Warm temperatures accelerate dead tree turnover

The most important driver of variation in snag fall rates, temperature, implies that climate change will dramatically alter deadwood structure and function. The strong effect of temperature on snag fall has seldom emerged from previous analyses, presumably because most have focused on snag dynamics in specific regions where temperature gradients are less dramatic [28–31,38,39]. At a continental scale, differences in overall snag fall rates between cooler Canadian forests (5.4% per year [11]) and warmer eastern United States forests (11.9% per year, this study) are consistent with a strong effect for temperature but confounded by differences in species composition and stand structure. After statistically controlling for the effects of other important drivers (such as species and wood properties), we substituted the effect of temperature across space for the effect of warming through time to predict that a model northern hardwood forest could lose more than 20% of snag C by mid-century. Considering that snags are a keystone structure for many endangered species, decreased snag lifespans could compound stresses faced by some threatened species [41], while simultaneously enhancing habitat for species that depend on recently fallen logs. Changes in species composition for deadwood associated organisms could affect forest biodiversity management plans. Moreover, snag fall rates are a major control on the timing of changes in net C balance following disturbance, including diebacks associated with increasing temperature [4,7,17]. We focused just on the effects of snag fall on the C cycle; however, warmer forests likely exhibit higher tree mortality, faster C emissions from both standing and down deadwood, and a faster transition from net C sinks to sources (Fig 1). Reduced C residence time in cooler forests that store C in snags today represents a quantifiable prediction for differences in C turnover among biomes [42,43].

The strong effect of temperature on snag fall contrasted with the weak effect of average wind speed, suggesting that variation in exposure to chronic mechanical stress poorly explains snag fall rates in eastern North America. This result contrasted with an analysis of drought-killed aspen (Populus tremuloides) in northwestern Canada, where snag fall rates were significantly higher in locations with faster wind speeds [28]. Snag sensitivity to wind throw may be species specific [44,45] or more strongly influenced by gusts. While we cannot directly quantify the effects of specific localized storms in our dataset, we note that blowdown events could contribute to the large variation in snag fall rates attributed to unmeasured spatially correlated processes. Specifically, elevated snag fall rates in the upper Midwestern United States could reflect the frequent incidence of strong convective storms in that region [46]. However, snag fall rates were not higher along the Gulf Coast where hurricanes generated a measurable influx of deadwood during the study interval [47]. Presumably, new down deadwood along the Gulf came directly from living trees that had blown over, raising the possibility of a distinct forest climate feedback involving stronger storms, higher mortality, and faster decay [17,48]. Living trees that fall because of blowdown, landslides, or other catastrophic events are likely to influence forest C cycling in important ways. More detailed analyses are necessary to identify what factors determine differences in blowdown rates of living and standing dead trees during storms.

Wood durability accurately predicts snag fall

Snag dynamics also depended on species’ wood durability. Wood durability contrasts with more widely known wood mechanical properties, such as hardness and modulus of rupture, in that wood durability is only weakly correlated with wood density and measures long-term rather than initial wood strength [49]. For example, a common standard ranks species’ wood durability by comparing the useful lifespan of heartwood to that of sapwood from a weakly defended reference species [50]. Builders can use this information to predict the service lifespan of untreated outdoor wooden structures used as construction materials. In temperate settings, the difference between the first and second wood durability categories roughly corresponds to a 5-year difference in expected service lifetime [51]. In our model, the same difference in species wood durability corresponded to a 4.5-year difference in snag half-life. This close correspondence between the dynamics of buildings and standing dead trees highlights the ecological importance of this understudied wood functional trait. Larger wood durability datasets, involving more tree species and alternative ways of incorporating uncertainty between sources, may inform other important ecological phenomena.

The biological basis for variation in wood durability is largely unknown, but could reflect decay inhibition by secondary compounds that pigment heartwood and reduce wood permeability [36,49,52]. To the extent that wood durability corresponds to overall decay resistance, snags that fall quickly may decompose faster as logs [35]. Furthermore, wood durability might also predict resistance to internal decay in living trees, identifying species that are likely to be hollow and more prone to falling while alive [53]. For both of these reasons, species with less durable wood could be targets for interventions to enhance C storage and manage risks of hazard tree failure. Even species that produce durable wood currently may produce less decay resistant wood in the future. CO₂ fertilization tends to accelerate tree growth, and fast growing trees tend to produce less durable wood [34,35]. In a recent wood decay experiment involving both plantation-raised and old growth scots pine, less durable wood decomposed more quickly across a range of temperatures and fungal strains [37]. Faster growth with higher CO₂ raises the possibility that trees may produce less durable wood, so that they decay and fall more quickly (Fig 1). Lower concentrations of decay-limiting extractives in more recently deposited heartwood would support this conclusion.

While wood durability was an important predictor of snag fall, the large residual interspecific variation (Fig 4) indicates that other unmeasured species traits also play important roles determining whether snags break or uproot. For example, in a tropical dry forest, interspecific differences in tree fall rates were associated with tree height, rooting depth, and wood density [44]. Surprisingly, intact wood density was not a significant predictor for snag fall in eastern North American forests, indicating that standing dead tree failure depends less on initial wood mechanical strength and more on how strength changes with decay [19]. However, the range of wood density in the temperate zone is much narrower than what occurs in the tropics [52], raising the possibility that tropical forests show more variation in snag fall rates driven by increased variation in wood density.

Effects associated with breaking forces weaken with decay

Significant effects for stem diameter and stand density indicate that exposure to breaking forces also influences snag dynamics. Large stem diameters, which can slow wood decay and increase mechanical resistance to breaking forces [9], were associated with slower rates of snag fall. The slow fall rate for larger snags may moderate some effects of climate change on forest C cycling. Large trees store more C [54] and may be more sensitive to mortality during drought [55]. Yet, as dead trees, they tend to stand longer, providing more time for proactive management strategies like salvage harvest while their wood remains intact. Large, relatively persistent snags may also preserve habitat for some important wildlife species [1].

While tree size influences both wood decay and toughness, stand density more directly affects exposure to breaking forces. We found that snags in denser stands showed slightly slower falling rates. This result is consistent with the expectation that dense stands shield snags from breaking forces and could indicate that future forests with lower stem densities may have higher rates of snag fall. However, in dense forests affected by dieback, the protective effect of stand density could be offset by accelerated decay with increasing abundance of wood decay fungi and wood boring beetles [28] and the possibility for a domino effect [27]. Nevertheless, the effect of stand density, like the effects of diameter and temperature, declined with advancing decay class, which suggests that wood decay eventually overrides the effects of other factors.

Wood decay, forest disturbances, and distinctive snag legacies

Ultimately, integrating snag dynamics into risk projections for changing forests requires understanding how deadwood decay and toughness vary with different mechanisms of tree mortality. Harvest, drought, pest outbreaks, and fires promoted forest C loss at the same order of magnitude as fossil fuel consumption in western North America [56]. While the relative roles of drivers vary, harvest is the biggest single driver of global forest C loss [57]. In northern forests, experimental harvests raised soil temperatures, driving faster wood decay [58] and accelerated fall rates for both naturally recruited and artificial snags [32]. Both experimental results in managed forests are consistent with our analysis of snag dynamics in unmanaged forests across the eastern United States, where increased temperature and decreased stand density both accelerated tree fall. Whatever standing deadwood remains in managed locations is likely to fall and decay more quickly. Identifying whether these connections extend to recreational or urban forests has major implications for mitigating hazard tree risks.

While faster wood decay and snag fall might accelerate C emissions after harvest, forest dieback during drought may influence snag dynamics differently. Wood decay is strongly moisture limited [9]. Consequently, prolonged drought is likely to slow wood decay and snag fall. While drought should slow wood decay and snag fall, in some cases, drought is not always the proximate cause for tree mortality. Often, drought and heat stress interact to trigger pest outbreaks or fires [42]. These change agents can have distinctive impacts on snag fall and C dynamics [4,59]. Wood boring beetles can structurally compromise snags and spread wood decaying microorganisms, increasing rates of snag fall [60]. Fire may play an even more complicated role. Intense forest fires partially consume existing snags while creating new snags by killing trees. Charred snags have different exposure to breaking forces and decay resistance. In California, snags in burned stands fell faster due to a combination of structural weakening and increased exposure to wind [31]. Both of these results are consistent with our analyses of snag dynamics in unburned stands in the eastern United States. While burning is likely to reduce snag C, residual charred wood decays extremely slowly, which can enhance stabilized soil C depending on fire return intervals. In Oregon, reburned stands showed persistent differences in deadwood structure [39] although fire return interval had relatively little impact on snag abundance during an experiment in southern Florida [61].

Overall, warm temperatures and faster wood decay accelerate standing dead tree fall. Associated reductions in snag C residence time may increase the sensitivity of forests to climate forcing, especially if higher CO₂ reduces tree species wood durability. Forests with fewer snags and more logs may have different habitat value for deadwood-dependent species. Where people use forests, risks associated with snag fall may change as well. Improving forest resource management will require a better understanding of how snag fall rates interact with other agents of forest disturbance, especially strong storms and fires, as well as a thorough examination of potential connections between the drivers of snag fall in unmanaged forests with living tree fall rates including those in urbanized areas.

Response variable: Snag fall

We evaluated the persistence of standing dead trees in permanent forest inventory plots resurveyed by the USFS Forest Inventory and Analysis (FIA) Program [62]. In 1999, FIA began recording the presence of snags, defined as main boles of stems at least 12.7 cm in diameter at 1.37 m height (diameter at breast height or DBH) and leaning <45° from vertical. Annual resurveys of 20% of the FIA plots yield an average resurvey interval of 5 years. We queried the FIA database for all resurveyed and unmanaged plots that included one or more snags during the initial survey. Standing dead trees are tracked through time by the FIA program until they no longer meet their associated definition (fallen or decayed/fractured below minimum size thresholds). Given the importance of identifying the distinctly different live and dead tree populations across the US for national resource assessments, care is given to accurately identify trees as they transition from live, to dead, to down dead via the field measurement variables of tree status and reconciliation codes [63]. As such, when the FIA field records identified individual snags as no longer qualifying as a standing dead tree on plots where there were no recent forest management activities (i.e., logging equipment knocking over dead trees), our study considered those snags as fallen. We limited our analysis to plots that had been revisited twice between 2000, when the new nationally consistent criteria for snag inclusion were adopted, and 2010, when most plots had been resurveyed.

Predictors of snag fall

We estimated climatic conditions at the reported location for each plot, which, for privacy reasons, is within 1 km of the actual plot location. We estimated average wind speed (WND) at 10 m altitude during the interval from 2000–2010 based on the National Atmospheric North American Regional Reanalaysis [64], which provides data at a resolution of approximately 32 km². We estimated average annual temperature (AT) over the 30-year interval from 1981 to 2010 based on PRISM, which estimates temperatures at a resolution of 800 m² [65].

For quantifying effects of factors that vary at the landscape, stand, and individual scales, we obtained additional covariates based on measurements taken during the FIA surveys. For topography, we included subplot slope (SL) and physiographic class. FIA defines physiographic class using a cross classification of forest habitats by hydrology and landform (S1 Table). At the stand scale, we included both average size and density of living trees. For an estimate of average size, we calculated quadratic mean diameter (QMD) of all living trees over 12.7 cm DBH. To estimate density, we calculated the number of living trees per hectare (TPH) over 12.7 cm DBH. At the scale of each individual snag, we included both DBH (DIA) and initial decay class (DC). FIA assigns snags to one of five progressive decay classes based on external conditions [62].

To quantify species-level variation in wood traits, we compiled two additional datasets. We drew our estimates of species-level intact wood density (DEN) from Miles & Smith [66]. We also compiled information on tree species wood durability (DUR S1 Dataset). Wood durability describes how untreated wood resists degradation by microbes and insects using an ordinal scale. This scale has a direct quantitative interpretation under certain materials standards. For example, the European standard [50] compares the lifespan of heartwood stakes to that of Pinus sylvestris sapwood. Sapwood stakes from Pinus sylvestris resist impact breakage tests for about 5 years when suspended above ground in temperate conditions [51] defining the lowest reference class, “Not durable”. The next class, “Slightly durable” lasts between 1.2 and 2x longer, “Moderately durable” wood lasts between 2 and 3x longer, “Durable” wood lasts between 3 and 5x longer, and “Very durable” wood lasts more than 5x longer. This standard is similar to the discontinued United States standard [67]. We compiled descriptions of tree species wood durability from several sources including the Forest Products Lab Wood Handbook [34] and Wood Properties Techsheets [68], the Natural Durability of Wood : a Worldwide Checklist of Species [49], the American Hardwood Export Council Guide to Species [69], the Wood Database [70], the USDA Fire Effects Information System [71], the Forest Trees of Florida Handbook [72] and the Utah State University Tree Browser [73]. For maximum consistency with the materials standards terminology, we defined the least durable classification within a source (e.g. “not durable”, “non-resistant”) as durability 0, the next category (e.g. “slightly durable to perishable”) as durability 1, the next highest (e.g. “moderately durable”) as durability 2, the second highest (e.g. “durable” or “resistant”) to durability 3 and the highest category (e.g. “very durable” or “very resistant”) to durability 4. For species described using a range of descriptors within a source (e.g. moderately to very durable), we scored them to intermediate categories (i.e. durability 3). We resolved variation among sources by taking the arithmetic mean for each species with more than one estimate.

Model structure

Our goal was to identify the factors that most strongly influence the rate at which snags fall. We did so by estimating the probability, p_ijkl, that any individual snag i (i = 1, 2, …, n = 99,213), remains standing (S_ijkl = 0 or 1) at the end of a census interval of length t (years) in a plot with a specified physiographic class j (j = 1, 2, …, m = 16), spatial location k (k = 1, 2, …, p = 151), and that had been previously identified as species l (l = 1, 2, …, q = 205). We assumed S_ijkl represents a single Bernoulli trial: (1) We then analyzed the annualized probability of standing p_ijkl as a linear function of covariates in a logistic regression framework: (2) t_ijkl is a row vector of continuous predictors and ρ is a vector of associated effects. We covariate-centered the continuous predictors (t’s) by subtracting their mean such that β_o is the annualized baseline log odds of remaining standing at “average” covariate conditions. We evaluated many candidate models including effects for tree DBH, stand QMD and TPH, plot SL, WND and AT (S1 Appendix). In addition to continuous covariate effects, u_j is the effect of physiographic class j, v_k is the effect of location k, and w_l is the effect of species l.

To account for variation among physiographic classes, we modelled their effects as coming from a population of potential effects: (3) The effects (u_j’s) are constrained to sum to zero relative to the baseline persistence rate, β_o, and is the variance among physiographic class effects.

To model the spatially structured location effects, we assigned each location to a 1.7° x 1.7° grid consisting of k = 1, 2, …, p = 151 cells. This grid resolution was small enough to capture coarse spatial variation while being sufficiently large to include multiple plots in most cells. We then estimated spatially structured effects as a Gaussian conditional autoregressive process [74], such that for grid cell k ≠ h: (4) δ_k defines the cells that are spatially adjacent to the focal cell k, a_kh is a binary operator that equals 1 only if cell h is adjacent to k, a_k+ is the number of cells adjoining k, and is the variance of the conditional Gaussian distribution. As with our estimates of physiographic class effects, we constrained these effects (v_k’s) to sum to zero so that the spatial autoregressive effects represent local deviations from the baseline persistence rate, β_o.

For modelling species-level effects, we incorporated known differences in species’ wood traits using a hierarchical approach: (5) μ_l is the expected species-specific deviation from the baseline persistence probability, β_o, and is the variance that describes variability among species. We constrained the species effects (w_l’s) to sum to zero and modeled the expected species-specific deviations (μ_l) as a function of wood durability (DUR): (6) β_DUR is the effect of wood durability on the probability of snag persistence, and is the average wood durability index, computed across all species for which DUR estimates were available.

Several species lacked DUR estimates. Rather than exclude them, we imputed their DUR values based on DUR estimates available for other species by assuming a hierarchical model for DUR based on a coarse taxonomy [33] whereby species are nested in family: (7) Where the logit transformation (see Eq 2) and rescaling projects the [0–4] interval-constrained DUR variable to the real number line; f(l) denotes family f associated with species l;μ_f is the family-level logit-transformed DUR value; and is the variance that describes variability in logit-transformed DUR among species within families. Then we modeled the family-level means as nested in division (i.e. gymnosperm versus angiosperm): (8) d(f) denotes division d associated with family f; μ_d is the logit-transformed DUR for either angiosperms or gymnosperms; and is the variance that describes variability in logit-transformed DUR among families within divisions.

Model implementation and adequacy

We fit the above model in a Bayesian framework, with diffuse normal priors, Normal(0,1000), for all of the continuous effect parameters (i.e., each element of ρ in Eq (2) and β_DUR in Eq 6) as well as μ_d in Eq (8). For the baseline fall rate parameter, β_o, we used a flat prior. We specified broad uniform priors, Uniform(0,100), for the standard deviation terms associated with each variance component (i.e. σ_P, σ_S, σ_f, σ_d). Finally, following [46], we specified a Gamma(0.5, 2) prior for the precision of the conditional Gaussian distribution for the spatially structured effects (i.e. , Eq (4)). Our use of minimally informative priors allows the data to largely inform the posterior distribution. Furthermore, no other studies have employed Bayesian techniques to estimate snag fall rates so informative prior distributions are lacking.

We implemented the above model using OpenBUGS, V3.2.1 rev 781 [75] (S2 Appendix). OpenBUGS employs Markov Chain Monte Carlo sampling for model parameters from the joint posterior distribution. We ran five parallel chains for at least 110000 iterations per chain. We assessed convergence of the chains to the posterior when the upper limit of the 95% CI Brooks-Gelman-Rubin (BGR) statistic [76] fell below 1.05. We then discarded the first 10000 samples as burn-in and thinned each chain by a factor of 40 to reduce within-chain autocorrelation. We repeated the process for the full dataset and for subsets of the original dataset representing each decay class (S1 Appendix).

We assessed model adequacy by comparing observed to expected snag persistence given the hierarchical model and posterior mean parameter estimates. For every observation S_ijkl, we rounded model-based expected values for p_ijkl^t to either 0 or 1. Then we calculated both the proportion of correct predictions of S_ijkl and the 95% confidence interval for the odds ratio of observed versus predicted values using the function “fisher.test” in R v 3.2.5 (package “stats”). We also plotted the expected versus observed counts at every level of the three categorical predictors (species, spatial grid and physiographic class).

Parameter interpretation

For continuous predictors (e.g., ρ in Eq (2) or β_DUR in Eq (6)), we computed a standardized effect size by dividing the posterior mean and 95% CI limits by the standard deviation of the corresponding covariate data. This provides an estimate for the relative effect of a standard change in the predictor near the mean of the data on the logit scale. Because we imputed wood durability (DUR) for some species, our estimate of the standardized effect size for this variable was averaged over 5000 draws from the posterior distribution for the missing DUR estimates. For the relative influence of categorical predictors, we also report the posterior mean and 95% CI for the standard deviation of the random effects (i.e., σ_P, σ_S, and σ_G).

To relate the aforementioned effects to expected changes in rates of snag fall, we computed average predictive comparisons [77,78]. This approach allows us to compare the difference (Δ) in the response of interest (y)—which is related to different indices of rates of snag persistence—with a specified change (ξ) in a predictor variable of interest (x) as: (9) r denotes an individual tree, as indexed by all observed unique combinations of i, j, k, and l (see Eq 1); θ represents a vector of estimated parameters (i.e., includes ρ, β_o, β_DUR, and all σ terms); and X is a matrix of all other predictors excluding the focal predictor, x (X and x involve different combinations of t in Eq (1) and DUR in Eq (6)). The specified values for ξ for each variable were chosen as reasonable estimates for the direction and magnitude of change in eastern North American forests by mid-century. For mean annual temperature in 2050, we used the estimated 2.4°C increase from the CMIP3 model projection for the RCP8.5 emissions scenario [79]. For diameter (DBH), we used the mean diameter growth for eastern North American forests in a 35 year interval as estimated by [80] to infer the associated change in DBH for a “typical” tree. For mid-century TPH, we assumed that most forests would lose trees due to stand thinning and mortality. We note that the proposed values for ξ are simply to put effects on a biologically relevant scale and that they are considered one at a time, in an additive fashion.

We expressed the difference in expected rates of snag persistence in two different ways. First we computed the average predictive difference in the annualized probability of remaining standing such that: (10) where p is the annualized probability of a snag persisting, which is predicted from Eq (2). Then, for snags in the earliest DC, we also computed the average predictive difference in half-life in years as: (11) where Pr(S|x,θ,X) is described in Eq (10). To incorporate parameter uncertainty, we report distributions for these average predictive differences based on 5000 draws from the posterior distributions of the parameters.

Finally, we examined the effects of projected changes in snag persistence on forest structure and function in a simple, spatially implicit, discrete time forest C model [81]. We modified the model by specifying separate detritus pools for litter, snags, and logs. We then parameterized the model to approximate the C budget of a well-characterized north temperate hardwood forest, Hubbard Brook Experimental Forest, using internal transitions estimated by [82], along with snag persistence times estimated from our model, and decay rates for litter, snags, and logs based on literature information [9,83,84]. We evaluated steady state snag C and maximum net ecosystem productivity (NEP) under contemporary conditions relative to those projected for 2050 based only on a change in snag persistence time due to projected temperature change. This approach estimates the sensitivity of snag C and NEP to snag fall rates per se without including changes in decay rate or potential feedbacks mediated by climate (e.g. Fig 1). We assessed uncertainty in our sensitivity estimate by repeating the simulation for 1000 parameter estimates drawn from the distribution of predictive comparisons. Model code is available in S3 Appendix.