A summer climate regime over Europe modulated by the North Atlantic Oscillation

. Recent summer heat waves in Europe were found to be preceded by precipitation deﬁcits in winter. Numerical studies suggest that these phenomena are dynamically linked by land-atmosphere interactions. However, there exists as yet no complete observational evidence that connects summer climate variability to winter precipitation and the relevant circulation patterns. In this paper, we investigate the functional responses of summer mean and maximum temperature (June–August, T mean and T max ) as well as soil moisture proxied by the self-calibrating Palmer drought severity index ( scPDSI ) to preceding winter precipitation (January– March, P JFM ) for the period 1901–2005. All the analyzed summer ﬁelds show distinctive responses to P JFM over the Mediterranean. We estimate that 10 ∼ 15% of the interan-nual variability of T max and T mean over the Mediterranean is statistically forced by P JFM . For the scPDSI this amounts to 10 ∼ 25%. Further analysis shows that these responses are highly correlated to the North Atlantic Oscillation (NAO) regime over the Mediterranean. We suggest that NAO modulates European summer temperature by controlling winter precipitation that initializes the moisture states


Introduction
The recent European climate is characterized by an increasing frequency of summer heat waves with potentially substantial societical and ecological impacts, e.g. the recordbreaking heat wave in 2003(van Oldenborgh et al., 2008;;Della-Marta et al., 2008).Climate projections point towards even higher-frequent and longer-lasting heat waves under increased greenhouse gas emission scenarios (Scherrer et al., 2005;Pal et al., 2004;Stott et al., 2004;Meehl et al., 2004).These past and projected heat waves highlight the importance of a detailed understanding of the mechanisms that contribute to the initialization and persistence of extreme heat conditions.Hot and dry summers in Europe are generally associated with a specific large-scale anticyclonic atmosphere circulation regime (Cassou et al., 2005;Fischer et al., 2007).It has also been noted that most of the hot and dry summers over Europe were preceded by pronounced deficits of precipitation in the winter and early spring (Della-Marta et al., 2007;Vautard et al., 2007).Using numerical experiments, Vautard et al. (2007) showed that the observed winter precipitation deficit and summer heat wave are dynamically linked via feedback loops between land and atmosphere.In these feedback loops soil moisture plays a critical role.In these studies, a deficit of precipitation and subsequent drier spring soils resulted in reduced latent cooling and a corresponding increase of summer air temperature, in agreement with other numerical experiments (e.g.Seneviratne et al., 2006;Fischer et al., 2007;Zampieri et al., 2009).
These investigations of individual heat waves have highlighted the role of land-atmosphere coupling, and also pointed to the importance of circulation patterns, in the generation of summer heat waves.An immediate question that Published by Copernicus Publications on behalf of the European Geosciences Union.
arises is whether this land surface feedback mechanism exists only for extraordinary hot summers or whether it exists more systematically.Schär et al. (2004) emphasized that an increase of interannual temperature variability in response to greenhouse-gas forcing might be an alternative causal mechanism for the occurrence of European summer heat waves.Subsequent numerical analysis by Seneviratne et al. (2006) suggested further that the increased interannual temperature variability is strongly related to the land-atmosphere coupling.
However, there exists as yet no complete analysis of available observational evidence that connects the interannual variability of summer temperature to winter precipitation.The present paper aims to fill this gap in our understanding by investigating the relations of summer mean as well as maximum temperature and winter precipitation using longterm observations.Furthermore, to investigate the possible mechanisms in more detail, a soil moisture analysis is also presented.The paper is organized as follows: in Sect. 2 the observational datasets used are described and the statistical technique is briefly introduced.Section 3 is dedicated to the results and, finally, Sect. 4 contains a discussion and the conclusions of this study.

Datasets
We use long-term observations of accumulated precipitation in January-March (P JFM ) and averaged daily mean as well as maximum temperature in June-August (T mean and T max respectively) for the period 1901-2005.Due to the sparseness of in situ soil moisture observations, the averaged selfcalibrating Palmer drought severity index in June-August (scPDSI; Wells et al., 2004), is used as a proxy of soil moisture.The scPDSI is based on soil water content in a rather complex water budget model involving water cycle interactions with temperature; therefore it is suitable for our study purpose of land-atmosphere coupling.Ideally one would use remotely sensed soil moisture observations (e.g. de Jeu et al., 2008), but these datasets are unfortunately not yet sufficiently long in time.The scPDSI dataset spans 1901-2002 on a monthly basis and ranges from −4 to +4 in the case of extremely dry and extremely wet conditions respectively (van der Schrier et al., 2006).All datasets, gridded at a horizontal resolution of 0.5 • × 0.5 • , are derived from University of East Anglia Climatic Research Unit (CRU; Mitchell et al., 2005).The serial mean values over each pixel are removed to obtain anomalies.Data values over mountain Scandinavia are not included in this study.

Coupled Manifold Technique
We use a technique, the Coupled Manifold Technique (CMT) recently proposed by Navarra and Tribbia (2005), to investigate the functional relations between fields of interest.Let S and Z stand for two fields and suppose they are linked by a linear relation.Their relation may then be expressed in terms of data matrices, as (1) S = S for + S free = B Z + S free .
A and B are matrix representations of linear operators that express the relations between S and Z.Using the Procrustes method (Richman et al., 1993), the CMT seeks for A and B with , where the primes denotes a matrix transpose operation.They are generally not equivalent: A expresses the influence of S on Z, and B expresses the influence of Z on S. Hence, the Z field can be separated into two parts using A: The Z for part is the portion of Z variability forced by S, and Z free is the portion independent from S. The same equation can be solved for S for and S free using B: S free = S − B Z.
The Z for and S for portions can be further decomposed by writing Eqs.(3) and (4) into the right-hand side of Eq. (1): The portions of ABZ and BAS represent the respective manifolds of Z and S that are fully coupled with each other.From the physical perspective there are two possible mechanisms to generate these fully coupled manifolds.One is the existence of an external mechanism that simultaneously influences both S and Z fields.The other is the possibility of reciprocal feedbacks between S and Z.Since S free is a portion free from Z, the AS free portion represents the Z variability purely forced by S, that is, it gives us the directional influence (henceforth the "forced manifold").Similarly, the BZ free portion represents the directional influence of Z on S. We emphasize that these relations are functional and can not be interpreted directly as causal chains.For a physical interpretation, we have to rely on existing knowledge of the physics and dynamics of climate.
In our case time lags exist between the analyzed climatic fields, and of interest are only the directional portions representing the variability of summer T mean , T max as well as scPDSI in response to that of P JFM .Given the winter precipitation field S and a summer field Z, we have to solve for the AS free portion in the Z for = ABZ + AS free equation.In our case we note that the ABZ portion contains very little variance of Z for , making AS free equivalent to Z for .This is particularly valid for T mean and T max , and statistically confirms the lack of an external mechanism exerting influence on both winter precipitation and summer temperature.The CMT in our study is applied to the coefficients of empirical orthogonal functions (EOFs) of the analyzed fields to simplify the computation (Alessandri et al., 2008), with 99% of the total variance of each field retained.To retain only the significant relations, each element in A and B is tested against the null hypothesis of being equal to zero at the 1% significance level using the Student t distribution described by Cherchi et al. (2007).

Responses of T mean and T max to P JFM
Figure 1a shows the percentage of the T mean variance forced by the P JFM variability, and that for T max is shown in Fig. 1d.These values are derived from the ratio of the forced T mean (T max ) manifold to the original T mean (T max ) fields, namely, the (pixelwise) ratio values of AS free /Z, where S is winter precipitation and Z is T mean or T max .We also tested where the percentage values are significantly different from zero at the 0.10 level.For each grid point, the null hypothesis of getting as high or higher variance fractions is tested through a Monte Carlo bootstrap method (10 000 repetitions of the CMT) by randomizing the order number of P JFM values on each grid point.The largest values are found over the Mediterranean for both T mean and T max , where we expect the largest sensitivity to land-atmosphere coupling (Seneviratne et al., 2006;Fischer et al., 2007;Zampieri et al., 2009); little forcing (low values) is obtained over northern Europe of 50 • N. Up to 5 ∼ 15% of the summer T mean variance over the Mediterranean appears to be forced by P JFM .The forced T mean variance by P JFM is up to 8% over western Mediterranean, averaged within the green rectangle in Fig. 1a.Over eastern Mediterranean, this value increases to 11% averaged within the red rectangle in Fig. 1a, passing the significance test.These values for T max are a bit higher.The forced T max variance is up to 10% over western Mediterranean and that value over eastern Mediterranean is up to 14%, averaged within the green and red rectangles respectively in Fig. 1d.Low values for both T mean and T max over North of 50 • N indicate little influence from P JFM .
The P JFM field and the forced T mean as well as T max manifolds, that is, AS free and Z in Eq. ( 5), are further decomposed into the associated spatial patterns and time coefficient series with the Maximum Covariance Analysis (MCA).Figure 1b  and c show the patterns of P JFM and its T mean response from the 1st MCA mode, which contains 95% of the total squared covariance.The time coefficient series of this P JFM − T mean mode, shown as blue lines in Fig. 3, have unit correlation (r > 0.99).This high correlation value indicates that the derived MCA mode is very robust.We emphasize that the unit correlation derived here is due to the data preprocessing with CMT, which constructs only the T mean variability forced by P JFM at significance level of 0.01.The time coefficient series of the 1st P JFM − T mean MCA mode without CMT exhibits a correlation of 0.40 (not shown), which is insufficient to conclude a significant and robust linkage.The same situation also holds in the following analysis of T max as well as soil moisture proxy of scPDSI.There exist only significant P JFM anomalies over the Mediterranean.The T mean responses are of the opposite sign, and appear to extend northward and eastward compared to the P JFM anomalies.This indicates that the summer T mean fluctuates to the anomalous states of winter P JFM but in an opposite phase.However, precipitation is nearly a white noise process with very limited memory due to the chaotic nature of atmosphere (Wang et al., 2010), and thus winter precipitation cannot really persist into summer solely through its atmospheric memory.For some very dry years, Vautard et al. (2007) has demonstrated with numerical experiments a causal mechanism of soil moistureatmosphere coupling that sustains the signal of winter precipitation into the summer temperature field.Our analysis suggests that this mechanism exists not only for a limited number of very dry years but perhaps more systematically.
The 1st MCA mode for T max analysis is shown in Fig. 1e  and f.This mode contains 96% of the total squared covariance, and the time coefficient series have unit correlation (r > 0.99).The P JFM pattern and its T max response are quite similar with those derived from T mean analysis.Furthermore, the time coefficient series of this P FJM − T max MCA mode exhibit nearly unit correlation (r > 0.99) with those of the P FJM − T mean MCA mode.These statistical properties suggest that the responses of both T max and T mean to P JFM are very likely to be driven by the same climate dynamics, i.e., the land-atmosphere coupling as earlier numerical studies suggested (Vautard et al., 2007).An open question regarding land-atmosphere coupling is whether this leads to amplified variability of climate extremes, such as heat waves, particularly in the context of climate change (Seneviratne et al., 2010).The robust relations derived here enable us to compare the magnitudes of T mean and T max responses to P JFM .The magnitude of the T max response appears to twice that of T mean , suggesting that P JFM exerts larger influence on T max than that on T mean , possibly through land-atmosphere coupling, in accordance with suggestions from the modeling studies.

The role of soil moisture
So far we have shown that T mean and T max fluctuate in response to winter precipitation over the Mediterranean.The underlying mechanism is thought to be the land-atmosphere coupling and an analysis of soil moisture variability would help us to find support for this hypothesis.For this purpose the same analytic framework as above is applied to P JFM and summer scPDSI as a soil moisture proxy.Fig. 3.The P JFM pattern shows distinctive anomalies over the Mediterranean, nearly identical to those patterns forcing T mean and T max .The scPDSI anomalies are of the same sign over the Mediterranean.Furthermore, the time coefficient series for this scPDSI analysis have unit correlation (>0.99) with those obtained from the temperature analysis (Sect.3.1).This indicates a strong possibility that the responses of scPDSI, T mean and T max to the P JFM variability are driven by the same climate dynamics.Although our analysis cannot directly infer the precise causal mechanisms involved, P JFM is very likely to influence T mean and T max via soil moisture from the physical perspective.That is, a negative precipitation anomaly in winter can easily result in summer heating because of resulting decreased latent cooling from soil moisture.The reverse relationship also holds, when a positive precipitation anomaly implies cooling.The MCA analyses for T mean , T max and scPDSI are summarized in Table 1.
Our observational relations corroborate the interactions between the water cycle and temperature established in previous numerical work (e.g., Seneviratne et al., 2006).Note that if we perform the same set of statistical analysis to the winter precipitation and summer minimum temperature, we do not obtain the same relations.This is physically reasonable because the minimum temperature is highly constrained by external forcings, e.g., solar radiation and atmosphere circulation, rather than land surface processes (Alfaro et al., 2006;Zhang et al., 2008).

Link to North Atlantic Oscillation
The North Atlantic Oscillation (NAO) is the dominating large-scale atmospheric circulation over the Atlantic-European sector in winter, with a marked influence on winter climate.In recent years, the NAO has been shown to influence summer climate over Europe, in a weak but significant way.Qian et al. (2003) observed a significant link between summer temperature over UK and the preceding later winter NAO index, where seasonal expansion in the Azores high pressure system may play an important role.The NAO is also reported to have prediction skill for some hydrological features in European summer, e.g., precipitation (Kettlewell et al., 2003) and discharge (Bierkens et al., 2009).The derived forcing P JFM patterns on T mean and T max as well as scPDSI in our analysis appear to resemble the NAO regime over Mediterranean, suggesting a hypothesis that the NAO variability modulates summer climate over Europe by controlling winter precipitation that subsequently initializes the moisture states of land-atmosphere coupling.
To obtain more insight into this possibility, we compared the winter NAO index and the time coefficient series derived from the above MCA analyses.We use the averaged values of NAO index in January-March for the period 1901-2005, based on the difference of normalized sea level pressures between Gibraltar, Azores and SW Iceland.As shown in Fig. 3, the time coefficient series from T mean , T max and scPDSI analyses are highly correlated to the NAO index with r = 0.65 (p < 0.05), indicating a significant relation between the NAO variability and the analyzed summer fields.The NAO variability is a north-south shift (or vice versa) in the track of storms and depressions across the North Atlantic Ocean and into Europe.The Atlantic storms that travel into Europe result in a dry Mediterranean Europe during a high NAO winter and the opposite happens during a low NAO winter (Hurrell et al., 2001).Based on the above analyses, we suggest that the NAO regime over the Mediterranean modulates European summer climate via precipitation and subsequently the initial states of land-atmosphere coupling.

Discussion and conclusion
The importance of soil moisture initialization in winter and early spring for the seasonal prediction of heat and drought waves in European summer has been demonstrated in recent years (e.g., Vautard et al., 2007;Fischer et al., 2007;Zampieri et al., 2009;Seneviratne et al., 2006;Ferranti et al., 2006).Although soil moisture is closely related to precipitation, a complete observationally based analysis of the relations between summer climate and preceding winter precipitation was not yet available.This was largely caused by the weakness of the expected signals in the fields of interest.
Traditional techniques for cross-correlation, such as MCA and CCA, are not capable of generating robust relations in the presence of strong background noise.
Using the CMT technique, we present in this paper the robust responses of summer T mean and T max as well as scPDSI to previous winter precipitation.Distinctive responses exist only over the Mediterranean area, where the temperature response is most sensitive to land-atmosphere coupling in regional climate models (Schär et al., 1999;Seneviratne et al., 2006).The P JFM variability accounts for up to 10 ∼ 15% of the total T mean and T max variance respectively for the period of 1901-2005; for the scPDSI this value amounts to 10 ∼ 25%.Although there is no direct evidence from our statistics, the robustness of the results suggests that the P JFM appears to influence T mean and T max via scPDSI.This agrees well with our understanding of the water cycle dynamics over land (Seneviratne et al., 2010).Therefore our findings are very likely to be a physical signal of land-atmosphere coupling, although we are generally cautious to infer physics from statistics alone.We are also not addressing the full picture of land-atmosphere coupling but only that part that is related to late winter precipitation.
In our statistical analyses, the summer climate responses appear to be extended towards north and east.This is also observed in numerical experiments.Vautard et al. (2007) suggested that the northward extension may be due to the southerly wind episodes carrying moisture northward.However, the eastward extension is probably due to the heat low response over Central Europe, blocking the inflow of moist maritime air from the Atlantic and reinforcing the northward extension dynamically, as addressed by Haarsma et al. (2008).Using a moisture tracer model, Bisselink and Dolman (2009) also found that advection is the most important contributor to precipitation over central Europe.
We suggest that the NAO regime over the Mediterranean modulates summer climate over Europe through controlling winter precipitation that then initializes land-atmosphere coupling.A positive phase of NAO tends to generate the possibility of a hot and dry summer, or vice versa.Thus there is a scope for improved seasonal prediction of heat and drought waves from the pressure pattern of winter NAO.A remarkable feature of the NAO is its prolonged positive phases in the past 40 years, possibly related to anthropogenic warming (Shindell et al., 1999).Our analysis suggests that this NAO dry pattern over the Mediterranean may have contributed to the increased frequency of heat and drought waves since then.
Although the focus of this study is the Mediterranean area, the out-of-phase T mean and T max responses as well as the in-phase scPDSI responses to P JFM appear to also happen in the UK.Over the Mediterranean area, there exist plenty of numerical supports that the directional relations we have obtained are likely to have physical origin.However, these responses observed in the UK need more careful interpretation.Qian et al. (2003) attributed the significant correlation between winter NAO index and summer temperature over the UK to atmospheric circulation.Meanwhile, correlation between the NAO index and summer precipitation was found by Kettlewell et al. (2003), which might suggest the existence of local hydrological processes over land.In our analysis, the CMT is capable of disentangling the directional influences between two fields; however, by taking into account the remote effect it is not capable of eliminating the nonphysical tele-correlation in one single field.Therefore we suggest that the observed summer T mean , T max and scPDSI anomalies over the UK are not necessarily related to the winter precipitation regime over the Mediterranean area.

GFig. 1 .
Fig. 1.T mean as well as T max variability forced by P JFM .(a) Percentage of T mean variance forced by P JFM (sig = 0.10 in the red rectangle).The spatial patterns of (b) P JFM and (c) its T mean response.(d) Percentage of T max variance forced by P JFM (sig = 0.10 in the red rectangle).The spatial patterns of (e) P JFM and (f) its T max response for the 1st MCA mode.All the relevant time coefficient series mutually exhibit unit correlation (r > 0.99), shown in Fig 3. Units are K for T mean as well as T max and mm for P JFM .

The
scPDSI variability forced by P JFM is shown in Fig. 2. Shaded values in Fig. 2a indicate the percentage of scPDSI variance forced by P JFM that can pass the significance test at the 0.01 level.The largest values of 10 ∼ 25% exist in the west Mediterranean.Shown in Fig. 2b and c are the spatial patterns of P JFM and its scPDSI response from the 1st MCA mode, containing 80% of the total squared covariance.The time coefficient series of this P JFM − scPDSI mode have unit correlation (r > 0.99), and are shown as green lines in

Fig. 2 .
Fig. 2. The scPDSI variability forced by P JFM .(a) Percentage of scPDSI variance forced by P JFM (sig = 0.01).The spatial patterns of (b) P JFM and (c) its scPDSI response for the 1st MCA mode.The MCA time coefficient series have unit correlation, shown in Fig. 3.

Fig. 3 .
Fig.3.The MCA time coefficient series and the NAO index.One blue line is used to plot the highly correlated time coefficient series (r > 0.99) from T mean and T max analyses.Green line indicates time coefficient series from the scPDSI analysis.The red line is the averaged NAO index in January-March.

Table 1 .
A summery of the MCA analyses between P JFM and the forced manifolds.