Human influence increases the likelihood of extremely early cherry tree flowering in Kyoto

In addition to the long FFD record we also use temperature observations at Kyoto's Meteorological Observatory during the instrumental period 1881–2021. Details on the FFD and temperature data are given in supplementary appendix 1 (available online at stacks.iop.org/ERL/17/054051/mmedia). Kyoto's FFDs are highly anti-correlated with the mean temperature in March, as illustrated in figure 1(b) (correlation coefficient of −0.85). A warming trend since the mid-20th century maps onto cherry tree phenology, as flowering times arrive earlier. Attribution research has demonstrated the significant role of anthropogenic forcings in rising mean and extreme temperatures in East Asia [25–28], primarily driven by greenhouse gas emissions. In Kyoto, in addition to greenhouse gases, the effect of urbanisation is also expected to be an important contributor to the warming [29].

To illustrate the urban effect (figure 2(a)) we compare Kyoto's temperature timeseries with those from the rural station of Kameoka in the north-western suburbs of Kyoto. It is evident that the originally almost identical timeseries begin to diverge after the 1940s when the city's urban expansion took off. The temperature difference between the two locations (figure 2(b)) shows that the urban bias continues to increase until about the end of the 20th century, at which point Kyoto's observation station appears to be surrounded by a predominantly urbanised environment. The temperature difference between the two stations smoothed with 30 year running means (red line in figure 2(b)) to minimise the effect of internal variability is used for the urban-correction of simulated temperatures, as we will explain later. The difference is also extended to future years, assuming the urban warming has already reached its peak. It should be noted that the urban warming plotted in figure 2(b) corresponds to the mean temperature from February to April, as the phenological model uses daily temperature data from mainly these three months. However, we find that the urban effect does not notably vary between individual months, as illustrated, for example, in supplementary figure 1 for the month of March.

Zoom In Zoom Out Reset image size

We conduct the attribution analysis with a suite of 14 models which contributed to the Coupled Model Intercomparison Project Phase 6 (CMIP6) Project [30] (supplementary table 1). We select all models that provide daily and monthly mean temperature data for two experiments, one with and one without human influence on the climate [31]. The former experiment (ALL) includes all external climatic forcings, i.e. changes in well-mixed greenhouse gases, aerosols, ozone and land use, as well as natural forcings that represent changes in volcanic aerosols and the solar irradiance. The second experiment (NAT) describes a hypothetical world without human influence on the climate and so includes natural forcings only. The models provide multiple ALL and NAT simulations and, to avoid giving high weight to models with larger ensembles, we limit the number of simulations used in the study to a maximum of 10 per model for each experiment. In summary, the daily data are derived from 51 ALL and 55 NAT simulations, while larger ensembles are available for monthly data (supplementary table 1). Both experiments start in year 1850, and we also consider future changes by extending the ALL simulations to the end of the 21st century with the 'middle of the road' Shared Socioeconomic Pathway 2 4.5 (SSP2 4.5) [32], a scenario that describes a plausible pathway given the current mitigation efforts [33].

For each model we extract temperature data at the grid-point located closest to Kyoto (35N, 135.7E). Although temperature spatial correlations are large enough to justify this approximation, the models have a relatively coarse horizontal resolution, of the order of a few hundred kilometres, which represents the wider rural area around the city rather than the urban environment where the FFD and temperature records are collected. Indeed, we find that the ALL simulations yield trends lower than the observed trend in Kyoto, but consistent with the lower trend in Kameoka (figure 2(c)). To account for the urban effect, we add the observationally estimated urban warming (red line in figure 2(b)) to temperature data from the ALL experiment. The urban correction is not applied to the NAT temperatures, as the natural climate is not influenced by any human forcings. Modelled temperatures also have mean biases relative to the observations, which are typically corrected by ensuring that the mean simulated temperature over a baseline period agrees with the observations[34]. Here we use the first 50 years of the observational period as a baseline (1881–1930), as it is long enough to provide a reliable bias estimate, but also early enough not to be considerably influenced by anthropogenic warming. We estimate the mean February to April temperature over the baseline years from the ALL simulations of each model and subtract it from the equivalent observed mean. We then remove this estimated bias from both the ALL and NAT simulations of the same model and repeat the procedure for all the models. The mean correction factor (supplementary table 2) is again found not to vary appreciably between months and, assuming it is constant, we also apply it to simulated daily temperatures between February and April. The ALL trends from bias-corrected data that also include urban warming are consistent with the observations (figure 2(d)) and the ensemble mean trend agrees well with the observed trend in Kyoto after correction.

As attribution results are derived from model experiments, it is essential to demonstrate that our multi-model ensemble can reliably represent the main characteristics of Kyoto's March temperatures, which are closely linked to the FFDs. Apart from historical ALL trends, already shown to be consistent with the observations, we also employ here some additional standard evaluation assessments [34] to examine whether the modelled variability and temperature distribution compare well with the observations. March temperature timeseries are illustrated in figure 3(a). The ALL simulations reproduce the observed long term warming, projected to continue in future decades. On the other hand, the NAT climate appears to be largely stationary, suggesting that Kyoto's warming trend is of anthropogenic origin. Variability over different timescales is commonly assessed with power spectra [34, 35]. Spectra from detrended temperature timeseries show that the observations lie within the range of the models (figure 3(b)). Finally, using the quantile–quantile plot (figure 3(c)) we affirm that the temperature distribution constructed with the ALL ensemble (both main body and tails) is in overall agreement with the observations. We therefore conclude that the CMIP6 data employed here are fit for the purposes of the attribution analysis. Evaluation assessments applied to the data without the urban effect correction also confirm that the models represent well the rural temperatures of Kameoka (supplementary figure 2).

Zoom In Zoom Out Reset image size

Phenological modelling of blooming dates ranges from simple regression methods to more sophisticated physical-based models like the growing degree days phenological models [36–39]. The latter represent the transition from a rest period to a forcing period when heat units are accumulated until a predefined heating requirement is met. Here we employ the Days Transformed to Standard temperature (DTS) model [40–42], an alternative accumulation model that proposes an exponential contribution of the daily temperature on growth during the development process. The DTS value for day j in year i, DTS_ij ^day, corresponds to the amount of daily growth at temperature T_ij relative to the to the growth at a standard temperature T_s and is approximated by the relationship:

$$\begin{align}{\text{DTS}}_{ij}^{{\text{day}}} = \exp \left[ {\frac{{{E_\alpha }{\text{(}}{T_{ij}} - {T_s}{\text{)}}}}{{R\, \cdot \,{T_{ij}}\, \cdot \,{T_s}}}} \right].\end{align} \tag{ 1 }$$

The parameter E_α (J mol⁻¹) describes the response of flower buds to temperature, and R is the universal gas constant (8.314 J mol⁻¹ K⁻¹). The FFD is estimated as the day when the accumulation of daily DTS values, after a starting date D, reaches the mean accumulation value over a calibration period with N years:

$$\begin{align}{\text{DT}}{{\text{S}}^{{\text{calib}}}} = \frac{1}{N}\mathop \sum \limits_{i = 1}^N \mathop \sum \limits_{j = D}^{{B_i}} {\text{DTS}}_{ij}^{{\text{day}}}\end{align} \tag{ 2 }$$

where B_i denotes the FFD in year i. The parameters of the DTS model for Kyoto's cherry trees were estimated in previous work by an error analysis [41] for the 30 year calibration period 1911–1940 (N= 30), when the influence of urban and other anthropogenic warming is small. The parameter values, also used in this study, are E_α = 56 kJ mol⁻¹ and D = 42 (11 February), while the mean DTS over the calibration period is DTS^calib = 32.75. Applying the DTS model to daily temperature data from the ALL and NAT simulations we calculate FFD values for all the simulated years.

Using the FFD data derived from the ALL and NAT experiments, we first compute the change in the FFD relative to the natural climate in a reference time-window, e.g. a decade. To estimate the FFD change we use the 51 × 10 annual estimates of the FFD extracted from the 51 ALL simulations for the reference decade. Assuming the NAT climate is largely stationary, the mean FFD without anthropogenic climate change is approximated simply by the mean of the 9405 estimates derived from the NAT simulations (171 years during the period 1850–2020 × 55 simulations). We then subtract the mean NAT FFD from each of the 510 ALL estimates and compute the anthropogenic signal as the mean of the resulting 510 values. The associated uncertainty, representing year-to-year departures from the mean anthropogenic change because of internal climatic variations, is represented by the 5th and 95th percentiles of a normal distribution fitted to the same 510 values. The procedure is also applied to data from the ALL experiment but without the urban correction, to examine the what the anthropogenic influence would be in a rural environment.

In the second part of our attribution analysis, we estimate the probability of extreme events (FFD = 85 or less). The natural world probability is estimated by fitting a Generalised Pareto distribution (GPD) to the 9405 FFD estimates derived from the NAT simulations. For the ALL climate, we compute the probabilities in 20 year rolling windows, from the present climate to the end of the century. In this case we use the ALL FFDs in time windows 2012–2031, 2013–2023, ..., 2081–2100, i.e. samples of 1020 values per time window. The ALL probabilities are also estimated for rural conditions in the same way from data that do not include the urban correction. Rural probabilities are also computed with the GPD, whereas the larger probabilities with the urban effect included are estimated with the normal distribution [43]. Uncertainties from sampling limitations are estimated with a Monte Carlo bootstrap procedure [43] that performs random resampling of the FFD estimates and recalculates the probability with the resulting alternative samples. The procedure is repeated 1000 times and the 5%–95% uncertainty range is then computed from the 1000 probability estimates.

Reconstructions of Kyoto's FFDs with the DTS model using daily temperature observations as input are in close agreement with the observed FFDs (figure 4(a)). The correlation coefficient between the reconstructed and observed FFD timeseries is 0.91 and the root mean square error is 2.5 d. The DTS model was calibrated over the period 1911–1940 and it is assumed that the estimated model parameters remain unchanged with time. This is an important assumption given the non-stationarity of the climate and the associated earlier occurrence of the FFDs with rising temperatures. It is evident, however, that the DTS model performs well outside its calibration period (figure 4(a)), with no degeneration in recent decades when the major shift towards earlier FFDs takes place. The good level of agreement between observed and reconstructed FFDs suggests that the DTS model does not introduce notable uncertainty in our analysis.

Zoom In Zoom Out Reset image size

Timeseries of observed and model-based FFDs are illustrated in figure 4(b). As there is no long-term change in the NAT climate, FFDs as early as in 2021 are rare. However, even without anthropogenic forcings, such rare events may still take place because of internal variability, and we find 6 years with FFDs in late March out of the 9405 years provided by the NAT simulations. To illustrate the mean shift of the flowering season with time we next plot the mean of the 51 ALL timeseries, for which much of the variability effect is removed (red line in figure 4(c)). We also compare the change in Kyoto with what it would have been without urban warming, using the ALL data but without the urban bias correction (light blue line). The figure also shows the mean FFD in the NAT climate calculated as the average of all the NAT year estimates (green line). Both urban and rural timeseries are consistent with the natural mean FFD in earlier decades but begin to diverge in later years towards earlier flowering seasons. Interestingly, the change comes several decades earlier in an urban setting, highlighting the importance of urban forcing on local scales. We will next quantify the anthropogenic contribution to changes in the FFD and will demonstrate a human-caused increase in the risk of extreme events like in year 2021.

We compute the change in the FFD relative to the natural climate in consecutive decades from 1850 to the end of the 21st century (figure 5) following the method in section 2.4. The mean change (diamonds in figure 5) provides a measure of the total anthropogenic effect on the FFDs and the 5%–95% range (vertical bars) represents interannual departures from the mean change due to internal variability. We find that after about the 1930s, human influence leads to earlier flowering, with an estimated shift of 11 d in the present climate and 17 d by 2100. The anthropogenic effect is detectable in the present decade above internal variability, as the range of the FFD change no longer encompasses zero. Without urban warming, the anthropogenic signal would emerge several decades later, at around the end of the 20th century, and would lead to a smaller shift of 5 d at present and 11 d by 2100. Hence, in a rural setting, the current anthropogenic change in Kyoto's flowering season would not be observed until the end of the century. Our results also indicate a slowdown in the shift of the FFDs during the course of this century, though the precise trajectory of future changes depends on the selected emissions scenario. We expect the human impact would be further reduced in a more sustainable development pathway than the middle-of-the-road SSP2 4.5, which stresses the importance of lowering greenhouse gas emissions.

Zoom In Zoom Out Reset image size

Risk-based attribution studies estimate the changing likelihood of extreme events from ensembles of simulations that represent different climatic conditions. Applying the same approach here, we construct probability distributions of the FFD for the present climate, the climate of the late 21st century and the natural climate without human influence (figure 6(a)). The present and future distributions are based on FFD data derived from the ALL experiment and extracted from two 20 year-periods used as a proxy of different climatic conditions, namely years 2012–2031, representing the climate of 2021 and of the present-day, and years 2081–2100, representing the late century. For the distribution of the natural world, we use all the FFD data from the NAT experiment. We also assess the effect of urban warming separately and construct distributions for rural conditions using data from the ALL experiment but without the urban correction (figure 6(b)). It is evident that the 2021 FFD would be extremely rare in the pre-industrial world as it lies well outside the main body of the NAT distribution. Without urban warming, the 2021 event would be more likely in today's climate, but still uncommon as it is positioned in the far tail of the rural distribution. However, the combined effect of urban and other anthropogenic warming has moved the distribution further towards earlier FFDs, increasing the likelihood of the 2021 record, while by 2100 the event lies much closer to the peak of the urban distribution and would therefore no longer be considered extreme.

Zoom In Zoom Out Reset image size

Following the method described in section 2.4, we next calculate the likelihood of extremely early FFDs in the natural world and in 20 year rolling windows from present-day to the end of the century (figure 6(c)). The return time (inverse probability) of extremely early FFDs in the natural climate is estimated to be 1200 years (uncertainty range: 620–2500). The shift of the FFD in a climate influenced by anthropogenic forcings except urban warming is still largely influenced by variability (figure 5, present-day range includes zero) and so the return time of extremely early events encompasses the NAT range (figure 6(c)). The FFD sample used to estimate the rural likelihood is also much smaller than the NAT sample and since the early flowering events considered here are very rare, the estimated return time uncertainty in the present-day climate is much larger compared to NAT. However, in Kyoto's warmer urban climate the anthropogenic signal has emerged more clearly, and the likelihood of extremes is estimated to have already increased 15 times (7–30) because of human influence, with the return time reduced to 83 years (65–110). As temperatures continue to rise, early flowering events are found to become very common by 2100 with a return time of 4.6 years (4.2–4.9). Their likelihood also increases in the rural climate, for which the return time at the end of the century is estimated to be 43 years (34–56). Our results suggest a rapid change in the likelihood of extremely early flowering dates in Kyoto attributed to human influence that also reflects an imminent change in our perception of climatological seasons.