2.1 Czech climate reconstructions

The recent study uses the following climate reconstructions based on documentary data and instrumental observations related to the territory of the Czech Lands.

• a.

  Temperature reconstructions

  • i.

    Series of monthly, seasonal (DJF – winter, MAM – spring, JJA – summer, SON – autumn), and annual temperatures of central Europe (1500–2007 CE) based on temperature index series of Germany, Switzerland, and the Czech Lands (1500–1854) as well as mean instrumental temperature series of 11 meteorological stations in central Europe (1760–2007) (Dobrovolný et al., 2010)

  • ii.

    Series of March–June (MAMJ) temperatures of the Czech Lands (1501–2008 CE) derived from a series of winter wheat harvest dates (Možný et al., 2012)

  • iii.

    Series of April–August (AMJJA) temperatures of the Czech Lands (1499–2015 CE) derived from a series of grape harvest dates (Možný et al., 2016a)

• b.

  Precipitation reconstructions

  • i.

    Series of seasonal and annual precipitation totals of the Czech Lands (1501–2010 CE) derived from documentary-based precipitation indices (1501–1854) and mean areal precipitation series of the Czech Republic (1804–2010) (Dobrovolný et al., 2015)

• c.

  Drought reconstructions

  • i.

    Series of seasonal and annual drought indices (standard precipitation index, SPI – McKee et al., 1993; standard precipitation evapotranspiration index, SPEI – Vicente-Serrano et al., 2010; Z index and Palmer Drought Severity Index, PDSI – Palmer, 1965) of the Czech Lands for 1501–2015 CE (Brázdil et al., 2016) derived from central European temperature and Czech precipitation reconstructions (Dobrovolný et al., 2010, 2015)

  • ii.

    Series of AMJJA SPEI of the Czech Lands (1499–2012 CE) derived from a series of grape harvest dates (Možný et al., 2016b)

For the purposes of this paper, all of the above series were taken from 1501 and extended until 2020 to cover the entire 1501–2020 CE period. For comparison with MAMJ and AMJJA temperatures by Možný et al. (2012, 2016a), MAMJ and AMJJA temperature series from temperature series of central Europe (Dobrovolný et al., 2010) were calculated. Similarly, the AMJJA SPEI series from Brázdil et al. (2016) was calculated for comparison with that of Možný et al. (2016b).

2.2 European climate reconstructions

To study the spatiotemporal representativeness of Czech climate reconstructions at the European scale, gridded European reconstructions are used.

• a.

  The reconstruction of gridded ($\mathrm{0.5}{}^{\circ }\times \mathrm{0.5}{}^{\circ }$) seasonal temperatures by Luterbacher et al. (2004) and Xoplaki et al. (2005) covering European lands (25^∘ W–40^∘ E; 35–70^∘ N) in the 1500–2002 period is spliced from temperature-sensitive natural and documentary proxy-based reconstructions before 1900 and instrumental measurements from Mitchell and Jones (2005) after that time (https://www.ncei.noaa.gov/pub/data/paleo/historical/europe-seasonal.txt, last access: 20 October 2021).

• b.

  The reconstruction of seasonal precipitation by Pauling et al. (2006) includes gridded ($\mathrm{0.5}{}^{\circ }\times \mathrm{0.5}{}^{\circ }$) totals for European lands (30^∘ W–40^∘ E and 30–71^∘ N) for the years 1500–1900 reconstructed from long instrumental precipitation series, documentary-based precipitation indices, and natural proxies (tree rings, ice cores, corals, and speleothems) combined with a gridded reanalysis for 1901–2000 after Mitchell and Jones (2005) (https://www.ncei.noaa.gov/access/paleo-search/study/6342, last access: 20 October 2021).

• c.

  The reconstruction of summer self-calibrated (sc) PDSI for The Old World Drought Atlas (OWDA) by Cook et al. (2015) includes gridded ($\mathrm{0.5}{}^{\circ }\times \mathrm{0.5}{}^{\circ }$) data derived from tree ring widths for the 0–2012 CE period (http://drought.memphis.edu/OWDA/Default.aspx, last access: 20 October 2021).

Moreover, from gridded values of the three mentioned gridded European reconstructions, the mean series for the “central European window” with geographic coordinates 45–54^∘ N and 5–23^∘ E was calculated and used for comparison with the Czech series applying 31-year running correlation coefficients.

2.3 External forcings and large-scale climate variability modes

The following descriptors of external forcings and large-scale internal oscillatory climate variability modes are used as explanatory variables in the attribution analysis.

• a.

  Greenhouse gas radiative forcing (GHGRF)

  Based on Meinshausen et al. (2011) annual data for the 1765–2020 period extended back to 1501 CE using the CO₂, CH₄, and N₂O concentrations obtained from the online database of the Institute for Atmospheric and Climate Science, ETH Zurich, and approximate formulas provided in the IPCC report (IPCC, 2001, Table 6.2)

• b.

  Solar activity (SOLAR)

  Annual values of total solar irradiance by Lean (2018), extended to 2020 by data at https://climexp.knmi.nl/data/itsi_ncdc_yearly.dat (last access: 20 October 2021)

• c.

  Volcanic activity (VOLC)

  Stratospheric volcanic aerosol optical depth (AOD) series in the 30–90^∘ N latitudinal band, adapted from reconstruction by Crowley and Unterman (2013), extended to 2020.

• d.

  North Atlantic Oscillation (NAO)

  Series by Luterbacher et al. (2001) are available for 1659–2001 CE in monthly time steps and for 1500–1658 CE in seasonal time steps. Beyond 2001, NAO index values were calculated from the standardized pressure difference between Iceland and the Azores using NCEP/NCAR reanalysis data (Kalnay et al., 1996).

• e.

  Atlantic Multidecadal Oscillation (AMO) and Pacific Decadal Oscillation (PDO)

  Annual values of multidecadal temperature variations in the AMO and PDO regions by Mann et al. (2009) were adopted for the 1501–2006 CE period and extended to 2020 by GISTEMP (Hansen et al., 2010) areal temperature means for their respective northern Atlantic and northern Pacific regions. To overcome problems with the strong mutual correlation of Mann et al. (2009) AMO and PDO temperatures, their common component (designated AMO+PDO) and difference (AMO–PDO) were used instead of the AMO and PDO series themselves (following from predictor analysis presented in Mikšovský et al., 2019). The common component (AMO+PDO) was further detrended by subtracting its component correlated with greenhouse gas radiative forcing to more reliably separate signals related to these two predictors.

4.1 Climate fluctuations in 1501–2020 CE

4.1.1 Series derived from temperature and precipitation indices

Fluctuations in annual temperature, precipitation, and drought indices in the Czech Lands during the 1501–2020 period exhibit great interannual variability and prevailingly small non-significant linear trends (Fig. 1). Only for mean annual temperatures is the increasing trend statistically significant at the 0.05 significance level (0.11 ^∘C per 100 years), when temperatures after relatively stable fluctuations grew from ca. 1890, and their increase was particularly enhanced starting in the 1970s. The last 30-year period of 1991–2020 experienced the highest temperatures in the whole series, while the coldest 30-year period was detected in 1829–1858 (Table 1). Very similar fluctuations characterize series of annual precipitation totals and SPI series derived from precipitation, experiencing no long-term trends. Both series agree in the wettest 1912–1941 period, while the driest 30-year episode occurred in the first 3 decades of the 18th century (with a small shift between the two variables). Three remaining series of drought indices show non-significant negative trends and agree in the driest 30-year interval of 1990–2019. For the wettest 30 years, the Z index and PDSI agreed with the precipitation series during 1912–1941 (for the PDSI with a shift of 1 year), while in the SPEI it was already in the second half of the 16th century (1569–1598). The means of all selected 30-year extreme periods differ significantly from the means of the corresponding entire 520-year series.

Figure 1Selected annual climate variables in the Czech Lands during the period of 1501–2020 CE: (a) fluctuations smoothed by a 30-year Gaussian filter with linear trends and their numeric values (temperature: ^∘C per 100 years, precipitation: millimetres per 100 years, drought indices: index value per 100 years); extreme 30-year periods of each series are indicated by coloured bands and the lowest and highest values of series by small circles. (b) Box plots (median, lower and upper quartile, minimum, and maximum) for 1501–2020 and the two most extreme 30-year periods. The temperature and precipitation series are expressed as deviations with respect to 1961–1990.

Download

Table 1The warmest and driest (a) and the coldest and wettest (b) 30-year periods in annual and seasonal series of climate variables (CVs) in the Czech Lands in 1501–2020 CE: T – temperature, P – precipitation, SPI, SPEI, Z-in (Z index), and PDSI – drought indices.

Download Print Version | Download XLSX

Pearson correlation coefficients of annual temperatures with five other variables during the whole 1501–2020 period give statistically significant values between −0.27 with precipitation and −0.61 with SPEI. In terms of 31-year running correlations, they became statistically non-significant to a greater extent from the 19th century on, even changing signs of correlation from negative to positive approximately around 1900 (Fig. 2a). The close relationship of precipitation to drought indices with correlation coefficients from 0.59 with PDSI to 0.97 with SPI is well reflected in 31-year running correlations above the 0.05 significance level (Fig. 2b). The correlations among the four drought indices are the lowest between the SPI and PDSI (0.61) and the highest between the SPEI and Z index (0.96). None of the 31-year running correlations between drought index series dropped below the significance level (Fig. 2c).

Figure 2The 31-year running correlation coefficients between annual series of (a) temperature (T), (b), precipitation (P) and (c) drought indices (SPI, SPEI, Z index, PDSI) in the Czech Lands in the 1501–2020 period. Correlation coefficients for the whole period are in brackets. Dashed lines indicate 0.05 significance levels: correlations above and below these levels are statistically significant.

Download

Similar features as in the case of annual series can also be detected in the corresponding seasonal series (Figs. 3–6). All temperature series agree in increasing 520-year linear trends (the highest in DJF at 0.27 ^∘C per 100 years and the lowest in SON at 0.06 ^∘C per 100 years), with all being statistically significant except SON, and in the warmest last 3 decades 1991–2020 (DJF 1988–2017) (Table 1). A greater diversity appears in delimitation of the coldest 30-year periods: in the past 3 decades of the 16th century (DJF 1572–1601 and JJA 1569–1598), in the 18th century (SON 1757–1786), and in the 19th century (MAM 1832–1861). Seasonal series of precipitation totals and SPI indicate zero linear trends and great variety in 30-year extreme periods. Distinct clustering in their occurrence appears only for the wettest intervals in DJF (1555–1584) and JJA (1568–1597), while the other two seasons have maxima at the end of the 19th century and the beginning of the 20th century (MAM 1885–1914) as well as during the first decades of the 20th century (SON 1910–1939). The driest 30-year intervals appeared during the entire 18th century and only for SON in the 17th century (1605–1634). Compared to the series of precipitation totals, the SPI series showed different driest 30-year intervals in DJF (1680–1709) and partly shifted the wettest 30 years in MAM (1894–1923). The three remaining seasonal drought indices experienced statistically non-significant negative 520-year linear trends. The driest last 3 decades 1991–2020 are typical for all seasonal Z index and PDSI series (also for MAM and JJA SPEI). The wettest seasonal 30-year spans for the PDSI appear between 1912 and 1942 except MAM (1888–1917). Analogous intervals in the case of the Z index overlap with those in the PDSI only partly (starting earlier), and in JJA, it even occurs during the 16th century in 1569–1598, in agreement with the SPEI. The second half of the 16th century also experienced the wettest 30 years for SPEI in DJF (1555–1584), while those for MAM nearly overlap with the Z index and those for SON with PDSI. In total, different from the Z index and PDSI, the driest periods were in DJF (1680–1709) and in SON (1605–1634). Means of 30-year periods differed from the corresponding entire 520-year means only for the wettest and driest SPI in JJA and for the wettest SPI and SPEI in SON.

Figure 3See text in Fig. 1; DJF.

Download

Figure 4See text in Fig. 1; MAM.

Download

Figure 5See text in Fig. 1; JJA.

Download

Figure 6See text in Fig. 1; SON.

Download

Relationships between seasonal temperature, precipitation, and drought index series in the Czech Lands can be described using Pearson correlation coefficients in the entire 1501–2020 period, which are all statistically significant at the 0.05 significance level except for temperatures with four other variables in DJF (Table 2). Seasonal temperatures show the highest negative correlations with the SPEI (MAM, JJA, and SON) and SPI (DJF). As expected, the seasonal PDSI series shows the highest positive correlations in all seasons with the Z index and precipitation series with the SPI. The seasonal SPEI series indicates the highest correlations with SPI in all seasons except JJA; the same appears for the Z index series with SPEI except DJF. The seasonal SPI series exhibits the highest correlations with JJA and SON precipitation, while in the two remaining seasons, it is the best correlated with the SPEI. The maxima of the highest correlation coefficients for all variables occur in JJA (0.991 between precipitation and SPI). The minima of the highest correlations appear in DJF (for temperature and precipitation), MAM (for SPI and PDSI), and SON (for Z index). Temporal changes in the shared variability expressed with the running correlations show very similar features in all seasons as those for annual series (Fig. 2), and they are not shown here.

Table 2Pearson correlation coefficients between seasonal series of temperature (T), precipitation (P), and drought indices (SPI, SPEI, Z index, PDSI) in the Czech Lands during the 1501–2020 period (coefficients expressed in italics in brackets are statistically non-significant at the 0.05 significance level; all other coefficients are statistically significant).

Download Print Version | Download XLSX

4.1.2 Series derived from phenological data

For Czech reconstructions based on phenological data (Možný et al., 2012, 2016a, b), both temperature reconstructions agree in the warmest 30-year interval in 1991–2020 but differ in the coldest 30-year period: 1671–1700 in the reconstruction for MAMJ from winter wheat harvest dates and 1835–1864 in the reconstruction for AMJJA from grape harvest dates (Fig. 7). The first of this series also shows a statistically significant increasing linear trend (0.16 ^∘C per 100 years). The AMJJA SPEI series exhibits the driest 30 years in 1991–2020, while the wettest period occurred at the beginning of the 20th century (1900–1929). For this series, the driest period experienced much higher variability than the wettest, particularly documented by the interquartile range. The means of all selected 30-year extreme periods differed significantly from the means of the entire 520-year period.

Figure 7(a) Variability of MAMJ mean temperatures reconstructed from the winter wheat harvest dates (Možný et al., 2012) as well as AMJJA mean temperatures and SPEI reconstructed from the grape harvest dates (Možný et al., 2016a, b) in the Czech Lands during the period of 1501–2020 CE. Annual values are completed with the 30-year Gaussian filter, linear trends, and their numeric values (temperature: ^∘C per 100 years, SPEI: index value per 100 years); extreme 30-year periods of each series are indicated by coloured bands and the lowest and highest values of series by small circles. (b) Box plots express the median, lower and upper quartile, minimum, and maximum for 1501–2020 and the two most extreme 30-year periods. Temperature series are expressed as deviations with respect to the 1961–1990 period.

In Fig. 8, one can assess the agreement between the two temperature reconstructions derived from different documentary data (phenological series by Možný et al., 2012, 2016a, versus temperature indices by Dobrovolný et al., 2010) on different timescales. Even if the overall correlations between the two types of reconstructions are quite high and significant (0.67 for series derived from winter wheat harvest dates and 0.82 for those derived from grape harvest dates), 31-year running correlations reveal that the common signal varies substantially over time. Generally, it is lower before 1800 CE when the two compared series are represented by reconstructed values. Very high and significant correlations were also found for a relatively long period from the second half of the 16th century to the mid-17th century, which could perhaps be related to the higher quantity and quality of the available documentary evidence.

Figure 8(a) The 31-year running correlations (top) and low-frequency signal expressed as smoothed series by the 60-year spline function (bottom) compared to MAMJ temperatures reconstructed from the wheat harvest dates (Možný et al., 2012) and those reconstructed from temperature indices. (b) The same as (a) but for AMJJA temperatures reconstructed from the grape harvest dates (Možný et al., 2016a).

Whereas the running correlations allow us to compare the common signal on the annual and decadal timescales, low-pass filtering of the series with the 60-year splines reveals common features of multidecadal variability (Fig. 8). The long-term trend is quite consistent for AMJJA temperatures derived from grape harvest dates and from temperature indices. In contrast, smoothed winter wheat harvest date series show much higher long-term variability compared to the index-based reconstruction before 1800. The reconstruction from winter wheat harvest dates is well expressed, especially the period of low temperatures corresponding to the well-known Late Maunder Minimum of solar activity (1675–1715). This cold period is not as well expressed in the index-based temperature reconstruction because this central European reconstruction may partly smooth local effects.

4.2 Wavelet analysis

While strictly periodic components are typically not dominant in central European climate series beyond annual timescales, the presence of noteworthy unstable periodicities has been previously reported for some climatic characteristics (e.g. Brázdil et al., 2012; Mikšovský et al., 2019). As seen from the wavelet spectra of individual Czech series (Fig. 9), there are indeed several period bands in which notable (and sometimes statistically significant) oscillations exist. In the case of multidecadal variability, periodicities of approximately 70–100 years appear in several signals, although they are rather intermittent in terms of amplitude. In the case of documentary-based data, these can be detected in both temperature and precipitation series, as well as in the series of drought indices (only SPEI shown here: wavelet spectra of SPI are generally similar to those of precipitation, while Z index and PDSI resemble SPEI in their spectral structure). The presence of the ca. 70-year periodicity is particularly pronounced during JJA in precipitation and drought index series, whereas its statistically significant manifestations in other seasons are limited to shorter subperiods. The existence of 70–100-year oscillations is also supported by their appearance in the wavelet spectra of temperature and SPEI series reconstructed from wheat and grape harvest dates (bottom row of Fig. 9), although, again, these test as statistically significant only in part of the 1501–2020 period.

Figure 9Standardized wavelet power spectra for temperature, precipitation, and SPEI in the Czech Lands for the 1501–2020 period. Statistical significance is highlighted at the 95 % level (black line); series preprocessed by removing the GHGRF-correlated trend component.

Download

On shorter timescales, periodic components in the Czech climate series are typically even more scattered. Most notably, in both index-based and phenology-derived series, oscillations at periods of approximately 16–30 years are detected over some shorter subperiods. While these are typically statistically non-significant on their own over most of the analysis period, there are indications of interesting similarities to the spectral characteristics of several explanatory factors involved in our analysis. These are examined in more detail in Sect. 4.3.

4.3 Attribution analysis

A combination of regression analysis and wavelet transform was used here to identify and quantify links between reconstructions of Czech climatic characteristics and several potentially influential explanatory factors (for visualization of their temporal variability during 1501–2020 see Fig. 10; their mutual correlations are provided in Fig. S1 in the Supplement). The predictors used in our analysis (Sect. 2.3) exhibit only mild collinearity (with the strongest correlation detected between GHGRF and SOLAR at r=0.45). The results of linear regression (summarized in Fig. 11 through standardized regression coefficients and their confidence intervals) are therefore not substantively affected by variability shared by different predictors. Note that, unlike in prior analysis presented in Mikšovský et al. (2019), the El Niño–Southern Oscillation (ENSO) was not included among the explanatory factors due to the largely negligible influence exhibited by the available ENSO reconstructions covering our target period. We also do not present results obtained separately for instrumental and pre-instrumental periods as was done in Mikšovský et al. (2019) because such division tends to magnify uncertainties pertaining to identification of slowly variable components in climatic time series. Furthermore, outcomes for PDSI are not shown due to the long memory component in this drought index, making proper pairing of predictand and predictors problematic without additional transformations.

Figure 10Variability in annual series characterizing external forcings and large-scale internal climate oscillations involved in the attribution analysis.

Download

Figure 11Standardized regression coefficients between individual target and explanatory variables as well as their 95 % (box) and 99 % (whiskers) confidence intervals. The results shown are for time series in seasonal time steps involving all seasons (YEAR), individual seasons analysed separately (DJF, MAM, JJA, SON), and annual averages (ANNUAL).

Download

As expected, due to the generally strong relationship between greenhouse gas forcing and temperatures worldwide, there is a prominent GHGRF-correlated component in the temperature series (corresponding to an approximately 1.8 ^∘C increase between 1501 and 2020). This link is also notable in the temperature-sensitive drought indices (SPEI, Z index), most prominently during JJA and SON (Fig. 11a). In the precipitation data, the GHGRF-related trend is typically non-significant (except in MAM), and the direction of the respective link varies with season.

While there is a statistically significant association between solar irradiance and Czech climatic characteristics in a limited number of cases (particularly in SON for temperature – Fig. 11b), this relationship disappears when the slow variability component is removed from the SOLAR series (i.e. when only solar variability at periods of approximately 11 years or shorter is used as a predictor). Also considering that cross-wavelet analysis suggests only an intermittent link between temperature and SOLAR and that mutual phases of the respective oscillations are highly variable in time (Fig. 12), the direct influence of solar activity in central Europe seems unlikely from our data, at least at decadal or shorter timescales.

Figure 12Standardized cross-wavelet spectra between series of temperature, precipitation, or SPEI and explanatory variables with prominent oscillatory components (all seasons). Arrows show local phase shifts of the two series (with right-facing arrows corresponding to identical phases); areas with statistically significant oscillations are enclosed by black lines (95 % confidence level, AR(1) process null hypothesis).

Download

The signature of volcanic activity is generally weak in the precipitation data (as well as in precipitation-dominated drought indices, especially SPI – Fig. 11c). There is, however, a clear (and statistically significant) tendency for colder conditions following major volcanic eruptions, manifesting through negative regression coefficients between temperature and volcanic aerosol optical depth. This link is strongest during JJA but non-significant during DJF and MAM.

Although the NAO represents one of the major weather drivers in central Europe, its effects are highly variable both seasonally and regarding the type of target variable. For temperature, a strong tendency towards warmer conditions is associated with a positive NAO phase in all seasons except JJA (Fig. 11d). For precipitation and drought indices, the links are typically weaker, with most significant responses detected for MAM and SON. The relationship between NAO and temperature is also detectable from the cross-wavelet spectra (Fig. 12) and wavelet coherence (Fig. S2 in the Supplement), with oscillations at periods of approximately 25 and 70 years being the most prominent and relatively consistent in terms of phase difference. Similar shared periodicities can also be found in relationships between NAO and precipitation or drought indices, albeit in slightly weaker form.

As shown in Mikšovský et al. (2019), there are notable links between variations in the central European climate and decadal and multidecadal oscillations in the northern Atlantic and northern Pacific. Expanding on these prior experiments, we used the detrended common component (AMO+PDO) and difference (AMO–PDO) of temperatures in the AMO and PDO regions provided by Mann et al. (2009) as potential explanatory variables here. In the case of shared AMO and PDO variability, linear regression reveals a significant link to Czech temperature during all seasons except DJF (Fig. 11e). On the other hand, precipitation and all drought indices exhibit a relationship with differences in AMO and PDO phases, which is most pronounced during the SON season (Fig. 11f). The cross-wavelet analysis further suggests the stability of the temperature to the AMO+PDO link around the period of approximately 70 years, at least from approximately 1650 CE on (Figs. 12 and S2). Another region of spectral similarity appears around a period of 25 years, but the relationship is more intermittent and unstable in terms of phase shift. For the link between precipitation and AMO–PDO signals, the primary band of shared periodicities seems to be located between ca. 8 and 16 years, but again, some variations in phase shifts do appear.

4.4 Spatiotemporal representativeness of Czech reconstructions

To show the spatiotemporal representativeness of Czech reconstructions of selected climate variables, they were compared with related gridded reconstructions for Europe. The seasonal central European temperature series by Dobrovolný et al. (2010), compared with European temperature reconstructions by Luterbacher et al. (2004) in the 1501–2002 period, shows the highest correlation coefficients (>0.60) in the large area extending from the British Isles to eastern central Europe in the west–east direction and from southern Scandinavia to the Mediterranean in the north–south direction (Fig. 13). This area is the largest during DJF, when it extends far to eastern Europe, and the smallest in SON, when it does not cover a part of eastern central Europe (particularly Poland). In JJA, the highest correlations also extend over the whole British Isles, the northwestern part of the Iberian Peninsula, the Apennine Peninsula, and southern Scandinavia. Comparing temporal consistency between the two types of series (a series for central Europe from Luterbacher et al., 2004, and Xoplaki et al., 2005, was calculated for the window limited by geographic coordinates 45–54^∘ N and 5–23^∘ E) shows very high 31-year running correlation coefficients during the entire 500 years except a steep drop in correlations close to the significance level in JJA temperatures approximately around 1750 CE (Fig. 14a). The overall statistically significant correlation coefficients for the entire analysed period are the highest for DJF (0.94), while in the remaining seasons, they are 0.88 (MAM, JJA) and 0.89 (SON).

Figure 13Spatial correlations between reconstructed seasonal central European temperature series by Dobrovolný et al. (2010) and gridded European temperature reconstructions by Luterbacher et al. (2004) and Xoplaki et al. (2005) in the 1501–2002 CE period.

Figure 14Running 31-year correlation coefficients between the seasonal Czech climate reconstructions and selected gridded reconstructions averaged over central Europe (45–54^∘ N, 5–23^∘ E): (a) central European temperatures (Dobrovolný et al., 2010) vs. temperatures according to Luterbacher et al. (2004) and Xoplaki et al. (2005) for the 1501–2002 period; (b) Czech precipitation (Dobrovolný et al., 2015) vs. precipitation totals according to Pauling et al. (2006) for the 1501–2000 period; (c) Czech JJA scPDSI (Brázdil et al., 2016) vs. JJA scPDSI according to Cook et al. (2015) for the 1501–2012 period. Numbers in brackets represent overall correlation coefficients for the entire common period in question.

Compared to temperatures, the comparison of seasonal Czech precipitation reconstructions by Dobrovolný et al. (2015) with gridded European precipitation reconstructions by Pauling et al. (2006) for the 1501–2000 period suffers from great spatial variability of precipitation totals (Fig. 15). Although a broad belt of positive correlations extends from western to eastern Europe, the areas with highest correlations are much smaller and oriented rather to the area located westerly of the Czech territory. The Czech precipitation reconstruction is most representative in SON, while the weakest agreement appears in MAM. The 31-year running correlations between the two types of series generally decrease from the beginning of the 16th century to the middle of the first half of the 18th century (even with values below the significance level for MAM and SON), with an increasing trend afterwards (Fig. 14b). However, in addition to these trends, some remarkable drops or increases in correlation coefficients also appear (such as a drop at the beginning of the 20th century in MAM totals and an increase approximately around 1725 CE in JJA totals). The overall correlation coefficient is the highest in JJA (0.67) and the smallest in MAM (0.50), but it is statistically significant in all seasons.

Figure 15Spatial correlations between reconstructed seasonal Czech precipitation series (Dobrovolný et al., 2015) and a European gridded precipitation reconstruction (Pauling et al., 2006) for the 1501–2000 CE period.

Due to the lack of existing gridded European reconstructions of drought indices from documentary data, the JJA scPDSI series (Brázdil et al., 2016) was compared with the same European series but reconstructed from tree rings in OWDA (Cook et al., 2015) during the 1501–2012 period. As follows from Fig. 16, there is only weak spatial consistency, with larger positive correlations around the Czech territory extending to the southeast and westerly as far as France and exhibiting rather a spotty character. It is also reflected in 31-year running correlations with the series of central European windows from Cook et al. (2015), wherein the drop in correlations appears in the second half of the 16th century and particularly during the 18th century, with values far under the 0.05 significance level (Fig. 14c). This is reflected in the low overall correlation coefficient between the two series, reaching only 0.40 (but statistically significant).

Figure 16Spatial correlations between reconstructed Czech JJA scPDSI series (Brázdil et al., 2016) and the gridded European JJA scPDSI reconstruction (Cook et al., 2015) for the 1501–2012 CE period.

Because the Czech climate reconstructions are spliced from “reconstructed” and “instrumental” parts (see Sect. 2.1 for details), questions about the effects of these two parts on spatial representativeness may appear. For this reason, the temperature reconstruction was compared spatially separately for two 150-year-long periods from both mentioned parts of the series (Fig. 17). Correlations are high and significant for both parts of the series covering a large area of Europe with latitudes from ca. 60^∘ N to the south and longitudes from ca. 25^∘ E to the west. Moreover, the area of significant spatial correlations was quite similar for all seasons (not shown). On the other hand, it is necessary to say that a very preliminary version of the Czech temperature–precipitation index series compiled from a significantly lower density of documentary evidence at that time was used in corresponding gridded European reconstructions by Luterbacher et al. (2004), Xoplaki et al. (2005), and Pauling et al. (2006).

Figure 17Spatial correlations between (a) JJA reconstructed temperatures (Dobrovolný et al., 2010) and temperature field reconstruction (Luterbacher et al., 2004) for the 1600–1750 period as well as between (b) JJA measured temperatures (Dobrovolný et al., 2010) and HadCRUT5.0 temperature field (Morice et al., 2020) for the 1851–2000 period.

5.1 Climate fluctuations and European context

Proxy-based reconstructions reflect the main features of climate fluctuations. However, they can also be affected by the quality and quantity of proxies. In addition, methods of chronology compilation and data analysis may play a role. While in the case of natural proxies (e.g. tree rings) these non-climatic factors may be controlled to some extent during the process of standardization, in the case of documentary evidence it is more problematic for obvious reasons (see e.g. a detailed discussion in Brázdil et al., 2010).

With respect to these facts, mutual comparison of different climate reconstructions is an important tool to highlight strengths and weaknesses of individual reconstructions and outline possible reasons for some peculiarities in their variability. In this study, the comparison was based on the correlation analysis as well as on the direct comparison of smoothed series to highlight common variability on decadal and multidecadal scales (see Figs. 2, 8, and 14). The following text summarizes the main features of such comparison that have been explained in detail in the original “reconstruction” papers. Moreover, we are trying to explain possible factors that may be responsible for the loss of common signals in some periods.

As for temperatures reconstructed from documentary indices, very high and statistically significant correlations follow from the comparison of central European temperature series by Dobrovolný et al. (2010) with gridded multi-proxy European reconstructions of seasonal temperatures by Luterbacher et al. (2004) and Xoplaki et al. (2005), recalculated only for central European window (Fig. 14a). But around the mid-18th century a deep decline appeared in correlations for JJA temperatures, as already discussed by Dobrovolný et al. (2010). One of the reason could be the quality and quantity of available data. The reconstruction is based on documentary-derived series of temperature indices for Germany, Switzerland, and the Czech Lands. However, complete series of German indices are only available prior to 1760 and Swiss indices prior to the 1810s, while the Czech indices continued to the mid-19th century. This could result in lower temperature variability (see Fig. 14 in Dobrovolný et al., 2010) and subsequently in a lower coherence with other proxy-based reconstructions in this period.

However, a closer look at relationships between the two compared reconstructions in Fig. 14a reveals another problem. Calculation of JJA temperature differences between reconstructions by Dobrovolný et al. (2010) and Luterbacher et al. (2004) shows positive differences before the mid-18th century and negative one afterward. This shift is responsible for a sharp decrease in running correlations. In order to evaluate this inconsistency, differences of these two series with regard to a completely independent JJA multi-proxy temperature reconstruction for the Alps by Trachsel et al. (2012) were calculated. For better comparison, the series were first transformed to have a mean of zero and a standard deviation of 1. While the differences with the series by Dobrovolný et al. (2010) were distributed more or less randomly around zero, the differences with the Luterbacher et al. (2004) series showed the same patterns as described above: positive differences before the 1750s (i.e. higher temperatures by Trachsel et al., 2012) and negative differences afterward. This indicates that the problem of lost coherence around the 1750s in Fig. 14a cannot be attributed to the Dobrovolný et al. (2010) reconstruction.

As for series derived from phenological data, MAMJ temperatures reconstructed from winter wheat harvest dates were compared with 11 late spring and summer temperature series in central Europe (see Fig. 6 in Možný et al., 2012). Better coherence was found with documentary-based and biophysically based reconstructions (harvest dates) than those based on tree rings. A significant drop in correlations appeared particularly in the second half of the 17th century and around the 1750s. This may be partly related to the problem in the data quality of the winter wheat harvest dates. These dates had to be recalculated from the harvest dates of other available cereals in periods when the winter wheat dates were not available. The distinct role may be attributed to the “social bias” in data related to the complicated social and political situation in the country (see discussion related to those periods in Možný et al., 2012, and also Fig. 8a in the current study).

Similarly, AMJJ temperatures reconstructed from grape harvest dates were compared with 17 European temperature reconstructions based on temperature indices derived from documentary data, grape harvest dates, tree rings, and multi-proxies (see Fig. 9 in Možný et al., 2016a). Possible inconsistencies were found in the first half of the 16th century and around 1650, 1750, and 1900. Four periods with potential social bias were identified in the last decades of the 16th century and then in the 1640s–1670s, 1750s–1780s, and 1850s–1910s.

The comparison seems to be more problematic in the case of precipitation, which is characterized by high spatiotemporal variability. For example, less spatially homogeneous Czech JJA precipitation totals were plotted against six similar European precipitation reconstructions (see Fig. 9 in Dobrovolný et al., 2015). Periods of quite similar precipitation fluctuations were revealed, particularly in the first half of the 16th century, in the 1630s and 1710s (dry decades), and approximately in the 1590s, 1690s, 1730s, and 1810s (wet decades).

Documentary-based reconstructions of drought indices in the Czech Lands were correlated against six different European drought series (see Fig. 6 in Brázdil et al., 2016). The overall patterns were the same as in Fig. 14c in this study. While there was good agreement, especially in the first half of the 16th and the 17th centuries, a drop in common variance appeared in the second half of the 16th century, in the 1650s–1750s, and after the 1950s.

Differences between reconstructions and loss of coherence between them may also result from natural climate variability. This especially applies to those covering a slightly different spatial domain or those reconstructing climate variables characterized by high spatial variability. As discussed in more detail in Možný et al. (2016a), some periods (e.g. Late Maunder Minimum in 1675–1715 – Frenzel et al., 1994) can be characterized by a higher frequency of meteorological extremes of the regional extent. Their more frequent occurrence in some regions may be conditioned dynamically (i.e. by different circulation patterns – see e.g. Wanner et al., 1995) and thus may be responsible for higher spatial climate variability and subsequently for lower correlations in comparison to related series on a central European scale.

An interesting aspect of lost common signal manifested by a decrease in running correlations below the 0.05 significance level can also appear in the “instrumental part” of the reconstructed series, as documented in Fig. 2a. Running correlations of annual temperatures with the other five climate variables are highly significant from the 16th century up to the early 19th century. These negative correlations are physically consistent as they show that higher temperatures usually correspond to low precipitation and vice versa. Approximately from the mid-19th to the mid-20th centuries correlations among all compared series are not significant. Despite the fact that annual means express some mixture of different seasonal patterns, this gradual loss of common signal may be interpreted as follows. The fact that before the 19th century the series are reconstructed from dependent (and thus less variable) temperature and precipitation indices can be reflected in significant correlations. The instrumental parts of series (target data) are mutually less dependent and more variable than indices. The same patterns as in annual values (Fig. 2a) are also well expressed in SON series and partly in MAM and JJA series, while they do not occur in DJF series (non-significant correlations over the whole period) (not shown). The stronger common signal (significant negative correlation) occurring during the last decades can be attributed to a clearly expressed opposite tendency of rising temperatures and decreasing drought indices. The same pattern does not change even when correlating the detrended series or when changing the length of the window for which running correlations were calculated.

5.2 Climate variability and forcings

While climate reconstructions based on documentary data exhibit distinct interannual and interdecadal variability, some doubts appear regarding the expression of low-frequency (long-term) signals in such series (e.g. Brázdil et al., 2010). In our current analysis, a possible indication of different representations of long-term variability comes from the results of the wavelet transform. Although spectra of univariate documentary-based Czech series do not exhibit a clear systematic tendency towards higher amplitudes of multidecadal oscillations in any specific subperiod (Fig. 9), diminished powers of shared oscillations in cross-wavelet spectra appear for some of the explanatory variables, particularly around the 70–100-year period band (Fig. 12). On the other hand, such behaviour may be related to specific features of the explanatory variables themselves, particularly lower variances displayed by the NAO and AMO+PDO series in the early parts of the 1501–2020 period (Fig. 10). The phenological data provide a somewhat different representation of long-term oscillations in the temperature and drought index series, with a notable contrast between the early and later parts of the analysis period and peculiar differences between cereal- and grape-based reconstructions (Fig. 9). While this heterogeneity may be partly climatic in origin or related to crop-specific responses to particular weather patterns, variations in the geographical structure of growing locations or changes in cultivars grown (Možný et al., 2012, 2016a, b) likely play a considerable role as well. Even so, the presence of distinct spectral similarities between index- and phenology-based reconstructions supports the existence of ca. 70–100-year oscillations affecting the Czech climate, despite discrepancies in their exact timing and amplitude.

The problem of potential misrepresentation of low-frequency variations particularly concerns the expression of temperature–precipitation patterns in the form of different ordinary degree scales used for the creation of a series of temperature–precipitation indices that are less sensitive to characterizing particularly extreme values. It is well expressed in long-term trends of the analysed 520-year series, wherein no statistically significant trends appear in seasonal and annual precipitation (see Brázdil et al., 2021, from 1961 CE), and in drought index series. Significant trends in temperature reconstructions are mainly the effect of sudden temperature increase from the 1970s (see Zahradníček et al., 2021, from 1961 CE). It appears not only in the reconstruction based on temperature indices (Dobrovolný et al., 2010) but also in those derived from phenological data (Možný et al., 2012, 2016a). On the other hand, in reconstructions based on phenological data, non-homogeneities due to social bias may appear. For example, Možný et al. (2012) considered this aspect in connection with the importantly warmer first half of the 16th century (for example, the use of the sickle for cutting requested more time, i.e. harvests started earlier) and the importantly cooler second half of the 17th century (total devastation of the Czech Lands after the Thirty-Year War, coinciding with the cold Maunder Minimum period – see e.g. Frenzel et al., 1994) in MAMJ reconstruction from winter wheat harvest dates. The earliest start of harvests in 1517–1542 CE, even comparable to 1971–2010, was confirmed by Brázdil et al. (2019a) by analysing long-term changes in the agricultural cycle in the Czech Lands.

The presence of linear trends, especially detected for the temperature series and its drought-related derivates, can be approximated very well by the variations in greenhouse gas concentrations and the resulting changes in radiative forcing. Despite this good formal match, note that statistical methods alone are unable to reliably confirm the causal nature of this relationship between long-term trends, and other approaches (such as simulations by dynamical models) are needed to verify causality.

Similar caution is needed in the case of solar activity: while the regression analysis suggested the possibility of a relationship to Czech temperatures, this link vanished after the slowly variable component was removed from the SOLAR series (which eliminated aliasing between GHGRF and SOLAR signals). It should be emphasized, however, that our (strictly linear) analysis does not exclude the possibility of more complex interactions between central European climate and solar activity, possibly detectable by more general methods.

Unlike changes in solar activity, volcanic activity leaves a distinct imprint in the reconstructed temperature series. Cooling following major volcanic material ejections into the stratosphere is most notable during JJA; on the other hand, it is only borderline statistically significant in the temperature-sensitive drought indices (especially SPEI) and not detectable from the precipitation series.

Unsurprisingly, a strong effect of NAO was detected in most of the Czech series analysed, but the strength of its impact varied seasonally (with JJA exhibiting the weakest connection). Prominent components of this relationship seem to be tied to periodicities of approximately 70 and 25 years, although the respective links are not completely stable in time.

Our analysis, involving temperature variability in the AMO and PDO regions as an explanatory factor, has confirmed the distinct influence of both shared AMO and PDO variability (especially identified in the Czech temperature series) and their difference (significantly influencing Czech precipitation and drought indices). The results of cross-wavelet analysis suggest that this AMO–PDO impact may be related to shared periodic oscillations in the ca. 70- to 100-year period band. Other spectral similarities (although manifested in a less coherent fashion) have also been detected over the approximately 16- to 32-year period band (especially for the common AMO+PDO variability) and 8- to 16-year band (for the AMO–PDO difference). However, substantial variance in mutual phases revealed by the wavelet spectra suggests that the nature of these potential links goes beyond simple linear responses, and a more complex analytical approach may be needed to fully unravel them.

Finally, note that the outcomes of the attribution analysis may also be subject to specific properties of the explanatory variables used, particularly in the case of the reconstructed indices of internal climate variability modes (NAO, AMO, PDO). This issue has been previously investigated in Mikšovský et al. (2019), wherein multiple independent reconstructions were used for each of these indices (including NAO reconstructions by Trouet et al., 2009, and Ortega et al., 2015, AMO reconstruction by Gray et al., 2004, and PDO reconstructions by MacDonald and Case, 2005, and Shen et al., 2006). The series based on data by Luterbacher et al. (2001) and Mann et al. (2009) were shown to carry the relatively strongest link to central European climate variability and were therefore employed in this current analysis. Even so, the problem of predictor-related robustness of the attribution analysis remains an essential one and an issue worthy of revisiting, especially when new relevant proxy-based data arrive in the future.