNA TZ refers to the Mid-latitude Western North Atlantic Transition Zone between the subpolar and subtropical gyres (see Brock et al. (2012) for discussion), which is subdivided here into shelf and two off-shelf sub-regions.

Domain or Sub-region	Abbreviation	Latitude (°N)	Longitude (°W)
North Atlantic	NA	6.5–74.5	−29.5–100.5
Northwest Atlantic	NWA	30.5–74.5	20.5–100.5
Labrador Shelf/Sea	LSS	52–58	48–54
Newfoundland Shelf/Slope	NSS	44–52	46–56
Mid-latitude WNA TZ: Shelf	MTZ-S	42–50	56–70
Mid-latitude WNA TZ: Off- Shelf West	MTZ-OSW	38–42	50–70
Mid-latitude WNA TZ: Off- Shelf East	MTZ-OSE	38–44	40–48

Trend analyses for the NA domain have been carried out on these two derived datasets for four periods: 1870–2011, as the longest common period of at least two of the datasets; 1900–2011, in order to provide a century-scale view of all three datasets; 1950–2011, in order to examine variability in the post-war era with more comprehensive sampling; and 1979–2011, in order to examine recent variability during the high-resolution satellite era. In addition, to provide an indication of the magnitude of the trends in various areas of particular interest, trends were also computed for spatially averaged data from the six sub-regions indicated in Fig. 1a and Table 1. Standard linear regression techniques were used in the trend analyses, and an indication of the trends’ statistical significance at the 95% confidence level is included for the various regions, with the number of degrees of freedom estimated from the number of data points. Because there appears to be some autocorrelation in the time series, the statistical significance (in relation to the higher-frequency variability in each time series) may be overestimated. Further, because there are substantial differences among the trends from the three datasets, it appears that there are additional sampling and/or methodological differences among the datasets such that it is not possible to estimate accurate and reliable trends without more in-depth analyses of the gridded datasets.

Fig. 1 (a) Map showing the NWA sub-region/domain and the five other sub-regions (Table 1) for which trends are presented in Table 4. (b) Map showing locations of the four DFO monitoring sites (P5 denotes Prince 5; EB, Emerald Basin; and S27, Station 27, as well as Bravo), the positions of the gridded data used in the comparison, and ocean areas discussed in the text (LS denotes Labrador Sea; FC, Flemish Cap; GB, Grand Bank; SS, Scotian Shelf; and GoM, Gulf of Maine).

EOF analyses were carried out on the different NA datasets for the four indicated periods, with or without prior detrending depending on the question being addressed. The EOF values for each analysis were normalized by dividing the values for all grid points by the maximum absolute value of that EOF for any point in the domain. For consistency, the corresponding principal component (PC) time series for each analysis were normalized by multiplying each of their values by the same maximum EOF value that was used to normalize the corresponding EOF pattern.

Most of the EOF analyses were also carried out for the NWA sub-region (Fig. 1a, Table 1) in order to examine the sensitivity of the results to the size of the domain considered. The results for the three leading EOF modes were similar for the different domains (Loder et al., 2013) so the NWA results will only be mentioned briefly here. The leading EOFs accounted for a slightly larger fraction of the variance on the NWA domain, confirming their importance to the area of primary interest.

b DFO Monitoring Datasets

Time series of upper-ocean temperature variability from in situ observations at four DFO monitoring sites (Fig. 1b) were used to examine the consistency of SST variability in the interpolated datasets with that reported by DFO (e.g., Colbourne et al., 2014; Hebert, Pettipas, Brickman, & Dever, 2014; Yashayaev et al., 2014). Monthly means of near-surface (approximately 0–5 m) temperature for individual years were available for most months since a particular “record-start” year at three DFO Atlantic Zone Monitoring Program (AZMP; http://www.bio.gc.ca/science/monitoring-monitorage/azmp-pmza-en.php) sites, specifically

• at Prince 5 (P5) in Passamaquoddy Bay in the outer Bay of Fundy since 1924;

• in Emerald Basin (EB) on the SS off Halifax since 1947; and

• at Station 27 (S27) on the Grand Bank (GB) off St. John's since 1946.

Annual means and summer (July–September) means for individual years were computed by averaging the available monthly anomalies (with respect to the long-term monthly mean) in each year or summer, then adding this anomaly to the long-term annual or summer mean.

The time series for the fourth DFO site (I. Yashayaev, Bedford Institute of Oceanography, personal communication, 2013) for the 10–30 m depth interval in the central LS from DFO's Atlantic Zone Off-shelf Monitoring Program (AZOMP; http://www.bio.gc.ca/science/monitoring-monitorage/azomp-pmzao/azomp-pmzao-en.php) were similarly constructed from anomalies about the annual cycle from observations collected in the vicinity of the former Ocean Weather Ship (OWS) Bravo by the OWS, DFO, and others. These time series provided annual (summer) means for most years since 1948 and for four (three) earlier years as far back as 1928, under the requirement of at least three observations in each year (summer). The subsurface interval was purposely chosen to avoid spurious variations related to seasonal changes and different depths below the surface with the sparse sampling.

In comparisons with the contributions from the leading gridded-data EOF modes to SST variability at these sites, a 3-point running mean filter (retaining means for windows with values from at least two years) was used to provide weak filtering of the pronounced interannual variability in the DFO time series.

c Potential Forcing Indices

Although we focus on the spatial and temporal structure of the SST variability in this paper, brief consideration will be given to the possible origins of the variability based on comparisons and correlations with other climate indices. Three long-term and larger-scale annual indices that will be considered are shown in Fig. 2a: global mean (land–sea) surface temperature (GMST), taken from the website of the Goddard Institute for Space Studies (GISS), National Aeronautics and Space Administration (NASA) (http://data.giss.nasa.gov/gistemp/graphs_v3); and the AMO index and detrended winter (January–March) NAO index taken from the Global Climate Observing System (GCOS) website of NOAA (http://www.esrl.noaa.gov/psd/gcos_wgsp/Timeseries). Decadal-scale variability is apparent in all three indices, with additional higher-frequency variability in the NAO.

Fig. 2 Time series of forcing indices considered in correlation analyses to explore the origin of the leading modes of SST variability: (a) AMO index and GMST anomalies in the upper panel and winter NAO index in the lower panel; and (b) sea surface height (SSH) PCs in upper panel and wind stress curl (WSC) PC1 and Bermuda-minus-Bravo SSH difference in lower panel, from the Wang et al. (2015) hindcast simulation of the NA during 1958–2004.

In addition, consideration is given to sea surface height (SSH) and wind stress curl (WSC) indices (Fig. 2b) from a recent hindcast model study of ocean variability in the NA during the 1958–2004 period (Wang et al., 2015), which also show decadal-scale variability.

3 Comparison of datasets at DFO monitoring sites

In this section, we compare the time series of upper-ocean temperature from the DFO monitoring sites with time series from the gridded datasets interpolated to the site (for OWS Bravo) or for a nearby position (for the others). The annual mean time series for the entire length of the DFO records and from the gridded data are shown in Fig. 3a and the summer mean time series in Fig. 3b. Some similarities are expected because some of the observations contributing to the DFO time series also contribute to the gridded datasets. Some differences are expected from factors such as differing source data, possible differing criteria for “surface” in the datasets, limited winter data at OWS Bravo, and the location of P5 in a coastal embayment.

Fig. 3 Time series of (a) annual and (b) summer means of upper-ocean temperature at the four DFO monitoring sites (black), and of SST in the vicinity of the sites from the three gridded historical datasets (coloured lines).

Trends in the various datasets are shown in Table 2, and correlations between the DFO and gridded time series after detrending are shown in Table 3. There are substantial similarities among the time series from the gridded datasets and between them and the DFO time series, at all sites. Correlation coefficients between the detrended DFO and gridded time series are between 0.66 and 0.83 for annual and between 0.45 and 0.83 for summer, all of which are highly statistically significant. Averaged across the four sites, the correlations for annual (summer) means are 0.76 (0.70), 0.73 (0.60), and 0.74 (0.69) for HadISST1, ERSST, and COBE, respectively, indicating little difference among the gridded datasets with regard to their agreement with the DFO time series but with slightly poorer agreement for the summer time series than for the annual time series.

Table 2. Trends in (a) annual mean and (b) summer mean upper-ocean temperature (°C per decade) over the entire period of record at the four DFO monitoring sites, together with the trends in the corresponding gridded SST datasets at nearby positions (i.e., of the time series shown in Fig. 3) and the average of the trends for the gridded datasets. P5 denotes Prince 5; EB is Emerald Basin; S27 is Station 27; N is the number of years with data. Trends significant at the 95% confidence level (without consideration of autocorrelation in the time series) are in bold.

Site	Period	N	DFO	HadISST1	ERSST	COBE	Average
(a) Annual
P5	1924–2011	87	0.10	0.16	0.02	0.05	0.08
EB	1947–2011	65	0.08	0.18	0.05	0.09	0.11
S27	1946–2011	66	0.10	0.07	0.08	0.17	0.08
Bravo	1928–2011	62	−0.06	0.03	0.07	0.14	0.08
(b) Summer
P5	1924–2011	86	0.05	0.11	0.05	−0.02	0.04
EB	1947–2011	65	0.23	0.26	0.11	0.06	0.14
S27	1946–2011	66	0.19	0.12	0.15	0.15	0.14
Bravo	1928–2011	61	0.03	0.09	0.05	0.07	0.07

Table 3. Correlation coefficients between the DFO and various gridded time series at the DFO monitoring sites (Fig. 3) after detrending, for annual means and summer means. Correlations significant at the 95% confidence level (without consideration of autocorrelation in the time series) are in bold.

Site	Annual Means			Summer Means
	HadISST1	ERSST	COBE	HadISST1	ERSST	COBE
P5	0.75	0.66	0.78	0.60	0.45	0.60
EB	0.71	0.71	0.72	0.63	0.50	0.61
S27	0.81	0.83	0.79	0.82	0.79	0.83
Bravo	0.78	0.71	0.68	0.76	0.66	0.72

There are also similarities between the annual and summer time series for each site (Fig. 3), but also some notable differences (e.g., EB) pointing to interseasonal differences in the variability. For P5, there is similar interannual to multi-decadal variability among the time series for both the annual and summer means, even though the temperatures in the DFO time series are lower than those in the gridded datasets by approximately 2°C and 4°C for the annual and summer means, respectively. Such a difference is expected because of P5's coastal location and spatially variable tidal mixing in the region. There is similar multi-year to multi-decadal variability among the gridded and DFO time series from each of EB and S27 for each of the annual and summer time series, even though the means and interannual variations differ in some cases. There is also general similarity among the time series for OWS Bravo, particularly on decadal and multi-decadal time scales, even though the subsurface summer DFO time series are low (as expected) and do not capture some of the multi-year features common to two or more of the gridded datasets (or, vice-versa, the gridded datasets do not show some of the interannual variability in the DFO time series).

The HadISST1 annual means are noticeably higher than the ERSST and COBE ones after about 1970 at P5 and EB, over the entire record (from 1948) at S27, and before 1975 at OWS Bravo (Fig. 3a). Similarly, the ERSST annual means are noticeably lower than the other gridded datasets after 1985 at S27 and over the entire record at OWS Bravo. Similar differences are apparent in the summer means at EB, S27, and OWS Bravo, but not at P5 (Fig. 3b). These variations raise the possibility of a systematic difference in the sampling or interpolation, or in the temperature structure (e.g., spatial gradients), during some periods and at some sites.

The DFO annual means have statistically significant trends of +0.1°C per decade at P5 and S27 (Table 2a), a near-significant (p = 0.11) trend of +0.08°C per decade at EB, and a negative but non-significant trend at OWS Bravo (note that these are over different periods, so their differences reflect a mixture of spatial and temporal variability). The P5, EB, and S27 trends are comparable to the trends of approximately 0.1°C per decade (since 1900 or 1950) in surface air temperature for the region, while the reduced OWS Bravo trend is similar to the pattern of reduced warming over the coastal and offshore subpolar NWA (e.g., Galbraith & Larouche, 2013; Hartmann et al. 2013; Peterson and Pettipas, 2013; Vincent et al., 2012). The DFO summer trends point to different seasonality at three of the sites, with larger summer (than annual) trends (approximately 0.2°C per decade) at EB and S27, and a smaller (non-significant) summer trend at P5 but again no significant trend at OWS Bravo.

A striking result is the substantial difference in the trends from the three gridded datasets at particular sites, with a spread of about 0.1°C per decade at all sites for the annual means and of up to 0.2°C per decade (at EB) for the summer means (Table 2). Because this spread is comparable to the magnitude of the trends in the DFO data (and also to the average trends in the gridded data for particular sites), there is a clear indication that caution is needed in the use of site-specific trends from the gridded datasets. For P5 and EB, the HadISST1 trends for both annual and summer means are larger than the DFO ones, and the ERSST and COBE trends tend to be smaller or comparable. For S27, the annual trends from the gridded dataset bracket the DFO one, whereas the summer trends are smaller than the DFO one. For OWS Bravo, the trends from the gridded dataset are all higher than the DFO ones for both annual and summer means, which may be because the actual SST trend is higher than it is at 10–30 m depth (as expected for surface-intensified warming). Overall, the differences between the DFO and gridded trends are of the same sign in the annual and summer means (with the exception of COBE at S27), but the sign of the differences varies with location and gridded dataset.

4 Long-term variability from HadISST1 and ERSST

Before examining further the spatial and temporal patterns of the trends and variability of the three gridded datasets, it is instructive to examine the patterns of variability in the two longest datasets over their common period (since 1870). Figure 4 shows the EOF patterns for the three leading (based on variance explained) modes of HadISST1 and ERSST for 1870 to 2011, and Fig. 5 the corresponding PCs, from analyses without prior detrending of the datasets. For each of the annual (Figs 4a and 5a) and summer (Figs 4b and 5b) means, there is substantial similarity between particular EOF patterns and PCs from the two datasets, especially for modes 2 and 3 (after switching, as explained below). For mode 1 which accounts for 35–40% of the variance in the annual means, the PCs from the different datasets are in close agreement, but there are differences in the detailed structure of the EOF patterns. A notable similarity between the PC1s and the AMO and GMST indices in Fig. 2a is apparent, supported by correlation coefficients (at zero lag) of approximately 0.95 between the PC1s and the AMO, and approximately 0.8 between the PC1s and GMST (both of which are highly statistically significant). Additionally, the spatially uniform sign (phase) of the EOF1 amplitudes provides support for this mode being largely a combination of broad-scale NA (e.g., AMO) and global (e.g., anthropogenic) temperature variability. The summer mode 1 patterns resemble the annual mode 1 patterns in both spatial and temporal variability, while the summer mode 2 and 3 patterns resemble the annual mode 3 and 2 patterns, respectively, also in both spatial and temporal variability. The latter apparent switching of modes simply reflects one mode accounting for a larger percentage of the variance in the annual means and the other a larger percentage in the summer means. These two modes will be discussed further in Section 6a.

Fig. 4 EOF patterns (°C normalized by the peak value) for the leading three modes for (a) the annual and (b) the summer means of HadISST1 and ERSST for 1870 to 2011 without detrending, with the percentage of explained variance indicated in the inset panels.

Fig. 5 PCs for the leading three modes for (a) the annual and (b) the summer means of HadISST1 and ERSST for 1870–2011 without detrending.

To further examine the relation of the mode 1 patterns to the GMST, Fig. 6 shows the anomalies of GMST (about its long-term mean) together with the anomalies of spatially averaged SST in the NA associated with these modes from the EOF analyses described above (all without detrending). There is remarkable similarity between the GMST anomalies and the NA SST anomalies, both in the existence of a long-term trend with a net change in the range of approximately 0.5°–1°C depending on the specific period considered and a multi-decadal variation with a period of approximately 60–70 years and amplitude of a few tenths of a degree Celsius. With the trend in GMST being larger than that in the NA SST, and the multi-decadal variation's amplitude being larger in the NA SST (and even larger in parts of the NA), there is clear support for the common paradigm (e.g., Ting, Kushnir, Seager, & Li, 2009, Ting et al., 2014) of a global warming trend of atmospheric origin and an amplified multi-decadal variation in the NA SST (e.g., the AMO). But there are also indications that multi-decadal variability in GMST (e.g., hiatus periods associated with aerosols and volcanos; Booth et al., 2012) may partly contribute to multi-decadal SST variability in the NA and that the latter may also contribute to the former (e.g., decadal-scale ocean variability contributing to hiatus periods in GMST; Drijfhout et al., 2014). This also points to the need to take into account multi-decadal variability in GMST in descriptions and interpretations of the AMV (e.g., Mann et al., 2014; Trenberth & Shea, 2006). It makes fundamental sense that large-scale ocean and atmospheric variability are interlinked and that, from the perspective of SST variability in the NWA, both larger-scale SST variability in the NA and air temperature (as well as sea ice) variability in the northern hemisphere need to be considered.

Fig. 6 Annual anomalies of NA SST associated with the PC1s from the EOF analyses on HadISST1 and ERSST from 1870 to 2011 and of GMST from GISS. None of the time series were detrended. The NA SST time series were computed by multiplying the PC1s in Fig. 5a by the spatially averaged amplitude over the NA of the corresponding EOF1 in Fig. 4a.

5 Trend patterns in four historical periods

In this section we return to the topic of the trends in the three gridded SST datasets. It is obvious from the large multi-decadal variability discussed in the previous two sections that careful attention needs to be given to the period of a record in relation to this variability, in order to obtain trends that are meaningful for climate change purposes.

The spatial patterns and magnitudes of the trends of the annual and summer means from the gridded datasets during the four different periods are shown in Fig. 7 for the NA, and the trends of the spatially averaged means in the sub-regions of particular interest (Table 1) are shown in Table 4. The large-scale patterns of the trends from the different annual mean datasets are generally similar for particular periods, but there are substantial differences, particularly between the COBE dataset and other datasets around Greenland in the 1900–2011 and 1950–2011 periods. As discussed in the next section, the latter can be traced to a suspect abrupt jump in the COBE SSTs in that region around 1979. As a result, the trends around Greenland in both the annual and summer means from the COBE dataset during these two periods are not considered reliable.

Fig. 7 Trends (°C per decade with the zero contour in white) in the three (a) annual mean and (b) summer mean datasets for the four periods examined. Note that a different temperature trend scale (colour bar) is used for the 1979–2011 periods than for the other three periods.

Table 4. Trends in (a) annual mean and (b) summer mean SST (°C per decade) averaged over the latitude/longitude boxes approximating the regions of interest (Table 1), from the three gridded SST datasets (HadISST1 (HD), ERSST (ER), and COBE (CB)) and for the three different common periods. Trends significant at the 95% confidence level are in bold. The average (Avg) of the trends over the three datasets is shown after those for the individual datasets. The asterisked (*), italicized values indicate trends affected by the suspect warming around Greenland in the COBE dataset (prior to 1979) and not used in the averages.

Region	1900–2011				1950–2011				1979–2011
	HD	ER	CB	Avg	HD	ER	CB	Avg	HD	ER	CB	Avg
(a) Annual Means
NA	0.03	0.05	0.07*	0.04	0.01	0.07	0.10*	0.04	0.26	0.24	0.21	0.24
NWA	−0.01	0.04	0.07*	0.02	−0.02	0.06	0.11*	0.02	0.33	0.29	0.23	0.28
LSS	0.02	0.06	0.07	0.05	0.12	0.09	0.20	0.11	0.45	0.50	0.38	0.44
NSS	0.01	0.05	0.06	0.04	0.13	0.06	0.11	0.10	0.33	0.32	0.26	0.30
MTZ-S	0.11	0.05	0.09	0.09	0.14	0.07	0.15	0.12	0.25	0.21	0.17	0.21
MTZ-OSW	0.09	0.03	0.05	0.06	0.08	0.04	0.00	0.04	0.24	0.20	0.18	0.20
MTZ-OSE	0.07	0.01	0.04	0.04	0.12	0.04	0.04	0.07	0.29	0.25	0.22	0.25
(b) Summer Means
NA	0.05	0.07	0.09*	0.06	0.09	0.10	0.13*	0.10	0.27	0.30	0.26	0.27
NWA	0.08	0.07	0.10*	0.08	0.13	0.10	0.11*	0.12	0.38	0.39	0.34	0.37
LSS	0.06	0.05	0.13	0.05	0.18	0.10	0.09	0.13	0.56	0.54	0.56	0.55
NSS	0.11	0.09	0.12	0.11	0.16	0.13	0.14	0.14	0.38	0.45	0.46	0.43
MTZ-S	0.14	0.08	0.12	0.11	0.22	0.14	0.16	0.17	0.41	0.32	0.34	0.36
MTZ-OSW	0.09	0.06	0.07	0.07	0.09	0.08	0.05	0.07	0.32	0.31	0.31	0.32
MTZ-OSE	0.08	0.05	0.07	0.06	0.11	0.05	0.06	0.07	0.33	0.30	0.28	0.30

Overall, there is broad-scale warming during all four periods and in all three datasets, as well as in each of the annual and summer means, but not in all areas (Fig. 7). In particular, there is a large area of net annual mean cooling in the northern NA south of Greenland in the HadISST1 and ERSST datasets during the two longer periods, and smaller areas of cooling in this region during the 1950–2011 period. There is an indication of these cooling areas in the COBE data for these periods but the above-noted suspect warming dominates near Greenland. This region of cooling is well known from the long-term change maps of global (surface) air temperature presented in the last two IPCC assessment reports (Hartmann et al., 2013; Trenberth et al., 2007), and it is usually attributed to regional atmospheric (e.g., NAO) or oceanic (e.g., AMOC) circulation variability. The cooling in this region is replaced by substantial warming during the 1979–2011 period, and the extent of the cooling area is reduced in summer in all three of the earlier periods (Fig. 7).

The extent and magnitude of the warming can be seen to increase progressively in the two most recent periods (Fig. 7), consistent with some combination of anthropogenic warming and the multi-decadal variability in the AMO and GMST indices (Fig. 2a). There is also more consistency in the patterns across the datasets since 1979, as expected with the improved data coverage during the satellite era. Nevertheless, strong spatial variability in the trends is present during each period, and it is noteworthy that the sign of their latitudinal gradient switches from negative (cooling in the north) in the three earlier periods to positive (greater warming in the north) in the most recent period.

The magnitude and pattern of the trends for both the annual and summer means during the 1870–2011 period are similar in the two longer datasets. However, it is noteworthy that, in the mid-latitude transition zone (MTZ) between the subpolar and subtropical gyres off Atlantic Canada (Brock, Kenchington, & Martinez-Arroyo, 2012; Loder, Petrie, & Gawarkiewicz, 1998), the warming is greater in HadISST1 than in ERSST in both the annual and summer means. The origin of this discrepancy is unclear and beyond the scope of this paper, so the variability in this region prior to 1900 will not be discussed further.

During the 1900–2011 and 1950–2011 periods, the average annual mean trends are 0.04°C per decade in the NA and 0.02°C per decade in the NWA (Table 4; based on the HadISST1 and ERSST datasets because the COBE values are probably degraded by the suspect data off Greenland); the trends from ERSST are larger (and statistically significant), whereas those from HadISST1 are smaller and slightly negative for the NWA (although not statistically significant). For the summer means, the trends for both periods and datasets are more positive (all statistically significant) and closer in magnitude, with overall averages for the NA (NWA) of 0.06° (0.08°) and 0.10° (0.12°)C per decade for the 1900–2011 and 1950–2011 periods, respectively. These larger values in summer are consistent with the reduced extent of the cooling area south of Greenland and the expected increase in surface warming resulting from the shallower mixed layers in summer.

For the shelf-slope sub-regions off Atlantic Canada, the annual mean trends increase from the 1900–2011 period to the 1950–2011 period: for the Labrador Shelf/Sea (LSS), the average (across datasets) increases from 0.05° to 0.11°C per decade; for the Newfoundland Shelf/Slope (NSS), from 0.04° to 0.10°C per decade; and for the shelf portion of the MTZ (MTZ-S), from 0.09° to 0.12°C per decade (Table 4). Although the spreads among the trends from the different datasets are 0.05°–0.11°C per decade, and the differences in the trends are probably at the margins of statistical significance, there is a clear pattern in these regional trends of greater net annual mean warming since 1950 than since 1900. However, a contribution to this from multi-decadal variability appears likely. The tendencies for the NA and NWA summer trends also hold for these shelf-slope regions: the summer trends are larger than the annual ones (particularly for the NSS during the 1900–2011 period), and the trends for the 1950–2011 period are larger than for the 1900–2011 period. There is a general tendency for the MTZ-S trends to be larger than those for the LSS and NSS. The MTZ-S trends are also larger than those for off-shelf parts of the MTZ (OSW and OSE in Table 4), but it is unclear whether this is an indication of a possible northward expansion of the subpolar gyre (e.g., Wu et al., 2012). There is a discrepancy by a factor of 1.5–2 in the HadISST1 and ERSST trends for the MTZ-S during these periods, particularly in the annual means. This difference between these two commonly used datasets is consistent with those mentioned earlier in relation to Loder et al. (2013), the P5 and EB trends (Table 2) and during the 1870–2011 period (Fig. 7), but its origin is unclear.

During the 1979–2011 period, there is a common pattern of dramatically increased warming over most of the NA in all three datasets, in both the annual and summer means (Fig. 7). However, as noted earlier, there is pronounced spatial variation with enhanced warming in subpolar and northern areas, and little (if any) warming in some areas in either the annual or summer means. For the Atlantic Canadian shelf-slope regions, the rates are highest in the LSS where they are 0.44° and 0.55°C per decade for annual and summer means, respectively, and decrease progressing equatorward through the NSS to the MTZ-S where they are 0.21° and 0.36°C per decade for the annual and summer means, respectively (Table 4). These increased warming rates in recent decades probably reflect, in part, anthropogenic global warming, but it appears that they also include a contribution from natural variability associated with the AMO (see next section; also, Galbraith and Larouche, 2013), such that they should be interpreted and used with caution.

It can also be seen from the detailed spatial patterns of the trends (Fig. 7) that there are variations on local scales among the gridded datasets and periods that will affect the trends at particular sites. These probably account for the differences at the DFO sites (Table 2) being of order 0.1°C while those (above) in the spatially averaged SST in the regions of interest (Table 4) are in the 0.05°–0.1°C per decade range. It is noteworthy that the trend values of the region-averaged SST means (Table 4) are considerably different from the region-averaged ones computed from the trends at the individual grid points, further reflecting the grid-scale and interannual variability in the datasets together with the non-linearity of the linear regression technique used in the trend computation. These points indicate that caution is required regarding the spatial representativeness of the trend values from the gridded datasets at particular positions.

6 Modes of variability since 1900

In this section we report on EOF analyses on the three gridded datasets for the 1900–2011 period, starting with the annual means and then proceeding to the summer means. The time series at all positions were first detrended, as a first approximation to removing the long-term anthropogenic warming signal described earlier in relation to Figs 5 and 6. Although there are indications that this signal includes multi-decadal variability, the linear trends in Fig. 7 are removed following common practice. By only including data since 1900, this approximation should be less problematic. Results from corresponding EOF analyses for the 1950–2011 and 1979–2011 periods can be found in Loder et al. (2013): those for the 1950–2011 period are similar to those for the 1900–2011 period, whereas those for the 1979–2011 period are different because the period is not long enough to capture the AMO-like multi-decadal variation that dominates in the other periods.

a Annual Means

The EOF patterns for the three leading modes from the detrended annual means for the 1900–2011 period are shown in Fig. 8a, after an adjustment of the COBE modes (see the EOF mode 2 subsection below). The PC time series for these modes are shown in Fig. 9a.

Fig. 8 EOF patterns (°C normalized by the peak value) for the (a) annual and (b) summer means of the three datasets for the 1900–2011 period, with the percentage of explained variance indicated in the inset panels (and the suspect annual COBE mode and its variance not included).

Fig. 9 PCs for (a) the annual mean and (b) the summer mean SST analyses of the three gridded datasets after detrending for the 1900–2011 period.

1 EOF Mode 1

The EOF1 patterns from the different datasets show considerable similarity, even more so than those from the 1870–2011 time series without prior detrending (Fig. 4). The anomalies have the same sign over almost the entire domain, although they have greater spatial structure than in the case without detrending (as expected with the removal of the spatially coherent basin-warming trend). The EOF1s still account for 31–35% of the variance (after adjustment of the COBE modes), even though this is less than in cases without prior detrending (again, this is as expected because the EOF1s without detrending include the large-scale long-term warming signal). The EOF1s have a quad-pole pattern with the largest anomalies in the predominantly subpolar region extending from approximately 40° to 65°N and peak anomalies in the general area of the North Atlantic Current to the east and northeast of the GB, which suggests that this mode may be related to gyre–gyre interactions (and possibly the AMOC). The smallest anomalies (but still of the same sign) are in the subtropical gyre in the western NA and in some high latitude areas (>65°N). These EOF1 patterns are similar to those for the AMO in Knight et al. (2005) and Wang et al. (2012; see Fig. 1 in each).

The PC1s (Fig. 9a) for the different annual mean datasets are very similar to each other, including a prominent multi-decadal variation that closely resembles the AMO index (Fig. 2). This variation shows a prolonged maximum from the 1930s to the 1950s, a prolonged minimum from the mid-1970s to the early 1990s, and a maximum from the late 1990s to the present. Because EOF1 accounts for more than twice as much variance as any other EOF during the 1900–2011 period (and also the 1870–2011 and 1950–2011 periods), it can be considered to be the predominant mode of SST variability in the NA during the past century on decadal to multi-decadal time scales. The peak (local) range of this multi-decadal variation is about 2°C in the subpolar NA, compared with a range of approximately 0.5°C for the basin-averaged AMO (Fig. 6). The correlation coefficients between the AMO index and the PC1s exceed 0.9 (Table 5). Considering these correlations and the similarities of the EOF1 spatial patterns to the AMO ones in Knight et al. (2005) and Wang et al. (2012), it can be argued that the overall EOF1 pattern and multi-decadal PC1 time variation in Figs 8a and 9a, respectively, are credible representations of the most commonly used version of the AMO (without contributions from GMST removed).

Table 5. Correlation coefficients between the annual mean SST PCs in Fig. 9 and potential forcing indices after detrending. The correlations with the AMO and winter NAO indices are for the 1900–2011 period, while those with the SSH and WSC indices from the Wang et al. (2015) hindcast simulation are for the 1958–2004 period. Trends significant at the 95% confidence level are in bold. Be-Br refers to Bermuda SSH minus Bravo SSH

SST PCs	AMO	Winter NAO	Hindcast SSH Indices				Hindcast WSC PC1
			Be-Br	PC1	PC2	PC3
PC1 HadISST1	0.95	−0.20	0.02	0.36	0.65	0.46	0.36
PC1 ERSST	0.92	−0.21	−0.02	0.37	0.68	0.46	0.36
PC1 COBE	0.91	−0.23	−0.04	0.40	0.61	0.47	0.35
PC2 HadISST1	−0.04	−0.52	−0.41	0.66	−0.36	0.01	0.53
PC2 ERSST	−0.01	−0.48	0.06	0.55	0.05	0.01	0.76
PC2 COBE	−0.02	−0.46	−0.34	0.65	−0.28	0.02	0.58
PC3 HadISST1	−0.09	0.16	−0.35	−0.05	−0.51	0.00	−0.44
PC3 ERSST	−0.01	−0.19	−0.48	0.22	−0.64	0.07	−0.12
PC3 COBE	−0.22	−0.10	−0.39	0.05	−0.66	−0.10	−0.22

With the interpretation that mode 1 represents the AMO, comments can be made regarding the contributions of the AMO to SST variability in the NWA. The LSS and NSS regions of interest lie in or near the broad area in the subpolar NA with the largest EOF1 anomalies, and the MTZ region lies in the gradient zone between this area and the subtropical one with the smallest anomalies. The area of relatively large AMO anomalies extends west of the GB (albeit with reduced magnitude) into the MTZ which distinguishes this mode from the previously noted NAO-forced variation in ocean temperature that has opposite phases in the subpolar gyre and west of the GB (Deser et al., 2010; Petrie, 2007). The PC1 variations indicate that the AMO contributed to below-average SSTs in the subpolar gyre and its extension to the SS and Scotian Slope between the mid-1970s and early 1990s and to above-average SSTs in this region since the mid-1990s. The EOF1 patterns indicate that the AMO has a weaker influence on SST to the southwest of the SS and GoM (e.g., in the Middle Atlantic Bight), but there are differences among the datasets in the detailed pattern in this area.

The contributions of this AMO-like mode to SST variability at the DFO sites, obtained by multiplying each PC1 time series by the corresponding EOF1 anomaly at each site, are shown in Fig. 10a, and the correlation coefficients between the AMO contributions and the (weakly) low-pass filtered versions of the DFO annual means are included in Table 6. It is clear that the AMO (first mode) makes a substantial contribution to the very low-frequency (multi-decadal) variability at OWS Bravo, S27, and EB but only a small contribution at P5. It accounts for 50–59%, 26–36%, 22–28%, and 2–3% of the total variance in the filtered time series at these sites, respectively. These modest fractions are not surprising considering the higher-frequency variability in the DFO time series at these sites (even after partial filtering of the interannual variability). The range of the multi-decadal AMO variations at the sites varies from approximately 0.8°C at OWS Bravo and S27, to approximately 0.5°C in EB and <0.3°C at P5. There is approximate consistency among the AMO contributions from the different datasets at Bravo, S27, and EB, but the differences are comparable to the range of the AMO variation at P5 (Fig. 10a).

Fig. 10 The low-pass filtered (a) annual mean and (b) summer mean anomalies of upper-ocean temperature (black circles and lines) at the four DFO monitoring sites, together with the contributions to local SST variability from AMO-like mode 1 during the 1900–2011 period in the three detrended gridded datasets (coloured lines: HadISST1 (HD), ERSST (ER), and COBE (CB)).

Table 6. Correlation coefficients between the low-pass filtered upper-ocean annual mean temperature anomalies from the four DFO monitoring sites (Figs 10 and 11) and the contributions of the first (AMO-like; left columns), second (NAO-linked; middle columns), and third (ocean-origin; right columns) modes from the EOF analyses of the three gridded datasets during the 1900–2011 period. Both the DFO and gridded dataset indices were detrended before the computations. Coefficients statistically significant at the 95% level are indicated in bold. See Table 4 for acronyms.

Station	PC1 (AMO-like)			PC2 (NAO-linked)			PC3 (Ocean-origin)
	HD	ER	CB	HD	ER	CB	HD	ER	CB
P5	0.17	0.16	0.13	0.37	0.36	0.35	0.19	−0.05	0.09
EB	0.47	0.53	0.49	0.24	0.26	0.21	0.17	0.06	0.05
S27	0.51	0.60	0.53	0.09	−0.14	−0.12	−0.09	−0.06	−0.21
Bravo	0.73	0.77	0.71	0.10	0.31	0.01	0.27	0.23	0.30

2 EOF Mode 2

The annual mean EOF2s (Fig. 8a) also show a quad-pole pattern that is similar across the datasets, and the PC2 variations are also similar (Fig. 9a), after the adjustment of the COBE modes. The dominant feature of the original EOF2 for the COBE dataset was a large anomaly along the Greenland coast, and the corresponding PC2 had an abrupt jump around 1979 which appeared to be unphysical and likely a methodological artifact (Loder et al., 2013). Consequently, this “suspect” mode has been omitted from the results presented in Figs 8 and 9, with the original COBE mode 3 presented here as the adjusted COBE mode 2, and the original COBE mode 4 presented as the adjusted COBE mode 3. With this adjustment, there is close similarity between the adjusted COBE EOF2 and PC2 and the EOF2s and PC2s from the other datasets. Similarly, there is close similarity between the adjusted COBE EOF3 and PC3 and those from the other datasets. The adjusted EOF2s account for 12–14% of the total variance in the different datasets after adjustment of the COBE modes (with the 15% variance of the contaminated mode not included in the total for COBE).

In the western NA, the predominant feature of the revised EOF2s is a dipole structure between the subpolar waters to the north and northeast of the GB and an area approximating the MTZ but including the GB (Fig. 8a). This out-of-phase temperature pattern, together with the positive PC2 values during the 1960s (Fig. 9a), are reminiscent of the extension of relatively cool subpolar water west of the GB associated with the period of negative NAO (and subpolar gyre relaxation) in the 1960s, at a time when the temperature north of the GB was above average (Deser et al., 2010; Marsh et al. 1999; Petrie and Drinkwater, 1993). The correlation coefficients between the detrended winter NAO index and the PC2s for the gridded datasets are about −0.52 (Table 5), which provides some support for the previous suggestions that periods of negative NAO result in cooling west of the GB (Petrie, 2007; Petrie and Drinkwater, 1993), and periods of positive NAO result in cooling in the LS (Yashayaev, 2007). However, the modest values of these coefficients indicate that the NAO is not the only factor contributing to the basin-scale EOF2 mode. The PC2s have more decadal-scale and multi-year variability than the PC1s, consistent with a more complex origin for EOF2. Correlation analyses with various SSH indices from a hindcast model study of NA variability during the 1958–2004 period (Wang et al., 2015) indicate that the SST PC2s have correlations of 0.55–0.66 with SSH PC1 and 0.53–0.76 with the PC1 of the WSC over the NA (Table 5), suggesting a contribution from broader-scale ocean circulation variability associated with large-scale wind stress forcing.

Based on the EOF2 patterns, this NAO-linked second SST mode can be expected to make out-of-phase contributions to SST variability in the LSS and MTZ-S regions of interest, with the variations in the NSS region less clear because of the dependence of the detailed structure there on the dataset. For the DFO monitoring sites, there are small contributions to the DFO SST variability at P5 (12–14% of the variance) and EB (4–7% of the variance) with ranges of approximately 0.5°C, but little contribution at S27 and Bravo (Fig. 11, Table 6). There are indications of small contributions from this mode to the cool conditions in the SS–GoM region during the 1960s in all three datasets and to the cool conditions in the LS in the early 1990s in two of the datasets. But, it is apparent that this mode does not account for the regional intensity of these events which is not surprising in view of the large domain used in the EOF analysis and the limited correlation of this mode with the NAO. It is noteworthy that the contribution at OWS Bravo from the corresponding NAO-linked mode 2 from the EOF analyses on the NWA domain (Loder et al., 2013; their Table 5) is larger (14–18% of variance) than that from the NA mode 2 discussed here, which points to some sensitivity of the detailed spatial and temporal structure of the mode to the domain used in the EOF analysis.

Fig. 11 Low-pass filtered annual mean anomalies of upper-ocean temperature (black circles and lines) at the four DFO sites, together with the contributions to local SST variability from the NAO-linked mode 2 during the 1900–2011 period in the three detrended gridded datasets (coloured lines: same acronyms as in Fig. 10).

3 EOF Mode 3

The annual mean EOF3s, accounting for 8–10% of the total variance after the COBE adjustment, are dominated by a tri-pole pattern (Fig. 8a). The largest anomalies occur in a zonal band that extends eastward from the GB and expands latitudinally in the 40°–55°W region. There are weaker opposite-sign anomalies in zonal bands to the north (including the LS) and south in the 20°–40°N range. The boundary area where the mode changes sign (between the southern and mid-latitude poles) generally lies near the GoM and south of the GB, whereas that between the mid-latitude and subpolar poles lies over or north of the GB. This raises the possibility that this mode is also connected with the previously described variability in the extension of subpolar water west of the GB (e.g., Petrie, 2007).

The PC3s have substantial decadal-scale and multi-year variability (like the PC2s), as well as notable interannual variability (more than the PC2s) (Fig. 9a). The correlation coefficients between the PC3s and the AMO and NAO indices have magnitudes ≤0.22 (Table 5), indicating that these influences are not primary factors in this mode. The interpretation and origin of this mode are unclear at this point, beyond the peak amplitude east of the GB indicating that the mode could also be related to gyre–gyre interactions associated with the North Atlantic Current. Support for this conjecture comes from correlation coefficients of −0.51 to −0.66 (depending on the dataset) with the second SSH mode from the Wang et al. (2015) hindcast simulation for the 1958–2004 period (Table 5) and, more specifically, of −0.35 to −0.48 with the SSH difference between Bermuda and OWS Bravo in the simulation.

The correlation coefficients between the PC3s and the low-pass filtered DFO annual anomalies (Table 6) indicate a small contribution (5–9% of the total variance) from the third mode to the observed variability at Bravo and little contribution (<4%) at the other sites. This result differs slightly from that in Loder et al. (2013) in which it was found that the similar mode 3 from EOF analyses on the NWA domain contributed to 3–11% of the variance at P5 (but <7% at OWS Bravo). This also indicates some sensitivity of the local contributions of the modes to the size of the domain over which the EOF modes are computed. The present results are thus inconclusive with respect to the importance of mode 3 to variability in annual mean SST off Atlantic Canada, but EOF3's centre of action to the east of the GB (Fig. 8a) suggests some importance in the outer GB and Flemish Cap region (in addition to that identified here in the LS).

b Summer Means

In this subsection we briefly discuss results from EOF analyses of the summer means during 1900–2011, with particular interest in the summertime persistence and any summer amplification of the annual mean modes.

The EOF patterns for the summer means are shown in Fig. 8b and the corresponding PCs in Fig. 9b (in this case, there was no suspect COBE mode in the leading three modes). There is substantial similarity to the revised EOFs for the annual means (Fig. 8a), but there are also considerable differences. In particular, although the summer EOF1 modes closely resemble the annual EOF1 modes, the summer EOF2s and EOF3s are switched in relation to the annual EOF2s and EOF3s (i.e., the summer second mode closely resembles the annual third mode, and the summer third mode closely resembles the annual second mode). This is consistent with the findings in Section 4 for the HadISST1 and ERSST modes during the 1870–2011 period (without prior detrending) and simply means that one mode is more important in the annual means and the other in the summer means. Comparison of the annual and summer PC2 and PC3 time series (Fig. 9) supports this interpretation. It can be seen from the inset values in Fig. 8 that EOF1 accounts for a lower percentage of the total variance in the summer (20–27%) than in the annual means (31–35%); the summer EOF3 also accounts for a lower percentage (10%) than its counterpart annual EOF2 (12–14%), whereas the summer EOF2 (11–14%) accounts for a higher percentage than the annual EOF3 (9–10%). This may point to intensification in the origin of the last mode in summer, but the significance of this is unclear.

Focusing on the lower-frequency variability, the summer PC1s confirm the predominant multi-decadal variation, but also have more multi-year variability than the corresponding annual PC1s and a larger range for this variation, at least at some locations (Fig. 9). The correlation coefficients between the summer PC1s and the detrended annual AMO index are 0.82–0.86 (compared with 0.91–0.95 for the annual PC1s; Table 5), suggesting that the AMO is not particularly associated with summer processes. In contrast, the correlations between the summer PC3s and the detrended winter NAO (−0.01 to −0.30) are substantially smaller than those for their counterpart annual PC2s (−0.46 to −0.52), supporting a non-summer origin for this mode (as expected with winter NAO forcing). The correlations between the summer PC2s and the AMO and NAO indices are weak and mostly insignificant, which is consistent with the corresponding correlations for the annual PC3s, pointing to some other origin for this mode. The relation of the summer version of this mode to the SSH indices from the hindcast simulation has not been pursued, because only annual indices were available for the latter.

Off Atlantic Canada, there are small differences among the spatial structures of the summer and annual EOF1s, but overall they are similar (Fig. 8). With the inclusion of the seasonal ice-covered areas off Labrador and in Baffin Bay in the summer HadISST1 domain, the broad amplified positive anomaly of EOF1 over the subpolar NA extends across the LS to the Labrador Shelf and Baffin Bay, similar to that for ERSST and COBE. The contributions from the summer EOF1 to the observed variability at the four sites (Fig. 10b) are comparable (in terms of percentage variance explained) to those from the annual EOF1. Collectively, this indicates that the AMO has been contributing to warming from the SS to as far north as Davis Strait during the past two decades, suggesting that the warming trend in the DFO observations in recent decades (e.g., Colbourne et al., 2014; Galbraith, Larouche, Chassé, & Petrie, 2012) and in the 1979–2011 SST trends (Fig. 7) for these areas includes a contribution from natural variability. The contributions of the summer EOF2 and EOF3 to the observed variability at the four sites (not shown) are less than for the corresponding annual modes, suggesting that other factors are more important to local variability in summer.

Before concluding this section, several comments are appropriate. First, large-scale modes contribute to regional SST variability as expected. Second, it appears that multiple modes contribute to the variability at some sites, even on similar time scales, again as expected (de Viron, Dickey, & Ghil, 2013). Third, a large portion of the variability in the DFO (annual mean and summer mean) time series cannot be explained by the leading modes examined here, probably because of a combination of the predominance of higher-frequency (e.g., interannual) variability in the time series and of some of the decadal-scale variability having only regional scales. Corollaries of this are that the large-scale modes are more important to multi-year and lower-frequency variability than indicated by the percentages of explained variance, and forcings such as the NAO may make additional contributions to local variability through modes not examined here (or through local effects).

7 Summary

Standard trend and EOF analyses for the NA have been carried out using three global gridded datasets of monthly SST: HadISST1, ERSST, and COBE. Analyses were carried out for four periods (1870–2011, 1900–2011, 1950–2011, and 1979–2011), and for annual and summer means. Time series from the gridded datasets were compared with observations from DFO monitoring programs at four sites off Atlantic Canada, and PCs from the EOF analyses were compared with potential forcing indices.

Although there are many similarities in the results from the different gridded datasets, there are also notable differences that are probably related to data sparsity in some areas and different methods, interpolations, and corrections to and sources of the data. In particular, a data quality issue was identified with the COBE SSTs around the Greenland coast, and there is a clear indication that the magnitudes and spatial structure of the long-term (1870–2011 and 1900–2011) trends from the various datasets need to be used with caution. Comparisons with the DFO time series indicate qualitative similarities in the trends and variability among the gridded and monitoring datasets at all four sites and also notable quantitative differences.

For the 1900–2011 and 1950–2011 periods, there are average annual mean warming rates of only approximately 0.04°C per decade and approximately 0.02°C per decade for the NA and NWA, respectively. The smaller latter value arises because of a large area of cooling or reduced warming in the northern NA south of Greenland. There is warming on the Atlantic Canadian shelf/slope in all the datasets, with average magnitudes (across datasets) in the range of 0.04° to 0.12°C per decade depending on location. The spread of the trends from the different datasets is approximately 0.1°C per decade for local sites and in the range of 0.05°–0.10°C with spatial averaging across sub-regions. During the 1979–2011 period, the trends are substantially larger (regional averages of 0.2°–0.4°C per decade), and the relative differences among the datasets are reduced, with the latter presumably because of increased data coverage. Summer trends are generally larger than annual trends but not in all shelf/slope areas.

The most robust results for the 1900–2011 period (and also for the 1950–2011 period; Loder et al., 2013) appear to be the remarkable similarity of the first EOF mode in the three annual mean datasets and the predominance of a multi-decadal variation in its PCs corresponding to a local range of up to 2°C in SST. For detrended data, this mode closely resembles the AMO (e.g. Wang et al., 2012) and accounts for 31–35% of the variance in the annual means for the NA. This mode represents a basin-wide coherent SST variation but with a large difference in amplitude between its maximum in the subpolar gyre (and extending weakly west of the GB) and its minimum in the subtropical gyre in the western NA. Implications are that the influence of the AMO is greater east of the GB than to its west and that the AMO contributed significantly to the amplified warming in the subpolar NWA during the past three decades. The contributions of this leading EOF at the DFO sites indicate that the AMO is an important factor in the multi-decadal variability of local SST, except at P5. The EOF analyses on two of the datasets for the 1870–2011 period, without prior detrending, indicate that the long-term and multi-decadal variations also resemble those in GMST, suggesting that there are both global- and basin-scale contributions to low-frequency variability.

The pattern of the second EOF in the NA annual means (after exclusion of the suspect COBE mode) is also similar across the three datasets, accounting for 11–14% of the total variance. It resembles the out-of-phase changes in upper-ocean temperature north and west of the GB described by Petrie and Drinkwater (1993) and Deser et al. (2010) and attributed to the NAO. The PC2s for the 1900–2011 period are moderately correlated (r ∼ −0.5) with the winter NAO index, such that we have referred to this as an NAO-linked mode. Somewhat surprisingly, the local contributions of this mode are only weakly correlated (r ∼ 0.3) with the DFO time series from P5 and EB and not significantly correlated with those from S27 and OWS Bravo. Although the weak correlations are at least partly a result of the strong interannual variability at the various sites (that is not captured in the leading EOF modes), it also appears that this (or any) mode alone is not able to account for the magnitude of the spatial and temporal structure associated with particular regional events such as the cooling in the MTZ in the 1960s.

The EOFs and PCs for the third mode in the annual means for the 1900–2011 period are also similar across the three datasets, accounting for 8–10% of the variance. This EOF has a prominent centre of action near the North Atlantic Current extending eastward from the GB and an opposite phase in and to the south of the GoM. However, in the present analysis, it makes only a small contribution to the variability in the DFO time series at OWS Bravo and little at the other sites, again probably due to a combination of the predominance of interannual to decadal variability in the local observations and single modes not being able to capture the intensity of episodic regional events. Its PCs are weakly correlated with the sea level difference between Bermuda and Bravo (r ∼ 0.4) in a hindcast simulation of NA circulation during the 1958–2004 period (Wang et al., 2015) and with the second mode of SSH variability (r ∼ 0.6) in that hindcast. This suggests that this mode is related to variability in ocean circulation and to interactions between the subpolar and subtropical gyres in particular.

The pattern and amplitudes of the summer EOF1s are similar to those for the annual mean EOF1s for all three datasets during the 1900–2011 and 1950–2011 periods, and the corresponding PC1s are similar across the datasets and periods. The PC1 correlations with the AMO are lower for the summer means than for the annual means indicating that, although summer SST variability in the NA includes a significant AMO-like component, the AMO primarily originates in other seasons.

The EOF2s and EOF3s, as well as the PC2s and PC3s, for the summer means are similar to those for the annual means during the 1900–2011 period, but their order is switched with respect to the percentage of total explained variance for the NA. For example, the NAO-linked mode ranks second in percentage variance explained in the annual means but third in variance explained in the summer means. Because these differences are small, the main point is that the NAO-linked and ocean-origin modes are similar in the annual and summer means.

In conclusion, it should be emphasized that, although our results should help place upper-ocean temperature variability off Atlantic Canada in an NA-wide and broader climate variability perspective, many questions and issues remain regarding the space–time scales and origins of the apparent natural variability and the magnitudes and spatial patterns of anthropogenic changes. Further analyses with observational and model data for other ocean and climate variables are needed. In the meantime, particular caution is required in the use of ocean temperature (and other variable) trends from available NWA datasets of limited duration because of the strong decadal-scale natural variability in the region.

Acknowledgements

We are grateful to many personnel over the years who have contributed to the collection, processing, and analysis of the various datasets used here. We would especially like to thank Roger Pettipas for providing the AZMP datasets and assisting with the analyses of various datasets, Igor Yashayaev for providing the AZOMP time series, Eugene Colbourne for the Station 27 time series, Augustine van der Baaren for earlier analyses of HadISST1 and ERSST data that provided motivation for the present study, and Ingrid Peterson for ongoing helpful discussions regarding SST variability and its origins. Finally, we thank Dave Hebert and Ingrid Peterson for providing internal review comments, and we also thank two anonymous external reviewers.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This work was carried out with support from the Atlantic Trends and Projections activity of DFO's ACCASP.