Long-Term Temporal Variability of Flowering Day of Red Spider Lily (Lycoris radiata)

Abstract

In Japan, the flowering of the red spider lily (Lycoris radiata) marks the autumn equinox. To evaluate the effect of climate change on Japanese people’s sense of seasons and this cultural ecosystem service, we examined the spatiotemporal variability of the flowering day (FD) of red spider lily at 9 sites (Maebashi, Choshi, Nagano, Kanazawa, Shizuoka, Tsu, Nara, Wakayama, and Okayama) over the past 60 to 70 years through its relationship with the autumn equinox. (1) Delaying trends were statistically significant (0.12–0.16 days per year) at 4 sites (Nagano, Tsu, Nara, and Wakayama). (2) Bayesian inference analysis with a beta distribution showed that the probability of FD being later than the autumn equinox has increased in the 2010s at all sites. (3) The year-to-year variability of FD was positively correlated with average temperature during the period of flower stalk elongation (late August to mid-September) at 7 sites (except Nagano and Shizuoka). These results suggest that the probability of FD being later than the autumn equinox will increase under further warming during the period of flower stalk elongation, thus affecting people’s sense of seasons and this cultural ecosystem service.

1. Introduction

The flowering of red spider lilies (Lycoris radiata) during the autumn “Higan” (Buddhist services performed during the equinoctial week) marks the autumn equinox in Japan. Japanese people call red spider lilies “Higan bana” (“flowering during Higan”) [1,2]. The lilies have been planted around cemeteries and temples as a symbol of Buddhism and along rice field ridges to deter pest animals [3,4,5,6,7]. In the past, people used the scaly bulbs as famine relief food and glue [3,4,5,6,8,9]. Alternative and local names for red spider lilies related to death, Buddhism, and poison number more than 1000 [6,7,9,10]. Once detested as an ominous flower, now, tens of thousands of lilies that are planted in riverbeds attract tourists [9] (Figure 1). For these reasons, Japanese people anticipate the seasonal change with the flowering of red spider lilies.

Flower stalk elongation in red spider lilies is accelerated at 20 to 25 °C but suppressed at 30 °C [11,12]. This explains why flowering in Japan was delayed in 2024 and 2025, when temperatures remained high during the period of flower stalk elongation (e.g., [13,14,15,16]). In Japan, the increasing trend of temperatures associated with global warming in August (1.33 °C per century) and September (1.49 °C per century during 1898−2005) is remarkable at weather stations unaffected by urbanization [17]. Future high temperatures during this stage due to climate change may further delay flowering, confounding Japanese people’s sense of the season and spoiling the cultural ecosystem service [18].

To reveal the effect of temporal changes in flowering day (“FD”) in red spider lilies on people’s sense of the season, it would be appropriate to examine the relationship between FD and the autumn equinox. However, long-term temporal changes in FD in red spider lilies have not yet been published. In Japan, although standardized phenological observations of typical animals and plants have been conducted since 1953 [19], the number of sites at which red spider lilies have been observed continuously dates from much later than that, and monitoring of red spider lilies was dropped in 2020 to reduce labor costs [19,20]. This situation hinders the development of our understanding of their phenological sensitivity to climate change in autumn, which has not yet been examined, unlike that in spring [21,22].

Towards this goal, we examined changes over time in the probability of the FD occurring after the autumn equinox (“FD_late”) by Bayesian updating of a probability distribution. A major advantage of Bayesian inference is that we can improve the accuracy of prediction by updating the prior probability based on past experience as new observations are obtained. Our aims were to (1) reveal long-term temporal changes in the relationship between FD and the autumn equinox and (2) suggest a proxy method for the collection of future FD records so as to deepen our understanding of the effect of climate change on people’s sense of seasons and on cultural ecosystem services in early autumn.

2. Materials and Methods

2.1. Red Spider Lilies

Red spider lilies are distributed broadly in Japan, China, Korea, and Nepal [23,24]. However, those in Japan are naturalized, having originated in the Yangtze River Basin, China [1,6,7]. They are a sterile triploid perennial herb containing poisonous alkaloids and reproduce via their scaly bulbs. Leaves flush in October after flowering. In the following April, the leaves fall, and flower bud formation starts. The flowers reach the gynoecium development stage in mid-June and the pollen formation stage in mid-August. In September, the flower stalk elongates and 5 to 8 red ring-shaped flowers bloom [2,12,25].

2.2. Observation Data

We used records of FD collected at observation sites across Japan and published by the Japan Meteorology Agency (JMA; [26]). FD was defined as 2 to 3 flowers (a single flower on 1 scape) in bloom. The observed plants grow at weather stations or in nearby parks (within ~5 km horizontally and ±50 m vertically) [27]. We selected 9 observation sites where the ratio of missing data was <25% (17 years) during 1953–2020 (67 years), and continuous observations were recorded since 2000 (Figure 2). The observation periods covered 1953–2020 in Kanazawa; 1964–2020 in Maebashi, Choshi, Shizuoka, Tsu, Nara, and Okayama; and 1966–2020 in Nagano and Wakayama.

Because temperatures during the period of flower stalk elongation significantly influence FD [11,12], we used daily mean temperatures from 23 August to 22 September (“temp_8–9”) at all 9 sites published by the JMA [28].

2.3. Relationship Between Autumnal Equinox Day and FD

The probability density function of FD_late follows a beta distribution: Be(α, β):

$$f\left(p\right)={\displaystyle \frac{p^{\alpha -1}{\left(1-p\right)}^{\beta -1}}{{\int }_0^1{p}^{\alpha -1}{\left(1-p\right)}^{\beta -1}dp}}\left(0<p<1,\hspace{0.17em}0<\alpha ,0<\beta \right)$$

where α − 1 represents how many times we observed an event, and β − 1 represents how many times we did not observe an event when we set the first prior distribution as uniform (i.e., non-informative; Be(1, 1)). The prior distribution—Be(α, β) ($mean={\displaystyle \frac{\alpha }{\alpha +\beta }}, variance={\displaystyle \frac{\alpha \beta }{{\left(\alpha +\beta \right)}^2\left(\alpha +\beta +1\right)}}$)—is updated to a posterior distribution—Be(α + r, β + n − r) ($mean={\displaystyle \frac{\alpha +r}{\alpha +\beta +n}}, variance={\displaystyle \frac{\left(\alpha +r\right)\left(\beta +n-r\right)}{{\left(\alpha +\beta +n\right)}^2\left(\alpha +\beta +1+n\right)}}$)—with new observed data—_nC_rp^r(1 − p)^n−r (i.e., likelihood)—by Bayesian inference, where n represents the total number of trials and r represents how many times we observed an event [29].

For instance, let the prior distribution be Be(7, 5) (FD_late = 6 in the first decade). This prior distribution is updated to the new posterior distribution Be(15, 7) with new observed data that FD_late = 8 in the second decade. Let this posterior distribution be the new prior distribution. This new prior distribution is updated to the new posterior distribution Be(17, 15) with new observed data that FD_late = 2 in the third decade.

For each site, to evaluate the decadal changes, we updated the prior distribution to the posterior distribution with new observed data every 10 years by Bayesian inference. We set the first prior distribution as uniform. For instance, in Maebashi, where data were available during 1964–2020, the posterior distribution was updated 6 times (i.e., 1964–1970, 1964–1980, 1964–1990, 1964–2000, 1964–2010, and 1964–2020) (Annual updating yielded the same posterior distributions). Although the autumn equinox can be 22, 23, or 24 September, we defined it as 23 September in normal years and 22 September in leap years.

2.4. Relationship Between FD and temp_8–9

To examine the factors involved in the year-to-year variability of FD, we calculated the linear regression between FD and temp_8–9 at each site.

2.5. Statistical Analysis

The analyses were run in R v. 4.5.1 software [30] using RStudio v. 2024.12.1 build 563, LibreOffice v. 24.2.6.2, and shell scripts.

3. Results

A statistically significant delay trend (0.12–0.16 days per year) was found in Nagano, Tsu, Nara, and Wakayama (Figure 3). In Nagano, Tsu, and Wakayama, with Bayesian updating, the probability of FD_late was highest in 1964/1966–2020 (Figure 4;

Table 1. Probability of occurrence values for Figure 4.

Site	Period	Posterior Distribution	Mean	Variance	Site	Period	Posterior Distribution	Mean	Variance
Maebashi	1964–1970	Be(1, 8)	0.111	0.01	Tsu	1964–1970	Be(1, 8)	0.111	0.01
	1964–1980	Be(1, 18)	0.053	0.002		1964–1980	Be(1, 14)	0.067	0.004
	1964–1990	Be(1, 28)	0.034	0.001		1964–1990	Be(2, 23)	0.08	0.003
	1964–2000	Be(2, 37)	0.051	0.001		1964–2000	Be(2, 33)	0.057	0.001
	1964–2010	Be(2, 47)	0.041	0.001		1964–2010	Be(4, 41)	0.089	0.002
	1964–2020	Be(4, 55)	0.068	0.001		1964–2020	Be(10, 45)	0.182	0.003
Choshi	1964–1970	Be(4, 5)	0.444	0.025	Nara	1964–1970	Be(1, 4)	0.2	0.027
	1964–1980	Be(5, 14)	0.263	0.01		1964–1980	Be(1, 14)	0.067	0.004
	1964–1990	Be(7, 22)	0.241	0.006		1964–1990	Be(1, 24)	0.04	0.001
	1964–2000	Be(12, 27)	0.308	0.005		1964–2000	Be(3, 32)	0.086	0.002
	1964–2010	Be(14, 35)	0.286	0.004		1964–2010	Be(6, 39)	0.133	0.003
	1964–2020	Be(20, 39)	0.339	0.004		1964–2020	Be(9, 45)	0.167	0.003
Nagano	1964–1970	Be(1, 6)	0.143	0.015	Wakayama	1964–1970	Be(1, 5)	0.167	0.02
	1964–1980	Be(1, 16)	0.059	0.003		1964–1980	Be(1, 15)	0.063	0.003
	1964–1990	Be(2, 25)	0.074	0.002		1964–1990	Be(3, 23)	0.115	0.004
	1964–2000	Be(3, 34)	0.081	0.002		1964–2000	Be(4, 32)	0.111	0.003
	1964–2010	Be(4, 41)	0.089	0.002		1964–2010	Be(7, 39)	0.152	0.003
	1964–2020	Be(8, 45)	0.151	0.002		1964–2020	Be(11, 45)	0.196	0.003
Kanazawa	1953–1960	Be(5, 5)	0.5	0.023	Okayama	1964–1970	Be(4, 5)	0.444	0.025
	1953–1970	Be(7, 9)	0.438	0.014		1964–1980	Be(4, 15)	0.211	0.008
	1953–1980	Be(7, 19)	0.269	0.007		1964–1990	Be(4, 25)	0.138	0.004
	1953–1990	Be(9, 27)	0.25	0.005		1964–2000	Be(4, 35)	0.103	0.002
	1953–2000	Be(14, 31)	0.311	0.005		1964–2010	Be(5, 43)	0.104	0.002
	1953–2010	Be(17, 38)	0.309	0.004		1964–2020	Be(8, 50)	0.138	0.002
	1953–2020	Be(21, 43)	0.328	0.003
Shizuoka	1964–1970	Be(3, 5)	0.375	0.026
	1964–1980	Be(3, 15)	0.167	0.007
	1964–1990	Be(3, 25)	0.107	0.003
	1964–2000	Be(4, 34)	0.105	0.002
	1964–2010	Be(4, 44)	0.083	0.002
	1964–2020	Be(6, 52)	0.103	0.002

). In Kanazawa, it was highest in 1953–1960. In Choshi, Shizuoka, and Okayama, it was highest in 1964–1970. With Bayesian updating, despite decadal fluctuations, the probability increased in 1953/1964/1966–2020 (Figure 4; Table 1). The linear regression between FDs and temp_8–9 was significantly positively correlated at all sites except for Nagano and Shizuoka (Figure 5;

Table 2. Statistical values of regression functions in Figure 5.

Figure	Site	Regression Function	R2	Testing Significance of Regression *
				Intercept	Slope
Figure 5a	Maebashi	y = 1.30x + 226.90	0.09	t = 17.50, p < 0.001	t = 2.38, p < 0.05
Figure 5b	Choshi	y = 2.12x + 212.15	0.18	t = 14.29, p < 0.001	t = 3.44, p < 0.01
Figure 5c	Nagano	y = 1.15x + 236.30	0.07	t = 18.10, p < 0.001	t = 1.96, p = 0.06
Figure 5d	Kanazawa	y = 1.08x + 238.42	0.08	t = 21.05, p < 0.001	t = 2.30, p < 0.05
Figure 5e	Shizuoka	y = 0.50x + 247.23	0.01	t = 15.79, p < 0.001	t = 0.80, p = 0.43
Figure 5f	Tsu	y = 1.44x + 224.62	0.12	t = 16.45, p < 0.001	t = 2.67, p < 0.05
Figure 5g	Nara	y = 1.36x + 229.11	0.13	t = 18.65, p < 0.001	t = 2.69, p < 0.01
Figure 5h	Wakayama	y = 2.59x + 194.35	0.23	t = 11.33, p < 0.001	t = 3.91, p < 0.001
Figure 5i	Okayama	y = 1.06x + 233.40	0.11	t = 22.04, p < 0.001	t = 2.55, p < 0.05

* Null hypothesis is that coefficient = 0 (H0: intercept or slope = 0).

). The slope of the linear regression was larger in Choshi and Wakayama than at the other seven sites (Figure 5; Table 2).

4. Discussion

At just over half of the sites, FD did not show a linear delay, but the probability of FD_late was likely to increase, especially in the 2010s (Figure 3 and Figure 4). The positive correlation between FD and temp_8–9 at most sites (Figure 5; Table 2) is consistent with previous reports based on manipulation experiments in which FD was delayed by high temperatures during the period of flower stalk elongation [11,12]. The fact that we could detect the decadal-scale temporal change of FD, which conventional linear regression analysis could not detect, is an advantage of Bayesian inference. However, in Choshi and Kanazawa, FD_late was frequent in the 1950s to the 1990s, when high temp_8–9 was less frequent than later (Figure 3). Thus, high temperatures during the period of flower stalk elongation do not alone explain the latter exception.

A previous report based on manipulation experiments showed different optimum temperatures during the flowering process: from the beginning of flower bud formation to the gynoecium development stage (April to June), 25 to 30 °C; after the formation of the gynoecium, 20 to 25 °C; and from the pollen formation stage to flowering (mid-August to September), 20 to 25 °C [12]. These results indicate that flower bud formation in red spider lilies is accelerated at high temperatures during the beginning of flowering but is suppressed during the later part of flowering. This was supported by another previous report based on manipulation experiments [31]. In addition, Mori et al. [12] showed an indirect effect of high temperatures after the formation of the gynoecium on flowering: when red spider lilies that form the gynoecium at 30 °C for 3 to 5 weeks are then moved to 20 °C, they flower 4 to 5 weeks earlier than those grown at 20 °C. To deepen our understanding of flowering in red spider lilies, we may need to examine these trade-off relationships between high temperatures and flowering in more detail. In addition, various conditions during observation, such as the surroundings and meteorological factors, might cause uncertainty in our analysis.

Despite the uncertainty caused by insufficient understanding of the flowering process and the systematic errors present in the data, we predict that future higher temperatures during the period of flower stalk elongation due to climate change are likely to increase the probability of FD_late. To validate this possibility, we need to analyze proxy data collected after 2020, when the JMA dropped phenology observations in red spider lily [19,20], by Bayesian inference. Useful proxy data include data collected by citizen scientists [32,33,34], flowering information published on websites (e.g., [35,36]), and text, images, and videos published on social media such as X, Instagram, and YouTube [37,38]. With images and videos, we can evaluate the flowering status of tens to thousands of individuals at the same site. This advantage reduces the uncertainty caused by the spatial representativeness of samples in conventional visual records (e.g., phenology records published by the JMA).

At the same time, we need to associate these proxy data with conventional visual records. However, it is not so easy to obtain proxy data at conventional observation sites. To resolve this issue, at sites where proxy data are available (mainly tourist spots), we need to retrieve phenological records from old newspapers and magazines, and then to compare the proxy data with conventional visual records. Unlike the retrieval of flowering records from historical diaries and documents (e.g., [39,40]), few studies have retrieved proxy data from before the social media era. To deepen our understanding of the sensitivity of flowering phenology in red spider lilies to climate change, we should retrieve proxy data from them.

5. Conclusions

We examined the change over time in the probability of years when FD of red spider lilies occurred after the autumn equinox (FD_late) during the past 60 to 70 years at 9 sites in Japan. The probability increased in the 2010s at all sites, with a significant linear delay at 4 sites. Except at 2 sites, FD was positively correlated with temperature during the period of flower stalk elongation (late August to mid-September). These results suggest that rising temperatures during the period of flower stalk elongation will increase the probability of FD_late, disrupting people’s sense of the seasons and cultural ecosystem services.