Determinants of Predictability in Multi-decadal Forest Community and Carbon Dynamics

Predictions from ecological models necessarily include five different uncertainties: demographic stochasticity, initial conditions, external forcing (i.e., drivers/covariates), parameters, and modeled processes. However, most predictions from process-based ecological models only account for a subset of these uncertainties (e.g. only demographic stochasticity). This underestimation of uncertainty runs the risk of producing precise, but inaccurate predictions. To address these limitations, we created a new generalizable ensemble state data assimilation algorithm that accommodates two common features of ecological data, zero-truncation and zero-inflation, and allows the estimation of process error and its covariance among multiple ecological variables. We then demonstrate the use of this novel algorithm by assimilating 50 years of tree-ring-estimated species-level biomass at Harvard Forest into a process-based forest gap model. Finally, we partitioned the variance in this hindcast to test long-standing assumptions in the ecological modeling community about which uncertainties dominate our ability to forecast forest community and carbon dynamics. Contrary to >40 years of research relying on stochastic forest gap models, we found that demographic stochasticity alone massively underestimated forecast uncertainty (0.09% of the total) and resulted in overconfident and biased model predictions. Similarly, despite decades of reliance on unconstrained “spin ups” to initialize models, constraining initial conditions with data led to the largest increases in prediction accuracy. Counter to conventional wisdom from modeling other Earth system process, initial condition uncertainty declined very little over the forecast time period. Process variance, which heretofore had been difficult to estimate in mechanistic ecosystem model projections, dominated the prediction uncertainty over the forecast time period (49.1%), followed by meteorological uncertainty (32.5%). Parameter uncertainty, which has recently been the focus of the modeling community, contributed a modest 18.3%. These findings call into question much of our conventional wisdom about how to improve forest community and carbon cycle projections on multi-decadal to centennial time scales. This foundation can be used to test long standing modeling assumptions across fields in global change biology and suggests a fairly significant reorientation of the modeling community toward better initialization of models with current observations and efforts to better quantify, propagate, and reduce process error. These approaches have the potential to improve both the accuracy of ecological forecasts and our understanding of the predictability of ecological processes.