We assembled a unique dataset of range-wide observations of Yucca available from iNaturalist. These provided the bases for determining presence and absence of flowers for our two focal species, and while flowering Yucca are likely to be more photographed, we still had sufficient absences for these common and iconic species to use in downstream models. These models were effective at predicting flowering phenology of the two focal species with generally high accuracy during the normal flowering season. Unlike other frameworks for predicting phenology, our approach is not to estimate an onset, median or termination timing of a phenophase. Rather, our goal was to determine whether climate and daylength covariates provide a basis for predicting the probability of open flowers during normal and anomalous bloom periods. This approach is enabled by having dense reporting of presence-absence data generated by growing community science resources.
Model fitting revealed that probability of flowering is determined by complex interactions between climate and daylength, suggesting the critical importance of climate context. This importance for arid-adapted Yucca is in line with other studies of phenology of desert plants, from Saguaro in the desert Southwest of North America9 to lilies in arid environments in Africa25. Yet, we were surprised that flowering of Yucca does not necessarily always rely on increased precipitation. We expected precipitation would be a critical limiting factor in desert environments, and for Mojave yucca in particular, precipitation is a strong driver positively influencing the odds of flowering during spring and summer. However, for both species, flowering odds are also relatively high during cold and drier conditions earlier in the season (Fig. 3). These results may relate to precipitation falling as snow rather than rain during late winter—another avenue for further exploration in follow up studies.
Joshua trees and Mojave yucca have different growth forms and are of vastly different sizes at maturity, and therefore, may be expected to react differently to climatic drivers. However, our findings indicate that the same interacting climate variables drive flowering phenology for both species, and the overall shape of their seasonal phenology curves are similar (Fig. 2). The main differences in our models are likely attributed to adapted differences in overall bloom timing; in particular, Y. schidigera generally blooms later in the year than Y. brevifolia (Fig. 2). For example, Y. brevifolia only blooms under highest daylength conditions if it is unusually cold and wet, while Y. schidigera flowering odds can often exceed 50% under warm and wet conditions when daylength is long (Fig. 3). Conversely, while Y. brevifolia has better odds of blooming in cold, dry conditions early in the year, even under wet conditions it can bloom with odds above 25%. In contrast, Y. schidigera is rarely in bloom in colder, wetter early season conditions.
We also note the importance of including a polynomial term for daylength (Table 1), which always dramatically improved models (Table 1). The outcome, clearly visible in Fig. 3, is that responses of phenology are strongly non-linear across gradients. In sum, our work corroborates the importance of context dependence, finding that daylength, temperature, and precipitation interact in complex, nonlinear ways to influence flowering times.
A key question we sought to answer in this work is whether we predict anomalous flowering events. Accelerating climate change means that species will experience conditions outside the range experienced for centuries. How species respond phenologically to these novel conditions is an area of active research26, but the focus has predominantly been on using yearly anomaly data, e.g. warmest years on record or via warming experiments27. Our efforts here are trying to predict a seasonal anomaly, where plants seasonally flowered outside of their presumed normal periods (e.g. in fall rather than spring). A key question of interest was whether the fall-winter bloom in 2018–2019 was itself triggered by anomalous climate conditions mirroring those of the usual bloom period.
We examined this question by testing whether models, which were fit using data from years with a known normal blooming period, were able to predict presences and absences during the 2018–2019 fall-winter season. Our results show that we can predict absences with low error rates (4.7–6.7%, Table 2). However, these models had much higher rates for false positives (32.1–50.2%, Table 2). Our model predicts more anomalous blooming than actually observed. This suggests that, while Yucca might have been triggered to bloom by atypical cooler and wetter conditions, there are still factors not included in our models that limited the extent of anomalous blooming.
It remains possible that co-evolution between Yucca and their obligate pollinator or florivore community28 may extend to how phenology is cued. It may also be that Yucca are responding not only to instantaneous climate conditions, such as mean photoperiod, but whether days are shortening or lengthening, and if so, it may be that the modeling approach used here is not sensitive enough to capture these types of more dynamic seasonal cues. Our work may also point to out-of-normal season blooming simply being more common and widespread than previously suspected, given broadly suitable climate conditions. A next step is to use growing community-science reporting of Yucca plants in flower to determine the rate of seasonally anomalous flowering from dense, range-wide community science observations enabled via resources such as iNaturalist.
Finally, we note the value of examining predictive power of models using climate measurements over shorter and longer temporal windows. While these different climate accumulation windows are by nature highly autocorrelated, we found that data from the longer temporal window led to modest improvement of models based on AUC statistics. It is likely that the longer temporal window captures more information about GDD and overall water input in the environment. For example, a classic paper by Beatley5 that focused on shrub phenology in the Mojave showed fall and winter rains were precursor triggers of phenological events in spring. We also note congruence with the findings of Clair and Hoines13, who showed strong positive correlations with the 30-year averages of temperature and precipitation and fruit and seed mass in Joshua trees, although they focused on broader temporal scale questions. Connecting these longer timescale and broad spatial phenology studies, and aligning lags over different time-scales is a frontier area in phenology research. Finally, our findings should not necessarily be extrapolated more broadly for arid-adapted plants, and traits such as perenniality or woody versus herbaceous habit with associated differences in costs for growth and reproduction, may condition thresholds for needed accumulation of heat or water.
We close by noting that phenology modeling is often treated as a one-off exercise where models are built, and results shared. We argue that the accelerating growth of new data resources and flexible modeling frameworks provide a means for models to iteratively improve. One key step towards this goal is faster annotation of phenology state. Here we hand-coded two key states in Yucca photographs1. These carefully vetted classifications can now provide the basis for more automated approaches for annotating photographs, e.g. via machine learning29. These new results can be fed into current models to test and improve model performance.
Expanding data resources for modeling flower presence is one key step, but the development of improved phenology models that include more fitness-relevant responses is also important, such as number of flowers or fruits, potentially in relation to vegetative biomass. Individual yucca plants, for example, do not bloom annually even in favorable conditions, because vegetative growth must precede production of a heavy, high-cost inflorescence10. More sophisticated species-level models that link the full range of environmental conditions populations experience across their range with seasonal vegetative and reproductive biomass proxies are uncommon, mostly due to data limitations. Rather, studies typically focus on single, local areas or transect approaches across broad-scales30, with associated limitations for further prediction or forecasting.
We argue that the ability to develop ecophysiological-guided, range-wide models are in reach, using the same community science photographs that so far have only been used to generate simple states such as open flower presence. Such models hold promise in helping to provide a basis for improved understanding of mechanisms underlying flowering, and better detection of anomalous blooming events and their consequences. For example, we don’t yet know if anomalous blooms produce fewer or greater flower numbers, as compared to normal periods. Do these blooms ultimately lead to the production of fruit and, if so, how much? Such next-step approaches are particularly critical and necessary, because as we experience more unusual weather phenomena and novel conditions, phenology prediction and understanding the consequences of phenological changes becomes even more challenging. As weather forecasting was improved by assimilating more data and building better process parameters, our hope is that similar methods with richer data types can improve the most difficult phenology prediction challenges.
Source: Ecology - nature.com
