A modeling framework for forest aboveground biomass estimation in mixed forests of the southeastern United States using airborne lidar and PlanetScope data

Forest Aboveground Biomass Density (AGBD) estimation supports forest carbon accounting and informs carbon monitoring, reporting, and verification. Despite the demonstrated potential of airborne light detection and ranging (lidar) and satellite imagery, accurate AGBD estimation in disturbance-prone, mixed forests remains challenging. To better understand the applicability of these data in disturbance-prone forests, we sought to determine an optimal modeling framework for AGBD estimation. We utilized 70 PlanetScope (3 m), 34 airborne lidar, and 3 ancillary predictors (elevation, slope, and aspect) with field-estimated AGBD across 5 sites in the southeastern United States (US) with different kinds of wind disturbance (tornado, hurricane, straight-line wind): Bankhead, Mountain Longleaf, Oakmulgee, Weeks Bay, and Flagg Mountain. We evaluated: (1) five established variable selection methods; (a) all predictors, (b) top 5 predictors from Random Forest (RF), (c) top 10 predictors from RF, (d) Least Absolute Shrinkage and Selection Operator (lasso), and (e) Recursive Feature Elimination (RFE), and (2) compared 2 modeling algorithms; (a)RF and (b) Bayesian-based Gaussian Process Regression (GPR) for AGBD estimation. Results show that lasso and RF-based variable selection methods outperformed RFE, while GPR outperformed RF. Model accuracy (R 2 = 0.29–0.73; Root Mean Squared Error (RMSE) = 16.29–75.14 Mg/ha) was highest in the undisturbed (Bankhead) and lowest in the wind-disturbance-prone site (Oakmulgee). Findings demonstrate that AGBD estimation is more reliable in undisturbed landscapes, while such frameworks may be inadequate in disturbance-prone, dynamic landscapes. Our study offers optimal modeling frameworks for disturbance-prone mixed forests and advances the synergistic use of lidar and PlanetScope for AGBD estimation. • Built site-specific AGBD frameworks for disturbance-prone sites, fusing airborne lidar and PlanetScope (20 m). • Model accuracies are highest in undisturbed southeastern US forests, and lowest in disturbed forests. • RF-based and lasso feature selection outperformed RFE across sites.