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

Abstract

11Predictions from ecological models necessarily include five different uncertainties: demographic stochas-12 ticity, initial conditions, external forcing (i.e., drivers/covariates), parameters, and modeled processes. 13 However, most predictions from process-based ecological models only account for a subset of these un-14 certainties (e.g. only demographic stochasticity). This underestimation of uncertainty runs the risk of 15 producing precise, but inaccurate predictions. To address these limitations, we created a new generaliz-16 able ensemble state data assimilation algorithm that accommodates two common features of ecological 17 data, zero-truncation and zero-inflation, and allows the estimation of process error and its covariance 18 among multiple ecological variables. We then demonstrate the use of this novel algorithm by assimilating 19 50 years of tree-ring-estimated species-level biomass at Harvard Forest into a process-based forest gap 20 model. Finally, we partitioned the variance in this hindcast to test long-standing assumptions in the 21 ecological modeling community about which uncertainties dominate our ability to forecast forest com-22 munity and carbon dynamics. Contrary to >40 years of research relying on stochastic forest gap models, 23 we found that demographic stochasticity alone massively underestimated forecast uncertainty (0.09% 24 of the total) and resulted in overconfident and biased model predictions. Similarly, despite decades of 25 reliance on unconstrained "spin ups" to initialize models, constraining initial conditions with data led to 26 the largest increases in prediction accuracy. Counter to conventional wisdom from modeling other Earth 27 1 system process, initial condition uncertainty declined very little over the forecast time period. Process 28 variance, which heretofore had been difficult to estimate in mechanistic ecosystem model projections, 29 dominated the prediction uncertainty over the forecast time period (49.1%), followed by meteorological 30 uncertainty (32.5%). Parameter uncertainty, which has recently been the focus of the modeling commu-31 nity, contributed a modest 18.3%. These findings call into question much of our conventional wisdom 32 about how to improve forest community and carbon cycle projections on multi-decadal to centennial time 33 scales. This foundation can be used to test long standing modeling assumptions across fields in global 34 change biology and suggests a fairly significant reorientation of the modeling community toward better 35 initialization of models with current observations and efforts to better quantify, propagate, and reduce 36 process error. These approaches have the potential to improve both the accuracy of ecological forecasts 37 and our understanding of the predictability of ecological processes. 38 Running Head: Drivers of multi-decadal forecast uncertainty 39 nity ecology, Tobit Wishart ensemble filter (TWEnF) 41 1 Introduction 42 Understanding how different component uncertainties contribute to the overall uncertai...

A comprehensive assessment of in situ and remote sensing soil moisture data assimilation in the APSIM model for improving agricultural forecasting across the U.S. Midwest

“…The new method, hereinafter called the Generalized Ensemble Filter (GEF), comprises a fully numerical Bayesian approach to estimating the analysis distribution and an inflation scalar. The model resembles the approach presented by Raiho et al (2020) and Dokoohaki et al, (2022a) where Q is the estimated forecast inflation scalar and XA is a drawn sample from the analysis distribution. The estimation of XA and Q was completed using a Markov Chain Monte Carlo (MCMC) approach by leveraging the nimble R library (de Valpine et al, 2017).…”

Tackling unresolved questions in forest ecology: The past and future role of simulation models

“…First, they allow for understanding differences between models by relating the simulated pattern of each model to its underlying processes. This can identify model structural uncertainties, which have been highlighted as a major source of model uncertainties (Famiglietti et al, 2020;Lovenduski & Bonan, 2017;Raiho et al, 2020), and thus foster new model developments as well as novel empirical investigations. Although the increasing complexity of models makes the interpretation of model inter-comparison results challenging (Fisher & Koven, 2020; Appendix B), model benchmarking is facilitated by new tools of code and data sharing (e.g., Ram, 2013) as well as the availability of detailed standardized databases (Collier et al, 2018;Reyer et al, 2020).…”

Beyond ecosystem modeling: A roadmap to community cyberinfrastructure for ecological data‐model integration

“…Like calibration, the data assimilation methods that drive forecasting, through a formal fusion of data and modeled states (or both states and parameters), also require advanced statistical and computational expertise. Ecological models and data frequently violate the statistical assumptions embedded in assimilation algorithms developed in other disciplines (e.g., normality, homoscedasticity, independence); hence, [R27] many existing tools need to be reassessed and generalized by experts within community tools to appropriately meet the ecological model-data characteristics (Raiho et al, 2020).…”

“…By publicly archiving and reporting results community cyberinfrastructure enables comparisons of different forecasting approaches, future syntheses, and assessment of improvement over time. These features are integral to the vision for such an infrastructure and could then be coupled to, and build upon, existing community tools for workflow scheduling (Oliver et al, 2019) and data assimilation (Fox et al, 2018;Raiho et al, 2020;Pinnington et al, 2020).…”