Modeling and propagating inventory‐based sampling uncertainty in the large‐scale forest demographic model “MARGOT”

• Published in: Natural resource modeling Vol. 35; no. 4
• Main Authors: Audinot, Timothée, Wernsdörfer, Holger, Le Moguédec, Gilles, Bontemps, Jean‐Daniel
• Format: Journal Article
• Language: English
• Published: Hoboken John Wiley & Sons, Inc 01.11.2022; Rocky Mountain Mathematics Consortium
• Subjects: Accuracy; Agricultural sciences; Analysis; Bias; bootstrap; Covariance; demographic model; Demographics; Demography; error propagation; forest dynamic; Forest management; Forest reserves; Forest resources; Forests; Life Sciences; Mathematical models; matrix model; Modelling; Mortality; national forest inventory; Normal distribution; Nuclear factor I; Parameter uncertainty; Probability distribution functions; Project management; Sampling; Scale models; Silviculture, forestry; Simulation; Simulation methods; Statistical analysis; uncertainty; Variance; France; Europe; National forest inventory; Demographic model; Uncertainty; Matrix model; Bootstrap; Forest dynamic; Sampling; Error propagation
• ISSN: 0890-8575; 1939-7445
• DOI: 10.1111/nrm.12352

Models based on national forest inventory (NFI) data intend to project forests under management and policy scenarios. This study aimed at quantifying the influence of NFI sampling uncertainty on parameters and simulations of the demographic model MARGOT. Parameter variance–covariance structure was estimated from bootstrap sampling of NFI field plots. Parameter variances and distributions were further modeled to serve as a plug‐in option to any inventory‐based initial condition. Forty‐year time series of observed forest growing stock were compared with model simulations to balance model uncertainty and bias. Variance models showed high accuracies. The Gamma distribution best fitted the distributions of transition, mortality and felling rates, while the Gaussian distribution best fitted tree recruitment fluxes. Simulation uncertainty amounted to 12% of the model bias at the country scale. Parameter covariance structure increased simulation uncertainty by 5.5% in this 12%. This uncertainty appraisal allows targeting model bias as a modeling priority. Recommendations for Resource Managers • Uncovering the potential and limitations of large‐scale forest models are needed when deducing recommendations from forest resource projections under forest management and policy scenarios at regional, national, or continental scales. • Estimating simulation uncertainty in these models is crucial to assess their accuracy. The present study offers a generic methodological strategy for assessing parameter uncertainty in large‐scale forest models. • Users of the MARGOT model should consider that simulation uncertainties proved to be low at a national scale, but decennial wood stock increases as observed in the French forests over the period 1970–2016 were underestimated. • Assessing simulation uncertainty is also major for model bias appraisal. Better accounting for the controls of forest demographic processes (growth, regeneration and mortality) appears to be a priority for the development of MARGOT, and for other large‐scale models.