Precise and unbiased biomass estimation from GEDI data and the US Forest Inventory

Abstract

Atmospheric CO2 concentrations are dependent on land-atmosphere carbon fluxes resultant from forest dynamics and land-use changes. These fluxes are not well-constrained, in part because reliable baseline estimates of forest carbon stocks and the associated uncertainties are lacking. NASA's Global Ecosystem Dynamics Investigation (GEDI) produces estimates of aboveground biomass density (AGBD) that are unique because GEDI's hybrid estimation framework enables formal uncertainty calculations that accompany the biomass estimates. However, GEDI's estimates are not without issue; a recent validation using design-based AGBD estimates from the US Forest Inventory and Analysis (FIA) program revealed systematic differences between GEDI and FIA estimates within a hexagon tessellation of the continental United States. Here, we explored these differences and identified two issues impacting GEDI's estimation process: incomplete filtering of low quality GEDI observations and regional biases in GEDI's footprint-level biomass models. We developed a solution to each, in the form of improved data filtering and GEDI-FIA fusion AGBD models, developed in a scale-invariant small area estimation framework, that were compatible with hybrid estimation. We calibrated 10 regional Fay-Herriot models at the hexagon scale for application at the unit scale of GEDI footprints, for which we provide a mathematical justification and empirical testing of the models' scale-invariance. These models predicted realistic distributions of unit level AGBD, with equal or improved performance relative to GEDI's L4A models for all regions. We then produced GEDI-FIA fusion estimates that were more precise than the FIA estimates and resulted in a bias reduction of 86.7% relative to the original GEDI estimates: 19.3% due to improved data filtering and 67.5% due to the new AGBD models. Our findings indicate that (1) small area estimation models trained in a scale-invariant framework can produce realistic predictions of AGBD, and (2) there is substantial spatial variability in the relationship between GEDI forest structure metrics and AGBD. This work is a step toward achieving reliable baseline forest carbon stocks, provides a viable methodology for training remote sensing biomass models, and may serve as a reference for other investigations of GEDI AGBD estimates.