Multivariate estimation for accurate and logically consistent forest-attributes maps at macroscales

Abstract

Spatially explicit wall-to-wall forest-attributes information is critically important for designing management strategies resilient to climate-induced uncertainties. Multivariate estimation methods that link forest attributes and auxiliary variables at full-information locations can be used to estimate the forest attributes for locations with only auxiliary variables information. However, trade-offs between estimation accuracies versus logical consistency among estimated attributes may occur. This is particularly likely for macroscales (i.e., ≥1 Mha) with large forest-attributes variances and wide spacing between full-information locations. We examined these trade-offs for ∼390 Mha of Canada’s boreal zone using variable-space nearest-neighbours imputation versus two modelling methods (i.e., a system of simultaneous nonlinear models and kriging with external drift). We found logical consistency among estimated forest attributes (i.e., crown closure, average height and age, volume per hectare, species percentages) using (i) k ≤ 2 nearest neighbours or (ii) careful model selection for the modelling methods. Of these logically consistent methods, kriging with external drift was the most accurate, but implementing this for a macroscale is computationally more difficult. This extra cost is justified given the importance of assessing strategies under expected climate changes in Canada’s boreal forest and in other forest regions.