Predicting forest mortality and dieback is critical for understanding the impacts of climate change. However, this is very challenging at large scales due to the many causes and factors involved in tree mortality. Recent advances in vegetation models that include physical representations of plant hydraulics and stomatal control based on optimization theory allow predicting drought-induced tree mortality based on the degree of damage the vascular system experiences during a growing season. Mortality thresholds based on the percent loss of whole plant hydraulic conductance (PLC) have been validated under controlled experiments in which all plant traits required for parameterizing the model and environmental conditions were well characterized and measured. However, can we use PLC for predicting forest mortality at larger scales? The main challenges are that the trait data usually used are species mean traits and that gridded weather data may be to coarse to capture microclimate variability. Thus, we tested the potential of using a plant hydraulics model to predict the observed mortality in USA Forest Inventory and Analysis (FIA) dataset. We ran simulations for monospecific FIA plots of 14 widespread tree species across USA. We fitted multivariate logistic regression models to predict mortality both as percent dead basal area and as presence or absence of mortality. We used stand properties, weather variables and model outputs as predictive variables and used model selection techniques to find the best fit models.
Results/Conclusions
Logistic models performed worse for predicting percent basal area mortality than the binomial response of presence or absence of mortality at a plot. Logistic models that only included stand properties and weather variables were worse at predicting tree mortality than those that included plant hydraulic outputs such as PLC as a predictor. Best fitting logistic models were capable of explaining between 5-20 % of mortality. For most species, including plant hydraulic model outputs improved model predictions 5-10 %. These results are very encouraging since it shows that despite using a mean value per species for key plant traits that are vary variable (e.g., rooting depth, maximum conductivity, and tree leaf area to basal area) and a quarter degree downscaled hourly weather dataset, the plant hydraulics model simulation outputs were capable of improving mortality predictions by capturing environmental-biosphere interactions. The utilization of mechanistic plant hydraulic model outputs together with statistical logistic regression models improves predictions of the vulnerability of forests to climate change.