Seasonally recurring phenological events, such as animal migration and budburst or flowering of plants, have been shown to be highly sensitive indicators of the biological impacts of climate change. Based on previous studies, the IPCC’s AR4 estimated an overall trend towards earlier spring of between 2.3 and 5.2 days per decade since the 1970s. However, future phenological responses to climate change, and the associated effects on ecosystem health and biogeochemical cycling, remain highly uncertain; in particular, the uncertainties themselves have not been rigorously quantified. With this in mind, I present the results of a data-model fusion experiment that leverages a unique, 20-year phenological dataset of direct observations of budburst and flowering of a wide range of native woody species growing at the Harvard Forest in central Massachusetts.
My objective is to address the question of uncertainty in phenological forecasts. Using Monte Carlo techniques, I parameterize a range of standard phenological spring onset models for nine forest tree species. Models are evaluated against the data using both cross-validation approaches and a standard, information-theoretic model selection criterion. For each model, an ensemble of acceptable parameter sets—characterizing the posterior parameter distributions—are generated by exploring parameter space and identifying those combinations that yield model output consistent with the observational data, based on a chi-squared test at 95% confidence. Running the models forward in time (through 2050, using climate projection scenarios available from the archive of IPCC model results) with this ensemble provides a suite of probabilistic forecasts, with full characterization of prediction uncertainties.
Results/Conclusions
Differences among species in the model judged to be “best”, conditional on the observational data, show that the mechanisms controlling spring budburst vary among species. Akaike’s Information Criterion indicates good support for some common models, but little support for others. For individual models, species-specific parameterizations reveal highly variable sensitivities (across species) to temperature. Estimates of uncertainty related to model structure and model parameterization will be presented, based on variability across models and across the posterior parameter distributions, respectively. Specifically I will focus on how these uncertainties differ among species, and investigate the future conditions under which uncertainties are largest. I will conclude by discussing some ways in which the representation of phenology in the land surface component of coupled carbon-climate models could be improved.
