Professor University of Toronto Toronto, Ontario, Canada
Wood carbon (C) fractions are a key wood functional trait that is critical for refining estimates of tree- and forest-level C stocks, with previous work demonstrating that accurate wood C data may reduce errors in forest C estimates by as much as 8.9% in some biomes. Availability of wood C data for forest C estimation purposes has improved considerably in recent years, and as a result, there now exists multiple studies promulgating empirically supported “generalized” wood C fractions for use in forest C models. These data-driven wood C fractions represent significant improvements compared to wood C fractions commonly used in earlier forest C estimation methodologies (i.e., where a 50% wood C fraction has been widely used). However, decisions surrounding how to explicitly integrate large-scale wood C datasets into forest C estimation models remain unresolved. Here, we develop and evaluate multiple pathways for integrating our Global Wood Carbon Database (GLOWCAD), into forest C models. This analysis specifically employs phylogenetic imputation methods, mixed effects models, and machine learning techniques (i.e., convolutional neural networks), to identify the optimal methods for integrating wood C fraction data into forest C models, across different data availability scenarios.
Our analysis made use of over 3,500 wood C fraction observations across >850 species. Within species, wood C from stem wood strongly predicts wood C in other tissues, indicating that species-specific wood C fractions for all woody biomass components are well approximated by stem wood C. We also detected a significant phylogenetic signal in wood C concentrations across species, supporting the use of phylogenetic imputation methods for estimating missing wood C fractions across species. Linear mixed effects models and variance partitioning techniques support the use of “generalized” wood C fractions that could apply to trees of different taxonomic divisions (i.e., angiosperms and gymnosperms) and biomes (e.g., temperate and tropical). However, in a training dataset, the error associated with using division-by-biome-specific wood C fractions vs. phylogenetically imputed estimates was larger. Our results support the following decisions (beginning with the most accurate) for integrating wood C data (namely, GLOWCAD) into forest C models: 1) using species-specific wood C fractions where available; 2) using species-specific phylogenetically informed wood C fractions and; 3) using division- and biome-specific wood C fractions. Moreover, these steps represent statistically significant improvements over the highly generalized (e.g., 50%) wood C fractions currently employed in many forest C estimation models.