From pixels to parcels: Flexible, practical small-area uncertainty estimation for spatial averages obtained from aboveground biomass maps

Fine-resolution maps of forest carbon and aboveground biomass (AGB) effectively represent spatial patterns and can be flexibly aggregated to map subregions by computing spatial averages or totals of pixel-level predictions. However, generalized model-based uncertainty estimation for spatial aggregates requires computationally expensive processes like iterative bootstrapping and computing pixel covariances. Uncertainty estimation for map subregions is critical for enhancing practicality and eventual adoption of model-based data products, as this capability would empower users to produce estimates at scales most germane to management of the landscape: individual forest stands and ownership parcels. In this study we produced estimates of standard error (SE) associated with spatial averages of fine-resolution AGB map predictions for a stratified random sample of ownership parcels (0–2023 hectares) in New York State (NYS). This represents the first model-based uncertainty estimation study to include all four types of uncertainty (reference data, sample variability, residual variability, and auxiliary data), incorporate spatial autocorrelation of model residuals, and use methods compatible with algorithmic modeling (machine learning or nonparametric). We found that uncertainty attributed to residual variance, largely resulting from spatial correlation of residuals, dominated all other sources for all of the parcels in the study. These results suggest that improvements to model accuracy will yield the greatest reductions to total uncertainty in regions like the northeastern and midwestern United States where forests are divided into smaller spatial units. Further, we demonstrated that log–log regression relating parcel characteristics (area, perimeter, AGB density, forest cover) to parcel-level SE can accurately estimate uncertainty for map subregions, thus providing a convenient means to empower map users. These findings support transparency in future model-based forest carbon accounting and monitoring efforts.