Phenology is experiencing dramatic changes over deciduous forests in the
USA. Estimates of trends in phenology on the continental scale are
uncertain, however, with studies failing to agree on both the magnitude and
spatial distribution of trends in spring and autumn. This is due to the
sparsity of in situ records, uncertainties associated with remote sensing data, and
the regional focus of many studies. It has been suggested that reported
trends are a result of recent temperature changes, though multiple processes
are thought to be involved and the nature of the temperature forcing remains
unknown. To date, no study has directly attributed long-term phenological
trends to individual forcings across the USA through integrating
observations with models. Here, we construct an extensive database of ground
measurements of phenological events across the USA, and use it to calibrate
and evaluate a suite of phenology models. The models use variations of the
accumulative temperature summation, with additional chilling requirements
for spring phenology and photoperiod limitation for autumn. Including a
chilling requirement or photoperiod limitation does not improve model
performance, suggesting that temperature change, especially in spring and
autumn, is likely the dominant driver of the observed trend during the past
3 decades. Our results show that phenological trends are not uniform over
the contiguous USA, with a significant advance of 0.34 day yr
Plant phenology, such as the timing of spring budburst and autumn leaf fall, is sensitive to temperature variation (Körner and Basler, 2010; Polgar and Primack, 2011; Richardson et al., 2013) and thus exhibits a long-term trend with the changing climate (Badeck et al., 2004; Cleland et al., 2007; Gordo and Sanz, 2009; Jeong et al., 2011). Long-term changes in phenology may affect ecosystem carbon assimilation (Keenan et al., 2014), surface water and energy balance (Schwartz and Crawford, 2001), and forest composition and evolution (Forrest and Miller-Rushing, 2010). Emerging observations have shown advanced springs and delayed autumns over the Northern Hemisphere, especially in Europe, during the past several decades (Menzel and Fabian, 1999; Fitter and Fitter, 2002; Menzel et al., 2006; Gordo and Sanz, 2009). However, the extent of regional phenological trends in the USA remains uncertain as different studies present inconsistent and even opposite results (Table 1).
The uncertainty of the phenological changes in US forests could be attributed to genetic, geographic, and temporal factors. First, experiments have suggested that different species may have different phenological sensitivity to temperature (Vitasse et al., 2009). Some species may also require cold temperatures before budburst (called chilling requirement), leading to divergent responses of US plants to spring and winter warming at the community level (Cook et al., 2012) and the continental scale (Zhang et al., 2007). In addition, it is not clear whether other biotic and/or abiotic factors (e.g., humidity, photoperiod, tree age, and tree species) may play a role (Morin et al., 2009; Basler and Korner, 2012; Vitasse, 2013; Caldararu et al., 2014; Laube et al., 2014). Second, most deciduous forests in the USA are found at the mid-latitudes, where temperature increases have not been uniform, and are not as strong as those at high latitudes (Hartmann et al., 2013). Third, differences in the time frames used in different studies may lead to apparently inconsistent trends (Badeck et al., 2004).
Summary of studies estimating phenology trend in the USA for at least 20 years.
NDVI: normalized difference vegetation index; AVHRR: Advanced Very High-Resolution Radiometers; GIMMS: Global Inventory Mapping and Monitoring Studies
There are generally three approaches for estimating phenology at regional
and continental scales: ground networks, remote sensing, and numerical
modeling. Ground-based measurements can provide the most accurate
phenological dates, such as budburst, flowering, and leaf fall. Some records
last for decades and even centuries (Sparks and Menzel, 2002), making it
possible to study long-term phenological change. However, such measurements
usually have very limited spatial coverage. Ground-based networks, such as
the North American Lilac Network (Schwartz and Reiter, 2000),
improve the spatial coverage but focuses only on 1–2 species, which may not
represent the average phenological status of local plants. More extensive
networks, such as the North American Phenology Network (
Phenological models are useful tools for diagnosing causes of phenological changes and also for understanding the feedback of those changes to the Earth system (Richardson et al., 2013; Zhao et al., 2013). Evaluations of well-calibrated phenological models have shown high correlations between predictions and observations (e.g., White et al., 1997; Richardson et al., 2006; Delpierre et al., 2009; Vitasse et al., 2011). However, most of these state-of-art schemes are not evaluated at continental or even larger scales, thus limiting their applicability in dynamic vegetation models and climate models. Recent model–data comparisons have shown that the bias in the prediction of vegetation phenology is a large source of uncertainty in models of ecosystem carbon uptake (Richardson et al., 2012). This necessitates the development and evaluation of continental-scale phenology models with continental-scale observations.
In this study, we use an extensive data set of phenological observations to calibrate and evaluate 13 models (9 for spring and 4 for autumn) of deciduous tree phenology across the USA. We first calibrate each model using derived phenological dates based on the long-term ground observations of leaf area index (LAI) at four deciduous forests. We then examine modeled interannual variability and trends, along with regional phenological differences, using an extensive network of phenological observations. The phenology model best supported by the observations is then applied to (1) estimate the trend of both spring and autumn phenology of US deciduous forests over the last three decades; (2) compare our results with other approaches (ground network, remote sensing, and model based) to identify robust changes and assess discrepancies; and (3) examine the underlying drivers of both the observed trends and interannual variability.
We assembled and compared a suite of published models of spring and autumn phenology. Most of these models are built using cumulative thermal summations with constraining processes, such as chilling requirements and photoperiod limits. Model parameters were calibrated using long-term observations at four deciduous forest sites, with some model constants estimated based on literature values. An independent data set of ground measurements was compiled and used to validate the performance of these models. In total, phenological observations from 1151 sites were used for model validation. In this section we first present the observations used for calibration and validation, followed by a description of the various model formulations tested and simulations performed.
Decadal measurements of LAI from four US deciduous broadleaf sites are
collected from the Ameriflux network (
Ground measurements of leaf area index (LAI) used to calibrate the phenology model. The location of these sites is denoted on Fig. 1.
Simulation of spring and autumn phenology at four US
deciduous broadleaf forest (DBF) sites. The map shows the fraction of US
DBF derived from the Advanced Very High Resolution Radiometer (AVHRR). The
area with > 3 % coverage is the domain for this study. Five
triangles indicate the locations of sites whose long-term measurements of
meteorology and phenology are used for the calibration and/or validation of
the model: Harvard Forest (US-Ha1), Hubbard Brook Forest (US-HB1),
Morgan–Monroe State Forest (US-MMS), University of Michigan Biological
Station Forest (US-UMB), and Missouri Ozark Forest (US-MOz). Phenological
dates are recorded at US-Ha1 and US-HB1 during 1992–2012. Measurements of
leaf area index (LAI) and photos are used to derive phenology at US-UMB and
US-MMS for 1999–2012. Derived phenological dates at US-MOz are used for
model calibration but not validation and are not shown here. At each site,
two simulations are performed with the spring model S9 and autumn model A4 (refer to Fig. 2), driven by temperatures from either the in situ measurements
(blue) or the Modern-Era Retrospective Analysis for Research and Applications (MERRA) reanalysis (green).
Trend of each time series (units in day yr
We use > 75 000 records for deciduous trees to evaluate the temporal variation and spatial distribution of simulated phenology (Table 4 and S1–S2). Data from three out of the four calibration sites (US-Ha1, US-UMB, US-MMS, Table 2 and Fig. 1) are also used for validation; however, we use them in different ways. For calibration, we use the decadal average phenology derived from the multiple-year LAI measurements, so that every calibrated model can capture the spatial pattern of phenology events on the continental scale. For validation, we use year-to-year phenological dates estimated from date records, photos, and LAI at each year, so as to identify the model that best captures the temporal variations. Most of the phenological records are discrete and evaluation of the annual cycle of tree phenology is difficult. Following definitions in earlier literatures (e.g., Zhu et al., 2012; Richardson et al., 2013), we validate spring budburst date (or the onset of growing season, the dates D1 in Fig. S1) and dormancy onset date (or the end of leaf fall period, the dates D2 plus falling length L2 in Fig. S1) from phenology models with the site-level records. The dormancy onset date defined here is based on the canopy level instead of the bud dormancy examined in a recent review paper by Delpierre et al. (2015).
Phenological and climatological parameters for four
deciduous forest sites predicted by segmented regressions (Fig. S1) and the
selected phenology models (S9
The two New England sites, Harvard Forest
(
Ground phenology measurements of deciduous trees used to validate the model.
Data from ground networks were used to evaluate the model performance on the
continental scale. The USA National Phenology Network (USA-NPN) is a
nationwide project collecting standardized ground phenology observations by
researchers, students, and volunteers. The network has limited records
before 2009 but was significantly enriched thereafter. We select observations
during 2011–2012 for 52 deciduous tree species that are most common in the
USA (Table S1 in the Supplement). The derived phenological dates for individual trees are
averaged if they are observed at the same location (see Supplement). We also
used observations from the North American Lilac Network (NALN), which
provides records of the first leaf and first bloom dates of two lilac
species, common lilac (
Dozens of spring phenology models have been evaluated and inter-compared in the past two decades (Chuine et al., 1999; Linkosalo et al., 2008; Vitasse et al., 2011; Fu et al., 2012a, b; Migliavacca et al., 2012; Melaas et al., 2013). These models may have different formats and parameters, but are generally dependent on temperature and photoperiod and can be divided into two categories, spring warming (or one-phase) and chilling (or two-phase), based on their assumptions of how warm and cold temperatures control the phenology development (Migliavacca et al., 2012). Although regional studies have demonstrated that the one-phase models are as efficient as two-phase models at the site level (e.g., Vitasse et al., 2011; Fu et al., 2012a; Migliavacca et al., 2012), we consider that chilling requirement may be necessary for the phenology at the continental and global scales where divergent phenological responses are observed (Zhang et al., 2007; Cook et al., 2012).
The chilling models have different formulations based on the sequences
(sequential, parallel, or alternating) and forms (thermal summation or the
Sarvas function; Sarvas 1972) of chilling and forcing (Chuine et al., 1999). According
to these differences, Migliavacca et al. (2012) summarized and compared
eight models, listed as S1–S8 in Table 5, to fit phenology data at Harvard
Forest. The sequential models require that a chilling threshold (
Summary of phenology models with fit parameters calibrated against the long-term phenology at four US deciduous sites. The detailed parameters for the selected models, S9 and A4, are summarized in Table S3. Optimized parameters for other models are summarized in Table S4.
In a modified alternating scheme (S9), we decrease model complexity by
fixing some parameters based on literature values. First, we fix
We assume the green up process is linearly dependent on forcing
Autumn phenology is more uncertain than budburst because it is affected by
both temperature and photoperiod (Delpierre et al., 2009; Richardson et
al., 2013). Three models have been developed to predict leaf fall with
constraint from temperature and photoperiod, namely the continental
phenology model by White et al. (1997), the growing season index
(GSI) by Jolly et al. (2005) and the cold-degree day
photoperiod-dependent model by Delpierre et al. (2009). The
White et al. (1997) scheme is not compared in this study as it
depends on soil temperature, which is not available at some sites. Jolly
et al. (2005) calculated global phenology as the product of three segmented
functions, which depend on the upper and lower limits in temperature
(
We also construct a simple scheme based on cumulative cold degree days. The
scheme, named “CDD-photoperiod” (A4 in Table 5), calculates cold degree days
(CDD)
We perform both site-level and continental-scale simulations. For stand-alone
simulations (simulation 1), phenology models are driven with daily surface
air temperature sampled at each site (
For the regional simulation (simulation 2), we utilize daily surface air
temperature from MERRA to drive the selected model on a resolution of
1
We perform a sensitivity analysis (simulation 3) to evaluate the uncertainty
due to phenological schemes. In this run, we do not include chilling
constraint for the spring phenology by using a fixed and calibrated forcing
threshold
We analyze species-specific temperature sensitivity of tree phenology at
Harvard Forest (Sect. 4.3). Based on these results, we perform two
additional sensitivity tests to evaluate modeling uncertainties from the
intraspecific variations. In the first run (simulation 4), phenological
parameters are derived based on records of species with the lowest
temperature sensitivity for both spring (sweet birch,
The five sites we select to calibrate and evaluate models are all located in the eastern US, where > 90 % deciduous forests are located (Fig. 1). The site-level evaluations for nine spring models and four autumn models are shown in Fig. 2 and summarized in Table S5. For the spring phenology, the alternating approach (S7–S9) has higher correlations and lower RMSE compared to parallel models (S3–S6). The sequential approach with thermal summation (S1) shows the largest correlations and lowest biases. However, it requires fitting five parameters, increasing its AIC value relative to the alternating models. The three alternating models have comparable correlations and RMSE. However, the modified alternating model (S9) has the lowest AIC, suggesting that fixing some parameters based on literature does not weaken the performance but can reduce model complexity. For the autumn phenology, no models predict correlations higher than 0.5, indicating that missing mechanisms, such as accidental frost, strong wind and rainfall, may be required to improve the current model structures (Richardson et al., 2006; Schuster et al., 2014). The “CDD-photoperiod” scheme (A4) has comparable performance with that from Delpierre et al. (2009; A3) based on correlation and RMSE, and has lower AIC than the latter due to the lower number of fit parameters (Table 5). As a result of the site-level evaluations, we select the spring model S9 and autumn model A4 (parameters listed in Table S3) as the state-of-art schemes for the regional simulations.
Comparison of model performance in the prediction of phenological dates at four US DBF sites among (top) nine spring phenology models and (bottom) four autumn phenology models. The statistical metrics are correlation coefficient, root mean square error (RMSE), and the Akaike information criterion (AIC). Each point represents the mean values of the statistical metrics at four sites for one model. The error bar represents the range of the metrics. Each model uses the optimized parameters as summarized in Table 5 for the prediction. The red ones are the models used for the continental predictions. Detailed predictions at each site are shown in Figs. S4–S11. The values of correlation coefficients, RMSE, and AIC are summarized in Table S5.
Site-level simulations with models S9 and A4 capture both the interannual
variations and temporal trends of phenology at the validation sites (Fig. 1). Sites US-Ha1 and US-HB1 provide > 20 years of phenology
records. The observation–simulation correlations for budburst dates are
0.7–0.8 at these sites. Model performance is poor for autumn phenology, with
correlation coefficients between 0.2–0.4. Both observed and predicted
budburst dates at US-Ha1 show significant advances of
Sites US-UMB and US-MMS have relatively short observations for 1999–2012.
Missing in situ forcing values limit the model's spring phenology performance
compared to that using MERRA reanalysis. With MERRA forcing, the model shows
high correlations (
Comparison of the simulated
Phenology has a distinctive spatial distribution over US deciduous forests
(Fig. 3). Budburst occurs relatively later west of 105
We further evaluate the simulated year-to-year budburst dates with available
long-term records from NALN and USA-NPN networks (Fig. 4). The correlations
between modeled and observed budburst dates are larger than 0.3 for 47 out
of 59 sites, among which 26 are significant (
Correlations (circles) between the predicted budburst
dates and observed first-bloom dates from the North American Lilac Network
(circle) and first-leaf dates from the USA National Phenology Network
(squares). Simulations are performed with the spring model S9. The
correlation coefficients are calculated for individual trees with at least 6
years of observations during 1982–2012. Correlations with
Driven with the MERRA forcing, the model simulates a significant advance of
spring budburst dates in the central eastern USA during 1982–2012 (Fig. 5a).
The largest advance of 0.42 day yr
Trend in the simulated
The spatial pattern of the trend in forest phenology follows spatial
patterns of temperature changes in the past 3 decades (Fig. S14). Both the
reanalysis data and ground records show a significant spring warming of 0.75
Advanced spring and delayed autumn together increased the length of the growing season across the USA (Fig. 6). Relative to the 1980s, the growing season in the 2000s extends by 5.5 days (3.0 %) in the eastern states with dense forest coverage (fraction > 50 %). The model predicts larger extension of 6.4 days (3.9 %) in New England, 7.0 days (3.6 %) in states Illinois and Indiana, and 6.0 days (4.3 %) in the upper Rocky Mountains forests (Fig. 6). This magnitude is comparable to the trend of 2.1–4.2 days per decade in Eurasian and North American temperate forest estimated by other studies (Menzel et al., 2008; Jeong et al., 2011).
The
Most up-to-date estimates of the changes in US forest phenology are performed with remote sensing data. In a recent study, Buitenwerf et al. (2015) found an overall extension of the growing season over boreal and temperate forests during 1981–2012 based on the normalized difference vegetation index (NDVI) from satellite data. However, the exact phenological changes that underlie such overall greening differ among regions. For US forests, the longer growing season is primarily driven by later leaf-off dates, though regional advance of spring is also observed. Our results are generally consistent with their conclusions but with some deviations. For example, they observed later autumn in almost all the eastern US, where we predict delays only in the north and northeast (Fig. 5b). Such discrepancies reflect prediction biases, and may also be a consequence of satellite retrieval uncertainties (Table 1).
We further compare our results to recent reports from the literature, selecting all studies that examine phenological trends across the USA for at least 20 years (Table 1). All selected studies use the NDVI, however, they report different and even opposite trends. Such discrepancies may be attributed to the differences in the definitions of phenological dates (White et al., 2009) or the statistical algorithms in the extraction of the dates (Keenan et al., 2014). Here, we summarize their results on Fig. 7 so as to conclude the most robust changes for US forest phenology in the past 2–3 decades. Since the definition of phenological events varies among different studies (White et al., 2009), we qualitatively compare the simulations with the remote sensing retrievals so as to evaluate the ensemble spatial distribution of phenological changes in the past decades. For spring phenology, four out of seven studies predict advanced budburst or green-up dates in the east, while four predict delayed dates in the north (Fig. 7a). There are no evident phenological changes in the west, northeast, and southeast. Our results show similar changes in spring phenology to the ensemble of the remote sensing studies, except that we predict smaller delays in the northern states (Fig. 5a). In addition, our data-informed model simulates significant spring advances in the central US, while remote sensing studies largely disagree over this area. On the other hand, both the remote sensing studies and our results show that autumn phenology is significantly delayed in the west, north, and northeast (Figs. 5b and 7b). However, the examined studies also exhibit significant delays in the central states, in contrast to our results. In other areas, the trends are insignificant (southeast and east).
Comparison of phenology trend over the USA for
Interannual variations of phenological dates and their
responses to temperature changes during 1992–2011 for each DBF species at
Harvard Forest. The year-to-year
Estimates of trends in phenology are sensitive to the length of the examined
time frame due to relatively large internal climate variability (Badeck
et al., 2004; Iler et al., 2013). Our analyses show that interannual
variations may also cause large uncertainties in the estimated phenology
trend, especially on short decadal timescales. For example, Keenan et al. (2014) estimated a large advance of 0.48 day yr
We perform an additional sensitivity experiment (simulation 3) to examine
the impact of model structure on the phenology prediction. For spring
phenology, model validations have shown that the spring warming (one-phase)
models are as efficient as chilling (two-phase) models (Vitasse et al.,
2011; Fu et al., 2012a; Migliavacca et al., 2012). In simulation 3, we
remove the limit of chilling requirement on the forcing threshold
Our investigation of the roles of chilling and photoperiod is sensitive to the model structure, climate variability, and data availability. First, the similar performance between spring warming and chilling models might also result from the inaccurate representation of chilling/photoperiod mechanisms. For example, the chilling units used in our parameterization are calculated based on daily average temperatures, while Piao et al. (2015) suggested that leaf unfolding dates during 1982–2011 are triggered by daytime more than by nighttime temperature. The up-to-date autumn phenology model fails to capture interannual variability of dormancy onset (Fig. 2), suggesting that unknown processes are involved in the autumn leaf fall (Keenan and Richardson, 2015). It is unclear whether these processes are related to the variations of photoperiod. Second, the decadal changes in temperature may mask the role of chilling. The trend of winter warming is not significant for most areas in the USA (Fig. S14a), suggesting that chilling requirements have been fulfilled in the past 3 decades. However, it is unclear whether the winter warming will intensify in the future, which may slow the advancement of spring budburst. Third, we choose to calibrate the phenological parameterization at the level of PFT because species-specific measurements are usually incomplete in time and uneven in space. Such incompleteness may influence the accuracy of derived decadal phenological records used for both model calibration and validation. At the same time, PFT-level parameterization may be too broad for the vegetation modeling because it fails to capture intraspecific variations (Van Bodegom et al., 2012; Reichstein et al., 2014). Observations at the community level suggest that the budburst of some species is sensitive to autumn/winter and spring warming but with opposite signs (Cook et al., 2012). In the next subsection, we examine the records of 13 deciduous tree species at Harvard Forest.
Tree phenology and its responses to temperature changes have been shown to vary among species (Vitasse et al., 2009; Fu et al., 2012a; Archetti et al., 2013). In this study, however, we calibrate model parameters based on the long-term phenological cycle derived from LAI, which represents the mean growing seasonality averaged among species. We do not perform the species-specific simulation for the following three reasons. First, the species-level measurements are usually not available on the continental scale, which influences both model calibration and validation. Second, species-level modeling increases the complexity and computational costs while decreasing predictive reliability (Prentice et al., 2015). Third, investigations at both site level and continental scale show similar temperature sensitivity of tree phenology between the species-specific and species-aggregation approaches.
We analyze the temperature sensitivity of tree phenology for 13 deciduous broadleaf forest (DBF) species
at Harvard Forest (Fig. 8). We calculate the ensemble phenology based on the
basal area of each species (the dominant species are red oak (
We perform two sensitivity runs to evaluate the modeling uncertainties due
to intraspecific variations at the continental scale (Fig. S18). Simulations
with either the lowest (simulation 4) or the highest (simulation 5)
temperature sensitivity yield very similar phenological trends to that in
the control simulation (simulation 2). In the east, simulation 4 predicts a
spring advance by 0.33 day yr
We performed a model inter-comparison to identify a state-of-art scheme for
predicting tree phenology of US deciduous forests. An extensive database
of ground measurements, including long-term records of phenological events
at the site level and short-term records widely scattered on the national
scale, was compiled to evaluate the models. The selected models with the
lowest AIC values utilized the accumulative temperature summation, with
additional constraints of winter chilling on spring phenology and
photoperiod on autumn phenology. The 30-year phenology trend of US
deciduous forests was explored using the selected models. Consistent with an
ensemble of remote-sensing studies, the continental simulation showed a
significant advance of 0.34 day yr
Uncertainties in phenological predictions originate from drivers,
parameters, and model structures (Migliavacca et al., 2012). In this
study, we minimize uncertainties from meteorological forcings by utilizing
an updated reanalysis product and validate the gridded forcings with
site-based observations. For the model parameters, we calibrate model
parameters with long-term average phenology at four deciduous sites with
diverse spatial distribution. This approach was chosen because a
well-calibrated phenology model based on a single data set may have poor
performance against external data sets (Chuine et al., 1999; Richardson
et al., 2006). The validation shows that the predicted spatial pattern is
reasonable and the long-term average matches observations within sampling
uncertainty (Figs. 3–4). However, due to the data scarcity, all the selected
sites are located in temperate areas ranging from 38 to 46
Our model inter-comparison does not show a distinct advantage for a specific spring model, suggesting that the model formulation, such as sequential, parallel, and alternating, is not a dominant source of uncertainty for estimates of spring phenology. On the other hand, the evaluation of autumn phenology shows that models with cumulative cold summation and photoperiod limits may better capture the trend of the dormancy onset dates. However, the state-of-art autumn models still have large biases in capturing year-to-year variations. Missing mechanisms, potentially including biotic (e.g., tree age: Vitasse, 2013; Caldararu et al., 2014; and species: Vitasse et al., 2009) and abiotic (e.g. water stress: Jones et al., 2014; accidental frost: Schuster et al., 2014; strong wind and air pollution: Gallinat et al., 2015; and timing of spring flushing: Fu et al., 2014; Keenan and Richardson, 2015) factors, may jointly affect leaf fall in a process that is currently not well understood.
Given these uncertainties, our results show a significant advance of 0.34 day yr
Hubbard Brook phenology data were provided by A. Bailey at the USDA Forest Service, Northern Research Station, Hubbard Brook Experimental Forest. This project was supported in part by the facilities and staff of the Yale University Faculty of Arts and Sciences High Performance Computing Center. T. F. Keenan acknowledges funding from a Macquarie University Research Fellowship. Edited by: M. Williams