Is photoperiod a dominant driver of secondary growth resumption?

In their recent paper, Huang et al. (1) have amassed a fascinating time-series of xylem tissue formation across 826 individual trees, spanning 21 species and 79 Northern Hemisphere locations. Like many other aspects of plant growth, wood formation shows strong seasonality, but very little is known about the environmental triggers for seasonal growth resumption. With wood absorbing ∼15% of anthropogenic CO₂ emissions annually, this extensive dataset provides an exciting opportunity to refine earth system models, but it does not provide compelling evidence that photoperiod is the dominant season cue.

The crux of Huang et al.’s conclusions are derived from their observed positive relationship between the day of year of wood formation (DOY_wf) and the photoperiod on the day of year of wood formation (Photoperiod_wf), using an analysis that relies heavily on R² values (ref. 1, figures 3, S5, and S7). The authors used the response variable (DOY_wf) to calculate several of their predictor variables: Photoperiod_wf, cumulative forcing (FU_wf), and chilling (CU_wf) at DOY_wf. There are two major problems with this approach. First, it can lead to regression models that appear to fit exceptionally well for the wrong reasons. Where photoperiod covaries positively with day of year—as it does over the Northern Hemisphere in springtime—a positive relationship between DOY_wf and photoperiod on that day will always be apparent in individual sites. Similarly, FU_doy and CU_doy increase monotonically over the season, although less linearly than Photoperiod_doy. Second, models that require knowledge of the response variables to calculate the predictor variables cannot be used in forecasting.

To identify how the seasonality of photoperiod could influence results, we simulated phenological events from a process depending only on temperature and latitude (not photoperiod). Our simulations used the sites, years, and mean annual temperatures (MATs) from the dataset in ref. 1 (Fig. 1A). Expected relationships between MAT and DOY_wf are apparent (Fig. 1 C and D); however, photoperiod is a stronger predictor of phenology (Fig. 1 E and F), although it played no part in simulations. This trend is true among sites and years at a single latitude (Fig. 1 C and E) as well as across latitudes (Fig. 1 D and F).

Fig. 1.

Simulated phenology data, created using temperature and latitude, nonetheless appear to be strongly predicted by photoperiod, due to underlying seasonal trends. Phenology data were simulated using forcing unit (FU) thresholds identified in Huang et al. (1) (A). Dots show simulated dates of onset at a latitude of 48°; gray lines represent seasonal forcing accumulation across sites and years. Our simulated data varied forcing units required for onset of wood formation (FU) with latitude, matching trends in Huang et al. (1) (B). These simulations result in a significant relationship between MAT and onset date (DOY_wf ) across multiple sites with similar latitudes (∼48°; C) and across all latitudes in Huang et al. (1) (D), as expected given how they were generated. Although the simulated DOY_wf data were generated using only temperature (specifically, forcing), the relationship is even stronger between photoperiod and DOY_wf both with sites at the same latitude (E) and across latitudes (F). Ponts show dates of onset, with colors indicating latitude (legend in F). Lines in F show continuous patterns in photoperiod across the spring.

Second, we reanalyzed Huang et al.’s phenology data to test the roles of each of MAT, forcing, chilling, latitude, and day length in improving predictions of the timing of xylogenesis in new sites or years using climate data from ref. 2. Our spatial comparison supports the hypothesis that MAT cues xylogenesis timing—at least over broad geographic gradients. However, the temporal comparison shows a weak relationship between interannual temperature variability and xylogenesis timing, suggesting on a yearly basis plants may rely on cues beyond average temperature (Fig. 2). Neither analysis suggests a major role for photoperiod, forcing, or chilling alone, although interactive roles are possible (3). Disentangling the roles of photoperiod and temperature is challenging in the absence of experimental data (4). Nonetheless, we strongly caution against using the authors’ results in a forecasting framework.

Fig. 2.

Predictive performance of DOY_wf based on forcing, latitude, MAT, and photoperiod. We fit linear mixed models on randomly sampled subsets of the data, in each iteration dropping either 10% of sites (to evaluate the importance of environmental variables over spatial gradients in driving phenology; A) or a single year (for each site that included multiple years of data, to evaluate the importance of environmental variation in driving interannual variation in phenology; B). For latitude and MAT, we modeled the DOY_wf directly (e.g., DOY_wf ∼ MAT + random effects), yielding a prediction of DOY_wf for the evaluation subset. For forcing, chilling, and photoperiod, we used the threshold value of the environmental variable at xylogenesis as the response variable (e.g., FU_wf ∼ intercept + random effects), and then used daily meteorological data to translate the modeled thresholds into a predicted DOY_wf (e.g., the first DOY where FU_doy exceeds the predicted FU_wf for a given site, species, and year). Model predictive performance was summarized as the difference in predictive performance (root-mean-square error [RMSE] of predicted vs. observed DOY_wf) vis-à-vis the predictive performance of a null model with no environmental drivers for each of 100 train/evaluation subsets. A difference of 0 indicates the environmental drivers explain no additional variation beyond that of site and species identity; negative δ_{rmse(model-null)} values indicate better predictive performance than the null model, and positive indicate worse. The chilling model yielded far worse RMSEs than the null model (temporal mean δ_rmse = 102, 95% CI = 88–114; spatial mean δ_rmse = 100, 95% CI = 80–121).

Detailed methods along with all code can be found in Github at https://github.com/scelmendorf/wood_formation_phenology.