Process-based models outcompete correlative models in projecting spring phenology of trees in a future warmer climate
Introduction
Phenology is a key aspect of plant and animal life strategies because it determines the timing of growth and reproduction. Life cycle of species must be adapted to the local weather conditions and resources. As a consequence, phenology is one of the top controls of crop yield (Olesen et al., 2011; Nissanka et al., 2015), population dynamics (Anderson et al., 2013), species distribution (Chuine, 2010) and evolutionary dynamics (Duputié et al., 2015; Burghardt et al., 2015). Phenology also ultimately regulates many functions of ecosystems such as productivity (Richardson et al., 2012), ecosystem carbon cycling (Delpierre et al., 2009), water (Hogg et al., 2000) and nutriment cycling (Cooke and Weih, 2005).
Since the 1970s, spring phenology has been reported to advance in response to warming (Walther et al. 2002; Parmesan and Yohe, 2003; Menzel et al. 2006; Fu et al. 2014). For instance, it has been shown that the apparent response of leaf unfolding to temperature was –3.4 days per °C between 1980 and 2013 in temperate Europe (Fu et al., 2015). This advance in spring phenology events is due to the warming of springs as bud growth rate is positively and strongly related to temperature (see for review Chuine and Régnière, 2017). However, this trend has been slowing down by about 40% after 2000 (Fu et al., 2015). One of the most likely hypotheses to explain this slowdown is the warming of winters (Asse et al., 2018). Indeed, most of temperate and boreal trees have developed a key adaptation to winter cold: the inability to resume growth despite transient favorable growing conditions in terms of temperature (Howe et al., 2000). This particular physiological state, called endodormancy (Lang et al., 1987), establishes in fall and disappears in early to late winter after a certain exposure to cold temperatures. Therefore, the warming of winters is suspected to delay endodormancy release and be responsible for the apparent decrease in the response of leaf unfolding to warming after 2000. Ultimately, this lack of chilling temperatures might compromise budburst itself at some point if warming continues (Chuine et al., 2016). Such a situation is more likely to occur in populations inhabiting the warm edge of a species range and/or lower elevations in mountain regions, where species are already in suboptimal chilling conditions (Benmoussa et al., 2018; Guo et al., 2015). More than ever there is a need for more accurate projections of tree phenology for the upcoming decades to remove the large uncertainties that still remains.
Two main categories of predictive phenology models exist although there can be a continuum in-between: correlative and process-based models (for review see (Chuine et al., 2013; Chuine and Régnière, 2017). Correlative models statistically relate descriptors of climate to phenological variables (i.e. usually the occurrence dates of a phenological phase such as bud break or flowering). In correlative models, parameters have no a priori defined ecological meaning and processes can be implicit (process-implicit) (Lebourgeois et al., 2010). In contrast, process-based models are built around explicitly stated mechanisms and parameters have a clear ecological interpretation that is defined a priori. In this category of model, response curves are often obtained directly from experiments, contrasting with empirical relationships of correlative models. Consequently, process-based models often require a larger number of parameters to be estimated or measured, with the consequence of a higher level of complexity than correlative models. However, they provide greater insights into how precisely each driver affects the trait, and they are expected to provide more robust projections in new climatic conditions corresponding either to other geographical areas or other time periods (Chuine et al., 2016).
Among the most widely used process-based phenology models, are the so-called one-phase models that describe solely the ecodormancy phase, which follows the endodormancy phase. During the ecodormancy phase, bud cell elongation can take place whenever temperatures are appropriate, and the higher the temperature is, the higher is the rate of cell elongation during this phase. This category of models has been shown to be efficient in predicting accurately budburst date under historical climate (Vitasse et al., 2011; Basler, 2016). Another category of models, called two-phase models, describes additionally the endodormancy phase, and take into account the possible negative effect of winter warming on endodormancy release. This category of models is thus considered to provide more accurate projections in future climatic conditions (Chuine, 2010; Vitasse et al., 2011). However, it has been shown recently that this second type of models might suffer from flawed parameter estimation when dates of endodormancy release have not been used for their calibration (Chuine et al., 2016). Unfortunately, observations of endodormancy release date are very rare because they are very difficult to determine (Jones et al., 2013; Chuine et al., 2016). Models are thus usually calibrated using solely bud break or flowering dates (Chuine, 2000; Caffarra et al., 2011; Luedeling et al., 2009; but see Chuine et al., 2016). Some other models go further in the description of the processes by integrating a photoperiod cue (Schaber and Badeck, 2003; Gaüzere et al., 2017). Some studies indeed support the hypothesis that in photosensitive species, which might represent about 30% of temperate tree species (Zohner et al., 2016), long photoperiod might compensate for insufficient chilling (Caffarra et al., 2011a; Gaüzere et al. 2017).
There is now a large number of phenology models that differ in their level of complexity and in the types of response function to environmental cues they use (see for review Chuine et al., 2013; Basler, 2016; Chuine and Régnière, 2017). However, very few studies aimed at comparing their efficiency and robustness so far (Basler, 2016), especially in future climatic conditions (but see Vitasse et al., 2011; Chuine et al., 2016; Gaüzere et al., 2017), while this has been done multiple times for species distribution models (e.g. Cheaib et al., 2012; Higgins et al., 2012; Morin and Thuiller, 2009; Kearney et al., 2010) and crop models (Lobell and Burke, 2010) for example. By efficiency, we mean here the ability of the model to provide accurate predictions in conditions that have been used to calibrate the model (Janssen and Heuberger, 1995); and by robustness, we mean here the ability of the model to provide accurate predictions in external conditions (Janssen and Heuberger, 1995), i.e. other conditions than those used to calibrate the model. Model's robustness determines its transferability in time and space.
Process-based models are usually expected to provide more accurate projections for the future than correlative models because they describe causal relationships. The effect of each driver identified as affecting a particular trait value can be described by a causal relationship, sometimes involving other drivers as well (interaction between drivers). For this reason, process-based models have also an expected greater potential to deal with non-analog situations. However, the putative higher robustness of process-based models could also come from the fact that parameter values describing the causal relationships, or at least some of them, can be measured directly (forward estimation of parameter values) instead of being inferred statistically through inverse modelling techniques and data assimilation (backward estimation of parameter values). Yet, there has been no attempt so far to validate this widely accepted expectation.
Here, we aimed at comparing the efficiency and robustness of correlative vs process-based phenology models with contrasting levels of complexity, both in space and time. More precisely, we aimed at answering the following questions: (1) Are process-based phenology models more robust than correlative models? (2) If so, is it because they describe causal relationships or because they can be less dependent on statistical inference (i.e. back estimation of parameter values) and rely more on experimental measurements (i.e. forward estimation of parameter values)? (3) How do projections of both types of model differ in future climatic conditions?
Using observations of budburst dates collected over the Western Alps by a citizen science program during 8 years, and experimental data, we calibrated correlative and process-based phenology models with three levels of complexity for five major tree species: Corylus avellana (L.), Fraxinus excelsior (L.), Betula pendula (Roth), Larix decidua (Mill.) and Picea abies (L.). We then compared their predictions and projections over the Western Alps in historical climate and in future climate respectively.
The Western Alps are particularly interesting to evaluate phenology models because the elevation gradient provides a wide temperature range on a very short distance. In addition, the southern part of the Western Alps is nearly located at the warmest edge of the geographic range of the five studied species, where it has been shown that winter warming is already affecting endodormancy release and budburst Asse et al., 2018). Finally, temperature has already increased in the Western Alps twice as fast as in the northern hemisphere during the 20th century (Rebetez and Reinhard, 2008) and recent evidence indicates that the current warming rate increases with increasing elevation (Mountain Research Initiative EDW Working Group, 2015). Consequently, mountain summits might warm faster than lower elevation sites, so that the response of mountain ecosystems to climate change might be non-linear along elevation gradients. Therefore, ultimately, we aimed at answering a fourth question: ((4) How will climate change alter the budburst date of alpine species?
Section snippets
Phenological and meteorological data
We used observations of the budburst date, defined as the first day when 10% of vegetative buds of a given individual tree are opened (BBCH 07), of five common tree species: ash (Fraxinus excelsior L.), birch (Betula pendula Roth), hazel (Corylus avellana L.), larch (Larix decidua Mill.), and spruce (Picea abies L.). These species show different elevation ranges (from 150 to 1300 m a.s.l. for Corylus to 700–2100 m a.s.l. for Larix), which allowed us to compare the two types of model over a
Selection of the best models
Mixed effects models which best explained budburst dates were generally models including GDD5, and chilling as predicting variables (Appendix B). However, budburst dates of Corylus and Fraxinus were best explained by GDD0 together with chilling. Variance partitioning also indicated an important joint contribution of GDD0 or GDD5 and chilling for all species (Appendix B). However, over all 8 years and all locations, chilling did not explain a significant part of the variance in spring phenology
Process-based phenology models provide more robust projections than correlative models
Linear mixed models using chilling days and growing degree days (GDD) were more efficient than process-based models in predicting the budburst date in calibration conditions (i.e. sensu Randin et al., 2006). However, process-based models were more efficient than correlative models to project budburst date in external conditions (except for Picea for which the three model types revealed similar performance). Error in predicted dates and projected dates increased with the distance to the mean
Conclusion
Our results showed that (1) process-based phenology models are more robust than correlative model even when they rely entirely on backward estimation (inverse modelling) of their parameter values. (2) They also demonstrated that the robustness of process-based models could be increased, though not substantially, when their calibration could rely on forward estimation. Therefore, the robustness of process-based models seems to come primarily from the explicit description of causal relationships
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
We are grateful to all the volunteers and staffs of protected areas involved in the Phenoclim program for their help and support for collecting data on the study sites. We also thank Geoffrey Klein (Centre de Recherches sur les Ecosystèmes d'Altitude CREA, Chamonix-Mont-Blanc) for having prepared the temperature data used in this study. We thank the Conservatoire Botanique National Alpin for providing the species occurrence data, and two anonymous referees for their constructive comments on
References (75)
- et al.
Modeling climate change impacts on phenology and population dynamics of migratory marine species
Ecol. Modell.
(2013) - et al.
Warmer winters reduce the advance of tree spring phenology induced by warmer springs in the Alps
Agric. For. Meteorol.
(2018) Evaluating phenological models for the prediction of leaf-out dates in six temperate tree species across central Europe
Agric. For. Meteorol.
(2016)- et al.
Photoperiod sensitivity of bud burst in 14 temperate forest tree species
Agric. For. Meteorol.
(2012) - et al.
Shortcomings of classical phenological forcing models and a way to overcome them
Agric. For. Meteorol.
(2012) - et al.
Influences of cold deprivation during dormancy on carbohydrate contents of vegetative and floral primordia and nearby structures of peach buds (Prunus persica L. Batch)
Sci. Hortic. (Amsterdam).
(2005) A Unified Model for Budburst of Trees
J. Theor. Biol.
(2000)- et al.
A spatial modelling framework for assessing climate change impacts on freshwater ecosystems : Response of brown trout (Salmo trutta L .) biomass to warming water temperature
Ecol. Modell.
(2015) - et al.
Integrating interactive effects of chilling and photoperiod in phenological process-based models. A case study with two European tree species: Fagus sylvatica and Quercus petraea
Agric. For. Meteorol.
(2017) - et al.
Responses of spring phenology in temperate zone trees to climate warming: A case study of apricot flowering in China
Agric. For. Meteorol.
(2015)
Calibration of process-oriented models
Ecol. Modell.
Agricultural and forest meteorology: How accurately can minimum temperatures at the cold limits of tree species be extrapolated from weather station data ?
Agric. For. Meteorol.
On the use of statistical models to predict crop yield responses to climate change
Agric. For. Meteorol.
Validation of winter chill models using historic records of walnut phenology
Agric. For. Meteorol.
Deciphering and prediction of transcriptome dynamics under fluctuating field conditions
Cell
River flow forecasting through conceptual models part I – a discussion of principles
J. Hydrol.
Calibration of the phenology sub-model of APSIM-Oryza: Going beyond goodness of fit
Environ. Model. Softw.
Impacts and adaptation of European crop production systems to climate change
Eur. J. Agron.
Assessing the effects of climate change on the phenology of European temperate trees
Agric. For. Meteorol.
Simulation of phenological development of wheat crops
Agric. Syst.
Climate change threatens central Tunisian nut orchards
Int. J. Biometeorol.
“Prediction” or “Projection”?
Sci. Commun.
Modeling the influence of genetic and environmental variation on the expression of plant life cycles across landscapes
Am. Nat.
Modelling the timing of Betula pubescens budburst. II. Integrating complex effects of photoperiod into process-based models
Clim. Res.
Modelling the timing of Betula pubescens budburst. I. Temperature and photoperiod: A conceptual model
Clim. Res.
Climate change impacts on tree ranges: Model intercomparison facilitates understanding and quantification of uncertainty
Ecol. Lett.
Why does phenology drive species distribution?
Philos. Trans. R. Soc. B Biol. Sci.
Phenology is a major determinant of tree species range
Ecol. Lett.
Can phenological models predict tree phenology accurately in the future? The unrevealed hurdle of endodormancy break
Glob. Chang. Biol.
Selecting models to predict the timing of flowering of temperate trees: Implications for tree phenology modelling
Plant, Cell Environ.
Fitting models predicting dates of flowering of temperate-zone trees using simulated annealing
Plant, Cell Environ.
Plant development models
Phenology: An Integrative Environmental Science
Process-based models of phenology for plants and animals
Annu. Rev. Ecol. Evol. Syst.
Chilling and heat requirements for leaf unfolding in European beech and sessile oak populations at the southern limit of their distribution range
Int. J. Biometeorol.
Cited by (21)
Predicting spring phenology in deciduous broadleaf forests: NEON phenology forecasting community challenge
2024, Agricultural and Forest MeteorologyLengthening height-growth duration in Smith fir as onset becomes more synchronous across elevations under climate warming scenarios
2022, Agricultural and Forest MeteorologyCitation Excerpt :Furthermore, primary growth associated with bud development, apical meristem activation, and shoot elongation is intimately tied to carbon allocation and secondary xylem growth of stem wood formation (e.g. Zhang et al., 2019; Buttò et al., 2021). Most studies on tree phenological variation have focused mainly on bud burst, leafing, and some on timing and rate of radial growth (Yang et al., 2017; Li et al., 2019; Asse et al., 2020; Gao et al., 2022), whereas little attention has been paid to height-growth phenology. Height growth could interact tightly with phenological events of other organs (e.g., leaf and root growth) through hormone signals and source–sink communications, and eventually influence the growth trade-off and resource allocation among organs (Kearsley et al., 2013; Thomas 2013; Zani et al., 2020).
Climate change-induced reduction in agricultural land suitability of West-Africa's inland valley landscapes
2022, Agricultural SystemsCitation Excerpt :Still, some aspects warrant further discussions. The models relate descriptors of climate to inland valley rice locations empirically and processes can be implicit (Asse et al., 2020) and are estimated by response curves (Dormann, 2007). However, a previous study showed that response curves to both climatic and other biophysical predictors of inland valleys rainfed rice cultivation have a reasonable ecological explanation (Akpoti et al., 2020).
PhenoFlex - an integrated model to predict spring phenology in temperate fruit trees
2021, Agricultural and Forest MeteorologyCitation Excerpt :Sunset and sunrise times were computed based on geographic latitude, using equations by Spencer (1971) and Almorox et al. (2005), as implemented in the chillR package for R (Luedeling, 2020). We compared four models regarding their ability to predict the bloom dates for the two fruit tree cultivars: (a) the newly developed PhenoFlex model, b) the same PhenoFlex model but with the parameters of the chill-submodel set to the fixed values that are commonly used in phenological studies (c) a simplified version of the Unified Model, denoted as ‘StepChill’ model, which was recently identified as the best performing model in another comparative study (Asse et al., 2020), and (d) a machine learning algorithm, as a completely uninformed but potentially powerful alternative, which is increasingly used in agricultural research. We also evaluated the dataset with a set of ten phenology models contained in the phenor package for R (Hufkens et al., 2018).
Drought risk assessment for early maize growth in Northeast China based on a reconstructed phenological dataset
2024, Journal of Agronomy and Crop Science
- †
Lead author.