Search

Large-area growth estimates can be obtained by coupling growth model predictions with wall-to-wall remotely sensed auxiliary variables through a generalized hierarchical model-based (GHMB) inferential framework. So far, most GHMB variance estimators do not account for the residual errors of the submodels and their spatial correlations. This likely induces an underestimation of the true variance of the point estimator. In this study, we provide an example of large-area growth estimation obtained through the GHMB framework. To do this, we developed a new variance estimator that accounts for residual errors as well as potential spatial correlations among them. We tested this variance estimator through a simulation study and then used it to estimate the annual volume increment for a forest management unit in Quebec, Canada. Our results show that, contrary to our expectation, neglecting the residual errors of the different submodels leads to overestimating the true variance of the point estimator. We observed increases in the overestimation with small populations and spatially correlated residual errors. Our developed variance estimator corrected this overestimation and made it possible to derive reliable confidence intervals for annual volume increments at the population level. Key words: generalized hierarchical model-based inference, growth models, Landsat 8, variance estimation, simulation study, 1. Introduction Forest management requires information on current forest attributes and future growth. In mostjurisdictions in Canada, current forest attributes have been traditionally estimated using photo-interpretation and a stratified random [...]

Landscape-level studies such as those on forest management planning and carbon accounting rely on large-area growth projections provided by forest growth models. Nowadays, most of these models are individual tree-based models. The detailed input they require and their complexity are a challenge for the integration into a landscape-level study. A possible alternative consists of approximating the complex model through a meta-model. A meta-model mimics the behaviour of the original model, while being simpler in terms of input and computation. In this study, we developed a Bayesian meta-modelling approach that can be used to obtain a simplified growth model from an individual tree-based model. The approach was exemplified through a real-world case study, namely a forest management unit in the province of Quebec, Canada. Using a Markov chain Monte Carlo method, we managed to fit meta-models based on the Chapman-Richards equation or its derivative for the main potential vegetation types. This meta-modelling approach has the advantages of (i) being an effective method of upscaling, (ii) providing simple meta-models suitable for landscape-level studies, and (iii) ensuring a proper error propagation from the original individual tree-based model into the meta-model. Key words: meta-model, Bayesian statistics, Markov chain Monte Carlo, Metropolis-Hastings algorithm, individual tree-based models, upscaling, forest growth models, forest management planning Les etudes au niveau du paysage, telles que celles portant sur la planification de la gestion forestiere et la comptabilisation du carbone, reposent sur des projections de croissance sur de grandes surfaces fournies par des modeles de croissance forestiere. Aujourd'hui, la plupart de ces modeles sont des modeles bases sur des arbres individuels. Les donnees detaillees qu'ils requierent et leur complexite constituent un defi pour leur integration dans une etude au niveau du paysage. Une alternative possible consiste a approximer le modele complexe par un meta-modele. Un meta-modele imite le comportement du modele original, tout en etant plus simple en termes de saisie et de calcul. Dans cette etude, nous avons developpe une approche de meta-modelisation bayesienne qui peut etre utilisee pour obtenir un modele de croissance simplifie a partir d'un modele base sur des arbres individuels. L'approche a ete illustree par une etude de cas reelle, a savoir une unite de gestion forestiere dans la province du Quebec, au Canada. En utilisant une methode de Monte Carlo par chaines de Markov, nous avons reussi a obtenir des meta-modeles bases sur l'equation de Chapman-Richards ou sa derivee pour les principaux types de vegetation potentielle. Cette approche de meta-modelisation presente les avantages suivants : (i) elle constitue une methode efficace de saut d'echelle, (ii) elle fournit des meta-modeles simples adaptes aux etudes a l'echelle du paysage et (iii) elle assure une propagation correcte des erreurs du modele original base sur les arbres individuels vers le meta-modele. [Traduit par la Redaction] Mots-cles: meta-modele, statistiques bayesiennes, Monte Carlo par chaines de Markov, algorithme de Metropolis- Hastings, modeles bases sur des arbres individuels, saut d'echelle, modeles de croissance forestiere, planification de la gestion forestiere, 1. Introduction Landscape-level studies rely on large-area growth projections to assess the sustainability of ecosystem services such as wood supply, carbon sequestration, and water filtering. In this context, growth models [...]