The two annual wine production quantity series from Grevenmacher and Remich in the Moselle Valley (49.5° N, 6.35° E) of present-day Luxembourg are based on official records of a tax of one-ninth levied on the wine production. They cover the 1444–1786 period, though with a gap between 1684–1741 (except for the years 1698–1701) in connection with the French occupation of Luxembourg . It is unfortunate that much of the cold period coinciding with the Maunder Minimum of solar activity ca. 1645–1715; is missing in the Moselle Valley wine production data. However, the last 3 decades of the 16th century were at least as cold as any part of the Maunder Minimum . Thus, we still capture the full range of temperature variability.

While the tax was originally levied, when introduced in the 13th and 14th centuries, on every ninth basket of grapevines harvested, it soon instead became levied on every ninth barrel of wine produced. However, the number of barrels of wine produced closely followed the number of baskets of grapevines harvested. Thus, the number of barrels of wine produced mimics the quantity of the vine harvest. It is uncertain to what extent changes over time in tax exemptions affected the wine quantity at inter-annual and longer timescales. Furthermore, at times, the tax appears in fact to have been both higher and lower than one-ninth of the wine production. The degree of fraud and resistance to taxation might also have varied over time see also. Sufficient to say is that the long-term trend in the Grevenmacher and Remich series is not reliable. The source-critical problems associated with tithe records to estimate actual agricultural productivity variations have been extensively treated in the agrarian history literature e.g.. This research on tithes as a historical source for agricultural productivity has shown that long-term trends contain larger biases than short-term variability. Grevenmacher shows a long-term rather stable, or slightly decreasing, production trend, whereas Remich shows a long-term production increase (Fig. a). Nevertheless, the correlation coefficients between the Grevenmacher and Remich series for linearly detrended and 11-year high-pass filtered data, respectively, are r=0.55 and r=0.62. Both correlations are significant at a p=0.01 level (two-tailed test).

Modern observations show that temperature is the single most important climatic factor for viticulture in the Moselle region . The soil in Grevenmacher is limestone, and the soil in Remich is alluvial. Grain is the main crop, but terraces of vines are widespread too . The annual mean temperature during the 20th century was around 10 °C with around 650 mm annual precipitation and slightly over 1400 hours of sunshine. Years with less than 500 mm precipitation tend to produce the best wine. Spring frost is a known problem for viticulture in the region . The Moselle Valley during the early modern period was a rather marginal viticulture area (about 90 % of the wine was white) with often relatively low wine quality, with the wine of Remich considered superior to that of Grevenmacher. Consequently, the vine was grown to a much larger extent in Remich than in Grevenmacher .

We employ the volcanic stratospheric sulfur injection data from (lists provided by Michael Sigl, University of Bern, personal communication, 2022), an updated version of  that is utilised for the recommended forcing in state-of-the-art PMIP4 “past1000” climate model simulations . This dataset provides the forcing year, not the eruption year (which is sometimes unknown), and, depending on the seasonality of the eruption, the peak forcing year can be the same year as the eruption year or the following year. used composite ice-core data from Greenland and Antarctica, separately, to generate a robust list of dates of volcanic eruptions. We only use the dates for forcing events with aerosol loading in the Northern Hemisphere (NH) (flux is provided as the deposited sulfate found in the ice cores given in kg km⁻²). We restrict our list to those eruptions, since it is likely that even strong Southern Hemisphere (SH) extra-tropical eruptions do not contribute much to an aerosol loading in the NH.

For use in the superposed epoch analysis, we identify key years with NH volcanic forcing of various strengths. In the NH flux series, for the period we use in the analysis, there are 34 (27) eruptions listed with strengths above 0 kg km⁻² (the smallest eruption in the series is near 3.2 kg km⁻²), 28 (22) eruptions stronger than 5 kg km⁻², 16 (12) eruptions stronger than 10 kg km⁻², 11 (7) eruptions stronger than 15 kg km⁻², and 9 (6) eruptions stronger than 20 kg km⁻². Here, the numbers in parentheses give the count of eruptions that also cover complete wine data, in the sense that we only include eruptions not less than 5 years from a wine data gap. In our analyses, we use all the eruptions, as recent results suggest that weak eruptions can have significant climate effects . However, we also tested the effects of using only eruptions larger than the thresholds mentioned above. The eruption years included are listed in Table . In addition, we tested the sensitivity of the SEA by progressively excluding the strongest volcanic eruptions.

Instrumental climate data have only a short period of overlap with the Grevenmacher and Remich wine production series. The closest instrumental temperature station to Luxembourg in the homogenised European Climate Assessment (ECA) temperature dataset is the monthly averaged data for De Bilt in the Netherlands at a distance of about 300 km. Data from De Bilt overlap with the wine production series from 1742 to 1786, covering a period of only 45 years. The short overlap with the De Bilt instrumental record limits the reliability and confidence in results derived from it.

We therefore employ reconstructed series of temperature and precipitation, as summarised in Table . We use seasonally resolved gridded reconstructed temperature data from and (henceforth only cited as ) and seasonally resolved gridded reconstructed precipitation from back to 1500 (Table ). Both the and the reconstructions include instrumental data during the 18th century. Data are extracted from the local grid cell(s) covering Luxembourg. In addition, we use the seasonal resolved reconstructed temperature data from , which are available as a central European average for the same period. For the full period since 1444, we use the June–August temperature reconstruction by , as updated by , based mainly on tree-ring data but also containing the documentary-based June–August season data from the abovementioned reconstruction. Furthermore, we employ for the full period the solely tree-ring-based NTREND gridded field reconstruction of May–August temperature. For comparisons against volcanic-induced cooling in the Alpine region, we use the June–September temperature reconstruction from Lötschental, Switzerland, which is derived from a strong temperature signal in maximum latewood density tree-ring series . June–August soil moisture (drought) conditions for the entire period are obtained from the tree-ring-width-based gridded Old World Drought Atlas providing self-calibrated Palmer Drought Severity Index (scPDSI) values .

The episodic nature of the volcanic eruptions is an advantage when analysing their effects on variables with unknown, or even dubious, low-frequency variations and trends, such as wine production quantities. We simply compare the strength of the mean superposed standardised signal of the considered time series (wine production or climate conditions) in the years immediately following the eruptions with the mean superposed standardised signal in the years before (5 years). Using superposed epoch analysis SEA;, we calculate the average of this difference across K eruptions, which are covered by the considered time series – so the number of eruptions might vary slightly for different time series – and are stronger than the chosen threshold. We standardised the time series to zero mean and unit variance before conducting the SEA. The significance is estimated by a bootstrap procedure, assuming only that the eruptions are independent. If we have K eruptions, we pick K random years in the whole period under consideration and calculate the superposed epochs from these years. Note that the years with actual volcanic eruptions should be selected with the same probability as any other year. Otherwise, a bias would be introduced. We do this 3000 times to get a distribution of the SEA under the null hypothesis that volcanic eruptions have no impact on the considered time series (e.g. wine production). Comparing the original SEA with this distribution, we can calculate the p-value as the probability of having a lower SEA than the observed. We report significance in the SEA at both the p=0.05 and p=0.01 significance levels. Furthermore, our results are robust to changes in the details of the analysis, such as in the length of the epochs or whether the average is calculated over the whole epoch and not just the half preceding the eruptions. The SEA results are based on time series standardised to unit variance over the whole period. However, similar results are obtained when the standardisation is performed individually over running windows (corresponding to variance equalisation).

To further investigate the relationship between climate and wine production quantities, we calculated the Pearson correlation coefficient (r) between the wine production series and the various (palaeo)climate reconstructions. The significance of correlations was calculated using p=0.05 with a phase-scrambling test, which is more conservative than the common Student's t-test for details, see. To emphasise inter-annual variability, we employed 11-year high-pass filtering of the data. For studying low-frequency variability, we likewise employed an 11-year low-pass filter. This emphasises variability at decadal scales, while inter-annual fluctuations and noise are suppressed. The choice of 11 years as the filter cut-off length is chosen to match the timescale of the SEA. Although we have used a simple 11-year smoothing filter, other more complicated filters (e.g. Gaussian filters) provided almost identical results. During an exploratory data analysis phase, we also compared the use of the parametric Pearson correlation coefficient with the non-parametric Kendall and Spearman correlation coefficients, finding that they resulted in similar correlation patterns.

A main focus of this study is the effects of volcanic-forced abrupt cooling on the Moselle Valley wine production. Superposed epoch analysis (SEA) reveals a strong decrease in wine production the year following volcanic forcing events for both Grevenmacher and Remich. The SEA signal itself is shown in Fig.  with the blue curves, while the dark- and light-shaded bands indicate the p=0.01 and p=0.05 significance levels, respectively. This volcanic-related production decrease pattern is similar regardless of whether all volcanic events are included or only those exceeding certain thresholds for NH forcing flux (as demonstrated in the different panels in Fig. ). In addition, progressively excluding the strongest volcanic eruptions one by one had only a minor impact on the results (not shown). Conclusions were unchanged.

In general, we find a strong, significant, and consistent decrease in wine production in the years following volcanic forcing events for both the Grevenmacher and Remich series. For all thresholds, the signal (blue curve in Fig. ) is about 0.5 standard deviations and the p-value is below 0.05 and often below 0.01 (shaded regions in Fig. ). Slightly smaller signals are found for Remich compared to Grevenmacher. Including all eruptions, we find the largest signal in the first year after the eruptions. The (taxed) wine production was even zero, or close to it, in the first year following certain major volcanic events: 1481, 1601, 1674, 1767, and 1784. However, when increasing the threshold and only including stronger (and fewer) eruptions, we see a tendency for the strongest wine production decrease to be delayed. For the largest eruptions, the strongest signals appear 3–4 years after the eruptions.

The bands indicating the significance levels expand with increasing threshold due to the fewer eruptions included (Fig. ). When only including larger volcanic events, exceeding an estimated NH forcing flux of >10, a wine production decrease that is significant at the p=0.01 level is only found the third year after the volcanic event – and only for Grevenmacher. Weaker, less significant decreases are also detected the first and second years after the volcanic event, with the Remich series showing a broad decrease for years 1–3 after such strong eruptions.

We have performed SEA from volcanic forcing events on temperature and precipitation series (local grid cells when applicable) from the Moselle Valley (see Table ). Figure  summarises the strengths of the connection by reporting the smallest p-values (two-tailed test) taken over years 0–5, i.e. from the year of the eruptions and the following 5 years. These p-values basically reflect the strength of the SEA signals.

We find significant connections for the temperature series mainly in spring and summer. For example, p-values below 0.02 are found for the temperature series by and only for the March–May and June–August seasons. For temperature in spring and summer, the lowest p-values are found for lag 0 or 1 year (numbers shown to the right in Fig. ). The significant signals in the temperature series all correspond to coolings. Similar results are obtained using only eruptions larger than 5 kg km⁻², while the situation for higher thresholds (and correspondingly fewer eruptions) is more muddled. For precipitation, significant signals corresponding to wetter conditions are found in annual and September–November reconstructions.

We emphasise the weak and insignificant volcanic signatures in the tree-ring-based temperature reconstructions (EuroMed2k, NTREND, and Lötschental) for the central Europe region. This less distinct volcanic signature stands in sharp contrast to the strong and clear wine production quantity reduction after volcanic forcing events that reaches significance well below the p=0.01 level for several years even when including minor forcing events (Fig. ).

We have identified the 20 most extreme years (for each of the four seasons) for a selection of the climate series in Table . Thus, for temperature series we find the 20 warmest and coldest years, and for precipitation we find the 20 wettest and driest years. These extremes are identified from the raw series without detrending. We then conducted SEA to identify the responses in the two wine production series to these four kinds of extremes for each season. The results are shown in Table , which reports the p-values and the signs of the significant responses.

We find that the spring (MAM) stands out from the other seasons with significant responses for all four kinds of extremes. We also note that consistent directions of the responses are found, increasing wine production for warm/dry conditions and decreasing production for cold/wet conditions. The climate series from and show consistent results. Furthermore, we find almost identical results for the Remich and Grevenmacher wine series. For the other seasons, fewer significant and less consistent responses are found, indicating that the significance in these seasons might be partly due to chance.

Through correlation analyses we explored the long-term relationships between wine production and seasonal and annual temperature, precipitation, and soil moisture. We started the analyses with De Bilt instrumental series. The correlation between the De Bilt instrumental series and the Grevenmacher series, over their 45-year period of overlap, 1742–1786, is r=0.46 using 11-year high-pass filtered data for JJA temperature for both. For Remich, the result is r=0.29 (just below significance). Furthermore, the annual mean temperature for De Bilt shows significant positive correlations with Grevenmacher and Remich of r=0.46 and r=0.42, respectively, using 11-year high-pass filtered data. No other seasons, neither for linearly detrended data nor for high-pass filtered data, reveal any significant correlation between wine production and the De Bilt instrumental temperature series.

Using reconstructed temperature data covering the entirety, or the majority, of the length of the Grevenmacher and Remich wine production series shows similar, but more significant, correlation patterns as found using the short De Bilt instrumental series. Mostly significant correlations are found between temperature and wine production quantities between both 11-year high-pass filtered and linearly detrended data (Fig. ). These correlations are, in general, slightly stronger for Grevenmacher than for Remich. The most significant association is the positive relationship between spring, summer, and autumn temperature and between the annual mean and wine production (i.e. warm corresponds to high wine production and vice versa). Correlations are weak for winter temperatures, being entirely outside the growing season (Fig. ). Significant positive correlations are also found between (palaeo)climate series and the Grevenmacher series using 11-year low-pass filtered data. While some of the correlations are slightly stronger than those for high-pass filtered or linearly detrended data, the fewer degrees of freedom render a lower number of them significant. Notably, for Remich, not a single correlation using 11-year low-pass filtered data is significant, although most of the correlations are positive and of comparable strength to those for high-pass filtered and linearly detrended data.

Negative relationships, in general less significant than those with temperature, are found in the hydroclimate between precipitation and soil moisture and wine production (i.e. wetter corresponds to lower wine production and vice versa) for both 11-year high-pass filtered data and linearly detrended data. Significant relationships with hydroclimate are mostly restricted to the summer season and to the annual mean. We note that, while the correlation-based results provide the strongest signal for summer, the SEA-based results using extreme years instead show the strongest signal in spring. Nevertheless, both analyses reinforce the finding that warm and dry years and growing seasons benefited the wine production, while cold and wet years had an adverse effect. For hydroclimate, no significant correlations are found with either the Grevenmacher or the Remich series using 11-year low-pass filtered (palaeo)data.

Our results are robust to choices regarding the exclusion of weaker volcanic forcing events and to the exclusion of individual particularly strong volcanic forcing events. Although wine tithe data amount to zero, or close to it, in the first year following certain major volcanic events (1481, 1601, 1674, 1767 and 1784), this does not necessarily mean that the vine harvest was zero or close to zero. From grain tithes, for example, in the early modern Swedish Realm , we know that sometimes no tithes were collected during years with extremely low harvests. Thus, zero wine tithe indicates very low vine harvests rather than necessarily no harvest at all. Furthermore, we cannot rule out the possibility that the respective error term is skewed towards a higher resistance to taxation in years of bad harvests. Such non-linearities are very difficult to assess and can similarly affect other types of records. Nevertheless, the long-term trend in the wine tithe data is more uncertain than the short-term (inter-annual to inter-decadal) variability. Since the SEA method focuses on a 5-year period before and after each event, it is not sensitive to potential long-term trends in the data. The compositing in SEA constitutes an averaging process that serves as a filter enhancing the high-frequency response signal of interest while minimising noise and also accounts for long-term drifts see e.g.. We have also ensured that the response statistics to volcanic forcing are not driven by just a few major eruptions by investigating the effects of including all volcanic forcing events and of including only those exceeding different threshold values. The results are also robust to using forcings less than a threshold. We have also tested the effects of excluding only the 1783 Laki eruption (not shown) and found that this volcanic forcing event does not dominate the results, although the forcing of Laki in the Northern Hemisphere is multiple times stronger than that of any other eruption.

Regarding the climate–wine production relationships, we can emphasise that we have tested both linear and non-linear correlation methods (that is, Pearson correlation coefficient versus rank correlation methods such as Kendall and Spearman), although we only report the results of the first, as they all give similar results. The similar results imply that the non-linearity that is indeed present in the wine data (which cannot go below zero) and the various non-linearities present in palaeoclimate reconstructions have not prevented the linear Pearson correlation from rendering an accurate picture of how the series are correlated. The varying correlation strength between wine production and the different temperature reconstructions, even for the same season, is not surprising considering that the reconstructions have employed diverse input proxy data and different methods to combine the data and “calibrate” the reconstructions against instrumental temperature measurements (see, for example, ; ).

Reduced wine production for up to several years following larger volcanic forcing events can, in part, be attributed to the fact that strong eruptions cause aerosols to linger for a longer period in the atmosphere and induces feedback mechanisms in the climate system that prolong cooling e.g. sea-ice feedbacks;. In part, it can also be attributed to a biological memory effect, and cold and frost damage, of the vine stocks during years with very poor growth conditions . Note that, while we do observe an immediate signal following larger volcanic events, the SEA results suggest that the maximum response is delayed. Biological memory effects are very well documented for tree growth . In particular, tree-ring width depends to some degree on growth conditions in the previous year, and resulting climate reconstructions often show more auto-correlation than the instrumental target . As expected, auto-correlation values (AR1) reported for the time series used in this study are higher if biological archives are involved compared to documentary-based records (see Table ). However, while a limited immediate response to abrupt volcanic cooling has been reported for many tree-ring width records , as opposed to maximum latewood density (MXD) records , this seems not to be the case for the wine production series. Thus, in this respect, wine production shows a similar behaviour to MXD, potentially indicating that other physiological processes control fruit production than stem increment.

The wine production decline in Grevenmacher and Remich following volcanic forcing events contrasts to the weak volcanic response for central Europe in the state-of-the-art gridded NTREND tree-ring-based temperature reconstruction , the mainly tree-ring-based updated EuroMed2k temperature reconstruction , and the MXD-based Lötschental temperature reconstruction . The NTREND reconstruction for Europe is primarily composed of MXD data from tree rings, a temperature proxy commonly assumed to be minimally affected by biological memory . Despite remaining discussions regarding the precise quantification of volcanic signals in MXD-based reconstructions , there has been repeated evidence for significant cooling in MXD-based reconstructions following single events and groups of eruptions at local , regional , and hemispheric scales.

In contrast to a moderate volcanic cooling impact over central Europe, NTREND and EuroMed2k reveal strong and widespread cooling in response to tropical eruptions over Scandinavia, with a distinct transition between these two regions over the southern Baltic Sea region at around 55^∘N. However, NTREND and EuroMed2k do not contain tree-ring data in Europe between 50 and 60° N, making it difficult to determine the exact location of this transition. The gridded seasonal temperature reconstruction for Europe by , using predominantly documentary data, suggests significant summer cooling that extends as far south as the northern fringe of the Alpine Arc , which is in better agreement with the volcanic impact found here in wine production series from the Moselle Valley. The lack of climate-sensitive tree-ring records from central Europe (north of the Alpine Arc) is also expressed in reduced values for explained variance at the Moselle grid cell compared to the explained variance for e.g. northern Fennoscandia.

Comparison between wine production declines and tree-ring growth declines is complicated by the limited number of major volcanic forcing events during the comparatively short period covered by the two wine production series. Numerous studies of temperature-sensitive tree-ring data, from northern Europe and elsewhere, have shown a maximum growth decline (growing season cooling) during the summer in the year following the volcanic forcing event . Such a distinct cooling 1 year post-eruption is most evident in MXD data, compared to TRW data, owing to the typically stronger correlation with temperature and the larger biological memory in TRW . Thus, the post-volcanic decrease detected in this study in wine production quantities is very similar to the post-volcanic growth decline in MXD. Nevertheless, the most noteworthy feature in the wine production data is the strong and consistent response 1 year following the volcanic forcing event, even for a low NH flux threshold (>0 and >5), that is absent in tree-ring-based temperature reconstructions for central Europe. At the mid-latitudes, where the wine production data diverge from the tree-ring-based temperature reconstructions, no reconstructions show a significant summer temperature reduction in response to a low NH forcing flux threshold.

The Moselle Valley region is situated at the northern limit of viable viticulture areas in the same way as northern Scandinavia is situated close to the Arctic tree line. Hence, both the Moselle Valley grapevines and the Scots pine trees in northern Scandinavia are thus very sensitive to temperature drops and presumably also to reduced sunlight after volcanic forcing events. The high climate sensitivity in both data types is clearly related to the marginality of the respective locations for grapevines and conifer trees. The higher climate sensitivity of the grapevines of Grevenmacher compared to Remich may also be due to different soil conditions rather than merely to the small difference in latitude. We also suggest that the explanation for the stronger, and more persistent, climate response in the wine production data than in the tree-ring data is due to physiological differences. In addition to the consistent signal 1 year following the volcanic forcing event, there is a “late response” in the wine production data after big flux events (thresholds >15 and >20). This late signal is also present in the entirely tree-ring-based NTREND and (weakly) in the Lötschental temperature reconstructions (not shown). It is more difficult to come to a conclusion about this part of the signal because there are only few events falling into this category. While it would be interesting to study the climate (and volcanic) response to tree growth and wine production from the same locations, it would be challenging because temperature-sensitive and/or precipitation-sensitive tree-ring series are rarely available from vine growing regions, with the exception of regions close to the Alps. Nevertheless, such studies could provide insights into the decorrelation length of climate versus differences in the climate sensitivity of the vine growth and tree-ring growth.