Towards understanding predictability in ecology: A forest gap model case study

Abstract

AbstractUnderestimation of uncertainty in ecology runs the risk of producing precise, but inaccurate predictions. Most predictions from ecological models account for only a subset of the various components of uncertainty, making it diffcult to determine which uncertainties drive inaccurate predictions. To address this issue, we leveraged the forecast-analysis cycle and created a new state data assimilation algorithm that accommodates non-normal datasets and incorporates a commonly left-out uncertainty, process error covariance. We evaluated this novel algorithm with a case study where we assimilated 50 years of tree-ring-estimated aboveground biomass data into a forest gap model. To test assumptions about which uncertainties dominate forecasts of forest community and carbon dynamics, we partitioned hindcast variance into five uncertainty components. Contrary to the assumption that demographic stochasticity dominates forest gap dynamics, we found that demographic stochasticity alone massively underestimated forecast uncertainty (0.09% of the total uncertainty) and resulted in overconfident, biased model predictions. Similarly, despite decades of reliance on unconstrained “spin-ups” to initialize models, initial condition uncertainty declined very little over the forecast period and constraining initial conditions with data led to large increases in prediction accuracy. Process uncertainty, which up until now had been diffcult 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, a recent focus of the modeling community, contributed 18.3%. These findings call into question our conventional wisdom about how to improve forest community and carbon cycle projections. This foundation can be used to test long standing modeling assumptions across fields in global change biology and specifically challenges the conventional wisdom regarding which aspects dominate uncertainty in the forest gap models.