Field estimates of aboveground biomass are often referred to as ‘ground truth’ data in remote sensing studies, but without destructively sampling biomass in the field we do not know how accurate field estimates are. In this study, we demonstrate that allometric model selection yields on average a 19% difference in field plot estimates of aboveground biomass, and a 20% difference in the resulting county-level biomass map for Sonoma County.
Sonoma County is a particularly interesting area to conduct this study, as it hosts forests with some of the highest biomass densities in the United States. We see that the majority of discrepancies between allometric predictions occur in these high biomass areas, which is expected because more trees with smaller stems are destructively sampled for allometric model fitting, and indeed published allometries are often caveated by unknown uncertainties above a certain stem girth . For example, 91 of the trees in our field plots were larger than the largest tree destructively sampled in a respective species class, as reported by Jenkins, thus potentially contributing to greater errors for these 91 trees. Although these trees only represent 8% of those sampled across the county, they represent 18% of the total estimated biomass in our field plots. However, given a lack of other available models, generalized allometric models such as applied in this paper are typically applied regardless of tree size.
The Jenkins and Chojnacky models were expected to perform similarly, as they are largely based on the same datasets of destructively sampled trees. The two studies combined the meta-analysis differently, partitioning allometric models from the literature into different combinations based on generalized species classes  or theoretical taxonomic groupings and wood specific gravity . On average, these models produce similar biomass estimates and total county-level predictions, but discrepancies exist on a plot-to-plot basis depending on the species composition of a given plot. Most notably, the Chojnacky models produce greater estimates in high biomass plots, potentially because the models used by Chojnacky are more species-specific than the Jenkins models.
As with previous studies [4, 21], we found that the CRM field plot estimates were ~ 20% less than the Jenkins predictions. In a similar analysis in Maryland, CRM estimates in FIA plots were only ~ 11% less than the Jenkins estimates at the state level [11, 13]. The discrepancy between the differences seen in Maryland and Sonoma County could be due to the prevalence of conifer growth forms and larger trees found in Sonoma County, where we saw that estimates varied more in high biomass than low biomass areas.
There are several explanations for the differences between the Jenkins/Chojnacky and the CRM estimates seen both in this study and consistently observed in regional and national scale studies [5, 6, 14, 19]. First, the sample sizes used to construct the two sets of allometric models are different which could lead to systematic differences. Duncanson et al. [3, 7] demonstrated that allometric parameters are very sensitive to sample size, and that small sample sizes likely lead to an overestimate in biomass for a given DBH. The Jenkins and Chojnacky datasets are, on average, developed with smaller sample sizes than the FIA analysis . These smaller sample sizes yield model deviations because of probable differences in the destructively harvested size distribution, with the inclusion likelihood of large individuals in a sample decreasing with sample size. Similarly, CRM volume models applied to certain species in Sonoma County were developed with destructively harvested trees outside of the region, which may yield errors if trees in Sonoma County are growing in different climate conditions or have different resource limitations than those included in the sample .
Finally, others have suggested that the differences between Jenkins and CRM results are due to the inclusion of tree height in the CRM approach, which may better estimate stem volume. However, we do not see strong evidence of this here, as neither maximum nor mean lidar height were highly correlated to differences between CRM and Jenkins field estimates in the tallest forests in our study area. Indeed, the plots with maximum heights between ~ 25 and 45 m had the largest discrepancies, while the tallest trees sampled approached 70 m in height. This may suggest that Jenkins DBH-based estimates may be over estimating biomass in areas of moderate height (~ 25 m) and potentially underestimating biomass in tall forests (> 50 m), while CRM estimates constrain high predictions for a given DBH in relatively short forests and increase estimates for a given DBH in very tall forests. Notably, all of the plots with the highest discrepancies had > 80% canopy cover. As canopy cover is highly correlated to biomass, it is unclear whether the variability in estimates is because of high canopy covers or high biomass densities. Certainly the limited destructive sample for large tree sizes would explain the high biomass density discrepancies, but it is conceivable that destructively harvested trees were also preferentially extracted from open, easily accessible areas with lower canopy covers. This may have caused deviations in allometries in comparison to trees growing in closed canopy systems with relatively taller, smaller crowned individuals.
Our assessment of the drivers of variability between allometric model selection remains speculative. We see that discrepancies are largest in high biomass plots with high canopies covers and moderate heights. Whether these discrepancies are due to inadequate sampling across gradients of biomass, canopy cover or height in either CRM, Jenkins/Chojnacky, or all datasets remains uncertain. Only testing the different allometric approaches against an independent destructively sampled tree dataset can determine the underlying drivers, and such a dataset is currently unavailable for use in this study. However, these results highlight the importance of improving allometric models for biomass mapping. Fortunately, progress is being made in this field, both through the collection of larger destructively harvested tree datasets that can be used to fit improved models (e.g.  or through the derivation of new, non-destructively derived allometries based on terrestrial laser scanning (TLS)(e.g. ).