Climate change fingerprints in recent European plant phenology

A paper published in Global Change Biology in 2006 revealed that phenological responses in 1971–2000 matched the warming pattern in Europe, but a lack of chilling and adaptation in farming may have reversed these findings. Therefore, for 1951–2018 in a corresponding data set, we determined changes as linear trends and analysed their variation by plant traits/groups, across season and time as well as their attribution to warming following IPCC methodology. Although spring and summer phases in wild plants advanced less (maximum advances in 1978–2007), more (~90%) and more significant (~60%) negative trends were present, being stronger in early spring, at higher elevations, but smaller for nonwoody insect‐pollinated species. These trends were strongly attributable to winter and spring warming. Findings for crop spring phases were similar, but were less pronounced. There were clearer and attributable signs for a delayed senescence in response to winter and spring warming. These changes resulted in a longer growing season, but a constant generative period in wild plants and a shortened one in agricultural crops. Phenology determined by farmers’ decisions differed noticeably from the purely climatic driven phases with smaller percentages of advancing (~75%) trends, but farmers’ spring activities were the only group with reinforced advancement, suggesting adaptation. Trends in farmers’ spring and summer activities were very likely/likely associated with the warming pattern. In contrast, the advance in autumn farming phases was significantly associated with below average summer warming. Thus, under ongoing climate change with decreased chilling the advancing phenology in spring and summer is still attributable to warming; even the farmers’ activities in these seasons mirror, to a lesser extent, the warming. Our findings point to adaptation to climate change in agriculture and reveal diverse implications for terrestrial ecosystems; the strong attribution supports the necessary mediation of warming impacts to the general public.


| INTRODUC TI ON
Although plant phenology is one of the oldest forms of environmental monitoring, with phenological observations taken by ancient civilizations (Koch et al., 2007;Schwartz, 2003), it was only in the 1990s that its renaissance started with key publications on detection of climate change impacts on global vegetation. Keeling, Chin, and Whorf (1996) were the first to report a 7 day earlier start of the growing season based on long-term measurements of atmospheric CO 2 concentration, which Myneni, Keeling, Tucker, Asrar, and Nemani (1997) confirmed using normalized difference vegetation index satellite data from the Northern Hemisphere. Menzel and Fabian (1999) then provided the necessary ground truth by their analyses of long-term European phenological data. Many other papers, summarized in wellcited reviews and relevant chapters in IPCC reports (e.g. Rosenzweig et al., 2007;Walther et al., 2002), confirmed the extraordinary role of phenology as a bio-indicator of climate change. While research initially focused purely on the identification of changes, it then extended to communication of climate change to the general public (e.g. Van Vliet et al., 2003) as well as on various ecological consequences of these changes (Morisette et al., 2009;Thackeray et al., 2010). Nevertheless, at the beginning of this renewed research there was a latent accusation of a publication bias, or cherry picking, in the sense that only the most advancing records or extraordinary changes found their way into popular scientific journals. Therefore, the COST725 initiative collected all available European phenological data (later developed as the PEP725 database, Templ et al., 2018) and analysed more than 100,000 time series for climate changedriven changes (Menzel, Sparks, Estrella, Koch, et al., 2006, hereaf-ter referred as GCB2006). The GCB2006 study concluded that there was indeed a strong response in European phenology to climate change and that these changes matched the warming pattern. With more than 1,500 citations in the Web of Science Core Collection, this study has a high relevance for the scientific discourse on climate change impacts on the biosphere. It was also the backbone of the corresponding assessment of observed changes and responses in natural and managed systems of AR4 WGII of the IPCC (Rosenzweig et al., 2007) as well as of the subsequent paper of attribution of global impacts in nature to anthropogenic warming (Rosenzweig et al., 2008).
Thus, the well-accepted role of phenology as a climate change indicator is based on the formal attribution of shifted phenological onset dates to anthropogenic-induced warming (Rosenzweig et al., 2007(Rosenzweig et al., , 2008, demonstrating at the same time that the phenological trends are not just natural variability, for example, as shown by Guan (2014) for two phenophases which unequivocally corresponded to the respective winter/spring warming trends. Most studies, however, rely on proof of the sensitivity of phenological phases to temperature as the fundamental prerequisite of climate change detection and attribution, and have linked phenological changes to temperature increases (e.g. Cook et al., 2012;Thackeray et al., 2016). Since there is thorough evidence that spring phenological development is also partially triggered by fulfilment of winter chilling and by photoperiod (e.g. Laube et al., 2014;Tang et al., 2016;Vitasse & Basler, 2013), it is debated whether phenological changes still mirror the recent ongoing warming, or mirror it to the same extent as before. Therefore, an update of the assessment of phenological changes across Europe and their attribution to climate change is necessary. In contrast to the GCB2006 publication, this update will uniquely concentrate on plant phenology.
Major factors that may impact on the results of this update are different changes in (apparent) temperature sensitivity (e.g. . Species-specific increases in heat requirements related to a decrease in chilling have been especially identified as the main reason . However, there are also contrasting opinions for example, by Wang et al. (2017). Güsewell, Furrer, Gehrig, and Pietragalla (2017) showed that for Switzerland there are still no indications of a lack of chilling leading to altered temperature sensitivities. Changes in temperature sensitivities are reported as not uniform over time  and, in particular, a weakened temperature response was observed since 2000 (Fu et al., 2014). Spatial variations in temperature sensitivity have been linked to mean annual temperature and to seasonal temperature range (Lapenis, Henry, Vuille, & Mower, 2014;Menzel, Sparks, Estrella, & Roy, 2006;Wang et al., 2014). Both major backup systems to prevent a premature spring development (chilling, photoperiod) may finally lead to nonlinear responses of phenological onset dates to forcing temperatures (Jochner, Sparks, Laube, & Menzel, 2016). Temperature sensitivity has also been shown to vary with specific plant traits, such as evolutionary relatedness, and to support invasions and plant performance (e.g. Wolkovich, Cook, & Davies, 2014). For wheat and maize in agriculture there are hints of geographic differences among cultivars in sensitivity to vernalization, day length and temperature (Van Bussel, Stehfest, Siebert, Müller, & Ewert, 2015). However, there are still major knowledge gaps concerning the so-called 'false' phenological phases, such as sowing or harvesting in agriculture, whose timings are decided by farmers. The GCB2006 study and follow-up publications provided the first evidence that the phenological signal in agriculture was weaker than the signal for wild growing plant species (Bock, Sparks, Estrella, & Menzel, 2013;Estrella, Sparks, & Menzel, 2007;Menzel, Sparks, Estrella, Koch, et al., 2006;Menzel, Vopelius, Estrella, Schleip, & Dose, 2006). However, it can be anticipated that, sooner or later, agricultural management and/or choice of cultivars will also be adapted to the new potential growing seasons. The drivers of autumn phenology, such as leaf colouring and leaf fall, are still far from being completely understood Gallinat, Primack, & Wagner, 2016;Gill et al., 2015;Wu et al., 2018).
The aim of the current paper is to update the previous study on climate change fingerprints in European phenology and include as many plant species/phenophases as are available. Recent publications have exclusively focused on a few prominent spring leaf unfolding phases and results of those may be strongly related to the choice of analysed species, having different sensitivities and covering different seasons.
Phenology is controlled by seasonal patterns in the warming signal (Lapenis et al., 2014) and these will vary by location (spatial differences in the climate signal) and study period (temporal differences in the climate signal, e.g. Rutishauser et al., 2009). Consequently, this update is more than timely since the climate signal has not been stable, but temporal variations or trends in climate change have been reported, especially related to the so called hiatus or standstill period (Zang, Jochner-Oette, Cortés, Rammig, & Menzel, 2019). A publication has already indicated that there were no trends in spring and autumn phenology during this warming hiatus period (Wang et al., 2019). Another phenological regime shift in the mid-1980s was also related to discontinuous temperature changes (Reid et al., 2016).
Thus, this update study will focus on the following research questions by analysing the complete picture of plant phenological changes in Europe: 1. We expect that there is an inherent variation of trends and will thus concentrate on how do phenological onset dates/trends vary overtime and by season and what drives the strength of recent trends.
2. An important result of the GCB2006 paper was that farmers' activities exhibited weaker trends and that there was an unclear change pattern in autumn. We will therefore study recent differences in climate change signals among wild plants, fruit trees, agricultural crops and farmers' activities and expect 'false' farming and agricultural crop phenology to be more similar to wild species and fruit trees now, but leaf colouring and fall still unchanged.
3. Addressing the discussion on lack of chilling, altered warming patterns and differential responses in farming we ask whether, and for which phenological groups, there is still an attributable fingerprint of climate change in phenology.

| Phenological data
Complete original plant phenological observation data were retrieved individually from the European Meteorological Services of Germany (DWD), Austria (ZAMG) and Switzerland (MeteoSwiss). Data of these countries account for 96.3% of the PEP database (Templ et al., 2018), thus our results are comparable to any other based on PEP725 data. However, these national phenological databases are richer in sites, species and phenophases; therefore, although having a smaller spatial extent, are preferred for this study (see map in Figure S1a). Unfortunately, it was not possible to directly update the GCB2006 data set due to a lack of access to the recent phenological data of specific countries. However, data from Germany, Austria and Switzerland constituted around 96.7% of the GCB2006 data set.
Thus, it is justified to consider the new data set as comparable.
Out of the complete phenological data from these countries, we took observational data between 1951 and 2018 with time series (series per species/phase/station) longer than 29 years and ending in or after 2000. Duplicates were removed by averaging the respective DOYs (onset dates as day of the year) in the same year. Very early phenological events (e.g. hazel or snowdrop flowering in December) were allocated to the correct reference year by negative DOYs.
This data set, hereafter called Update, comprised more than 4.2 million observational records and almost 97,000 time series (Table 1).
It is important to notice that a direct comparison between GCB2006 and the updated results is not straightforward since some differences in phenophases and data selection procedure exist.
Thus a GCB2006s (s = simulated) data set was defined as a subset of Update containing identical species/phases/stations, but restricted to the GCB2006 period and series length (1971-2000, 15+ years).
This simulation of GCB2006 conditions in Update allowed studying the effects of ongoing climate change post-2000 as well as methodological aspects (15+ vs. 30+ year series), when comparing Update and GCB2006s results. In contrast, GCB2006 results are only reported and discussed with respect to possible effects due to some different sites/phenophases in GCB2006s.
In the previous GCB2006 study comprising ~100,000 series in the period 1971-2000 plant phenological series were categorized into four groups (see Table 2). For Update and GCB2006s, we further refined this categorization into nine clusters, taking into account BBCH coding of phases (Meier, 1997) as well as farming activities, and additionally defined four periods/seasons, different from those in GCB2006.

| Climatic and other auxiliary data
The E-OBS v19.0HOM gridded data set with a 0.25 degree regular grid (Cornes, Schrier, Besselaar, & Jones, 2018), which is the homogenized

| Analyses of trends
Phenological changes were determined as linear regressions of onset days (DOY) against year for all series (i.e. 30+ years in Update). For GCB2006s , similar to GCB2006, all 15+ year series were considered. p values of the linear regression slopes were adjusted for multiple comparisons using the false-discovery-rate (FDR; Benjamini & Hochberg, 1995). Proportions of (significant) negative/(significant) positive trends and mean slopes (Tr mean ) were determined for the different categories described in Table 2. The uncertainty range of Tr mean is the 95% confidence interval for estimating the mean (of the slopes), and not the 95% interval of the underlying data. Thus, the range is TA B L E 2 Categorization of phenophases in Update comprising nine clusters and four phenological periods/seasons as compared to GCB2006

| Trend modelling
In order to understand what the partial contribution of explanatory factors to the observed phenological trend was, we followed a modelling approach. Explanatory factors/variables were the nine clusters (cluster9, see Table 2) The effect of each variable is calculated by holding all other variables constant (except for interactions). The models were fitted in R statistical software version 3.6 using the mgcv-package (Wood, 2017), effects were calculated using the emmeans-package (Lenth, 2019), and results were visualized using the ggplot2-package (Wickham, 2016).

| Attribution of the phenological change pattern to temperature changes
Attribution followed the methods proposed by Rosenzweig et al. (2007Rosenzweig et al. ( , 2008. First, based on statistically significant trends only, the percentage of significant trends (out of all significant trends) matching the direction expected from climate warming was determined, both for the phenological groups of GCB2006 and of Update.
Second, we applied a spatial approach based on all sites for which we tested the effect of the corresponding temperature trends on all phenological trends. For this we divided all (significant as well as non-significant) temperature trends in terciles (below, average, above) on a seasonal basis and the phenological trends by sign and significance (negative significant, negative nonsignificant, positive nonsignificant, positive significant). A frequency analysis was performed on this cross tabulation using Chi-square tests and p-values were adjusted for multiple comparisons using the FDR. Pearson (standardized) residuals for each cell were determined to assess their relative contribution to the total Chi-square score.
All calculations were done in R statistical software (R Core Team, 2019).

| Phenological changes
For leaf unfolding and flowering, fruiting and farmers' activities, the percentage of negative trends slightly increased, for example from 87% (GCB2006s) to 89% advanced for leaf unfolding and flowering (

Phenological trend
= cluster9 + woodiness + pollination mode + longitude + latitude +elevation + s (start year) + s (number years) + woodiness * pollination mode + cluster9 * longitude + cluster9 * latitude +cluster9 * elevation, For the nine clusters, trends in terms of sign, proportions of significant trends and mean slope are summarized in Figure 1 (corresponding numbers are given in Table S1a,b). Except for FWv au which comprises leaf colouring and fall in fruit trees and wild species in autumn, the overall change pattern in all other clusters was advancing. All spring and summer clusters had higher percentages of negative and significant negative trends in Update than in GCB2006s (Figure 1a) (Figure 1c; Table S1a).
The percentage of delayed leaf colouring and leaf fall increased to 57% (26% significant) compared to GCB2006s (49% and 11%, respectively). The mean slope was +0.036 ± 0.007 days/year versus −0.015 ± 0.013 days/year in GCB2006s, thus more clearly indicating a later ending of the growing season.
Undoubtedly ~3/4 of the farmers' spring and summer activities (F sp sowing of spring cereals, F su harvest) were advancing (74% and 84%, respectively) and the percentages of significant advances clearly increased from GCB2006s to Update (14%-31% and 24%-53%, respectively). For farmers' sowing of autumn crops (F au ), more advanced than delayed series were observed; however, the picture was still similar to GCB2006s except that more trends (both advances and delays) were significant ( Figure 1a; Table S1a). Surprisingly, mean slopes became less negative, except for F sp which advanced more strongly in Update than in GCB2006s (−0.116 ± 0.005 vs. −0.075 ± 0.011 days/year).
About 84% of the series indicated a lengthening of the growing season (GS, from leaf unfolding to colouring) and 48% were significantly longer (Figure 1b) Table S1b). The average trend was almost zero with roughly equal numbers of the series showing a shortening or lengthening of the generative period.   year of 1989, slopes for leaf colouring and leaf fall (b3 in Figure 2a,

| Variation of phenological changes and decadal anomalies across seasons
Plotting mean trends over weeks of the year (Figure 3) clearly exhibits systematic variations with season. Only weeks covered by less data points at the beginning or end of record had wider confidence intervals of the mean trends. In autumn, trends of delayed leaf colouring and leaf fall (b3 in Figure 3a, FWv au in Figure 3b Farmers' activities in spring and summer (F sp , F su ) only exhibited earlier than average starting dates in the last and the two last decades respectively.

| Modelling of the slopes of the phenological trends
Since phenological trends varied with phenological group (Table 3) and cluster (Figure 1), year (Figure 2), season ( Figure 3) and length of the series (Table 3), these variables, as well as geographical coordinates and species traits, were used in the generalized addi-

| Attribution of phenological changes to warming
About 96% and 95% of the significant changes of leaf unfolding and flowering (b1) and fruit ripening (b2), respectively, were negative, thus indicating advancing onset dates with warming (see Table 3).
For farmers' activities, 83% of the significant trends were negative.
Significant leaf colouring and leaf fall trends (b3 in Table 3, FWv au in Table S1a) were 63% positive, hinting to delayed autumn with warming. This pattern is confirmed by the corresponding analysis for the spring and summer clusters (data from Concerning the spatial match of temperature trends to all phenological trends, results of the Chi-square tests indicated that for all phenological clusters, except crop vegetative development in spring (Cv sp ) as well as ripening phases of fruit trees and wild plant species in summer (FWg su ), there was at least one significant association of phenological trends with seasonal warming patterns in winter/spring/summer ( Figure 5; Figure S4).
Crop generative development in spring (Cg sp ) was significantly associated with winter and spring warming patterns, and the respective spring phases in fruit trees and wild plant species (FWv sp , FWg sp ) even displayed a stronger positive association (see Figure S4). For the latter two, counterintuitively average and above average summer temperature trends were associated with delayed (subsequent) spring phenophases. Above average warming in winter and spring was significantly associated with delayed leaf colouring and leaf fall (FWv au ). For farmers' activities, the Chi-square tests indicated that advancing spring activities (F sp ) were significantly linked with above average warming in winter, but below average warming in summer. For farmers' summer activities, above average warming in summer and autumn was associated with advancing summer trends (F su ). In contrast, below average warming in summer was connected to advancing autumn activities (F au ).  of trends analyses may also partly be driven by the set of species and phenophases included. In our case a considerable adjustment in the observational programme of the DWD in 1991 has led to a substantial reduction in the number of (agricultural) species/ phases when series were selected to end in 2000 or later. ZAMG data selection mirrors a similar decrease of observations in the last two decades (see Figure S1b).

| D ISCUSS I ON
In the following the results will be discussed with respect to the three guiding research questions. There is still a clear picture of phenological advance except for autumn. For the vegetative and generative phases of crops, fruit trees and wild plants, longer time series (30+ years in Update) led to ≥90% advancing trends in spring (Cg sp , FWv sp , FWg sp ), ≥81% for Cv sp and FWg su , and ~75% for the farmers' activities in these seasons (F sp , F su ). Accordingly, at least 30+ year series are needed for robust trend estimations (Dose & Menzel, 2004;Rosenzweig et al., 2007) as also confirmed by our trend modelling.
Longer series were also linked to a higher percentage of significant trends, but assuming comparable variability and change rates, greater statistical power would be expected to lead to an increase in significance.
Although the proportion of trends that was significant in- it is more likely that a reduction in forcing conditions has driven the decrease in the advance of spring and summer phenology (as reported by Güsewell et al., 2017 for Switzerland) than a lack of chilling (e.g. . Nevertheless, there are differences in this decreased sensitivity of warming with leaf unfolding and flowering of fruit trees and wild plant species exhibiting the strongest decline in trend strength. In autumn, leaf colouring and fall trends now predominantly (57%) indicated delayed onset dates with a mean positive trend.
However since 1951-1960 was characterized by earlier mean onset dates, it has to be checked whether a prolongation of the records into the past (1951)(1952)(1953)(1954)(1955)(1956)(1957)(1958)(1959)(1960)(1961)(1962)(1963)(1964)(1965)(1966)(1967)(1968)(1969)(1970)  There was still a lengthening of the growing season of ~0.26 days/ year and this lengthening only marginally decreased in the most recent decade confirming many other studies (e.g. Kolářová, Nekovář, & Adamík, 2014). In this respect our paper disagrees with the findings F I G U R E 5 Attribution of phenological trends for the nine clusters (see Table 2) to trends in seasonal temperatures by Chisquared test (absolute numbers are given in Figure S4). Pearson residuals indicate the relative contribution of a cell to the total Chisquare. Crossed out combinations were not significant (p-values adjusted for multiple comparisons of 4 * 9 = 36 tests by FDR). The size of the circle is proportional to the amount of the contribution, green indicates positive residuals, which specify a positive association between phenological and temperature trends, and light grey implies a repulsion or negative association [Colour figure can be viewed at wileyonlinelibrary.com] of Chen et al. (2019), likely because species other than trees were also incorporated in our spring signal. In contrast, the FS was shortened by −0.15 days/year due to a smaller advancing trend in farmers' spring activities than in the development of wild plant species, matching the previous findings of Estrella et al. (2007). In principle, farmers themselves should profit from this earlier start of the (abiotic) growing season, also by using more cold tolerant cultivars of maize and summer cereals, but according to Parker, Shonkwiler, and Aurbacher (2017)  of fruit trees and wild species. Nonwoody and insect-pollinated plant species advanced less than wind-pollinated species, which was also found in the western Mediterranean (Gordo & Sanz, 2009). Advancing trends reached their maximum rates when starting at ~1978, a finding which corresponds to the reported 1980s regime shift (Reid et al., 2016).
Although the change pattern varied over time, the (still) advancing trends could be attributed to warming. More specifically, based on percentages of significant trends matching the warming, advancing farming activities are likely, ripening phases in summer are very likely and phenological spring phases, such as leaf unfolding and flowering, are very or extremely likely to mirror the increasing temperatures. The spatial approach confirmed this attribution, however, to a lesser degree. Seasonal warming was significantly associated with spring phases' advance in fruit trees and wild plant species, generative phases in crops as well as farmers' activities mirroring winter and spring warming (descending order, F sp not with spring warming).
Advancing farmers' summer activities clearly mirrored summer and autumn warming patterns. The association in autumn was reversed with below average warming in winter and spring being linked to advanced leaf colouring and fall (see equally Zohner & Renner, 2019) and less warming in summer being linked to advanced farmers' autumn activities. Most interestingly, what has been described as carry-over effects (Sparks, Buras, Estrella, & Menzel, 2020;Zohner & Renner, 2019), is depicted by this attribution analysis, where below average warming in summer was linked to advancing phenological trends in spring (FWv sp , FWg sp , F sp ). Thus, formal attribution of phenological trends does not only shape a formal fingerprint in nature, but additionally fosters a deeper understanding of the drivers.
The results of the updated study are relevant in a number of ways. A stronger advance of early spring wind-pollinated species leads to an earlier start of the allergenic pollen season. Citizen scientists can still observe climate change in their backyard and these spring advances including a lengthening of the growing season are to a large extent attributable to warming, although our refined analysis on nine clusters showed the superiority of fruit trees and wild species in this respect. A tricky question to be investigated in the future is why summer ripening phases of fruit trees and wild species do not mirror any warming pattern, whereas summer harvesting in agriculture does.
Our results clearly underline that farmers' decisions (weaker trends and smaller trend changes in spring and autumn) may be driven by other factors as well, although farmers seem to respond/ adapt since their activities in spring were the only phase to exhibit stronger advancing trends in Update. A shortening of the crop generative period may have undesired consequences for yield, but will allow more intercropping or earlier sowing of winter cereals.
Results of our comprehensive analysis of the complete plant phenological data set in these three Central European countries underline that the ecological consequences of these changes are challenging to be assessed due to inherent variation of changes over time, season, with topography and plant traits, but this variation does not hamper climate change attribution.

ACK N OWLED G EM ENTS
We acknowledge the E-OBS data set from the EU-