Discussion
Our results demonstrate that the intrinsic limitations of phenological
data derived from herbarium collections—assuming other forms of bias
are not pervasive—do not preclude the development of accurate
phenoclimate models capable of predicting the timing of population-level
flowering onset or termination, and are only slightly less accurate than
predictions of median flowering date. Further, the accuracies of these
models are not likely to be closely tied to the magnitude of
phenological variation among individuals of a species, and can be
produced with similar quantities of data as more traditional models of
mean flowering phenology (Park and Mazer 2018). However, this study does
identify several limitations to the prediction of population-level
flowering onset and termination DOYs from herbarium data that may impact
the reliability of such predictions.
First, our simulations demonstrate that the accuracy of specimen-derived
phenoclimate models can be highly sensitive to biases in collection
timing within populations (Fig. 4). The frequency with which such biases
occur is not well documented, although they are more likely to be
problematic among species that flower at the beginning or end of the
local growing season in temperate climates, as collection activity is
frequently lower during inter and other unfavorable conditions (Daru et
al. 2017). Additionally, collection activity may be reduced during the
early portion of the flowering display among the earliest-flowering
species, as relatively few species in a given location or region are
likely to be vegetatively or reproductively active during those periods.
Similarly, collection activity may be low during the later portions of
the flowering periods of some late-flowering species that flower after
the majority of species have ceased flowering or gone dormant. Thus,
predicted timings should be viewed with greater caution when modeling
the timing of flowering onset or termination among early spring or late
fall-flowering species.
Second, model predictions will likely be less accurate among
long-flowering species, as longer individual flowering durations were
consistently associated with lower model accuracy across simulated taxa.
Long flowering durations also amplify the deleterious effects of biases
towards collection of specimens from specific portions of individual
bloom displays, as longer individual flowering periods necessarily
increase the magnitude of temporal bias that can be introduced by
collector preference towards recently opened or nearly completed
flowers. Fortuitously, herbarium specimens most frequently have been
documented to exhibit biases towards collection proximate to peak
flowering DOY (Primack et al. 2004, Davis et al. 2015, Panchen et al.
2019), which notably produced more accurate phenoclimate models of both
flowering onset and termination than unbiased collections, particularly
among long-flowering species. Thus, for species that exhibit charismatic
or notable peaks in their individual flowering displays, collector
biases may actually improve rather than hinder phenoclimate modeling
conducted using these methods.
Third, and finally, our study assumes that the climate stimuli to which
species exhibit phenological responses can be sufficiently captured by
available climate data to drive such models; thus, the magnitudes of
error presented here should also not be taken to represent expected
model accuracy when predicting phenological timings of real plant taxa,
as the simulations presented here corresponded to an ideal situation in
which all among-population variation in phenological timing could be
explained by a single climate variable. Under real-world conditions, we
may expect that some component of phenological variation will be
explained by aspects of local conditions that cannot be easily captured
using available climate data. Thus, our study demonstrates that
specimen-based phenological snapshots enable estimation of
population-level onsets and terminations despite noise and biases in the
timing of collection, but the accuracy of herbarium-based predictions in
actual plant populations will likely depend on i) the degree to which
available climate data capture the drivers of its phenological variation
over space and time, and ii) whether the most relevant climate factors
have been identified and incorporated into phenoclimatic models.
Consequently, phenological predictions of species that exhibit highly
stochastic phenological timing or are highly sensitive to variation in
aspects of the local environment that are not easily captured using
broad-scale gridded data are likely to be less accurate regardless of
what aspects of a given phenophase are being assessed. Similarly,
species that s exhibit spatial biases towards collection solely in
specific habitats or broad seasonal biases in collection effort are
likely to be less accurate. However, as many studies have indicated that
strong linear phenological responses can be captured from monthly,
seasonal, or annual temperature at moderate spatial resolutions
(Miller-Rushing et al. 2006, Gerst et al. 2017, Park and Mazer 2018),
this is unlikely to represent a major obstacle in modelling the
phenology of most plant species in temperate environments.