Shifting from Discrete to Continuous Reference Data: QSM-Derived Horizontal Tree Biomass Distribution for Deep Learning Biomass Estimation

Conventional modeling approaches for LiDAR-based above-ground biomass (AGB) estimation rely on discrete plot-level inventory aggregates. This methodology introduces boundary-effect uncertainties that may severely degrade model performance within small field plots. To solve this limitation, we evaluate a Horizontal Biomass Distribution (HBD) reference mapped continuously from Quantitative Structure Models (QSMs). We trained a sparse 3D U-Net on simulated broadleaved forest structures using three AGB reference types: a standard forest inventory (FI) plot-level aggregate, an edge-effect-free QSM plot-level aggregate, and a continuous HBD mapping. Evaluating training plot sizes scaling from 100 to 2500 m^2 , QSM-based models systematically outperformed FI approaches at small plot sizes. Specifically, for 100 m^2 plots, the HBD reference reduced the relative root mean square error (RRMSE) by 16.84 \pm 4.37 % and increased R^2 by 0.22 \pm 0.05 against the FI baseline. By replacing plot level aggregates with HBDs as AGB reference, this methodology corrects for edge-effects and shows that using an HBD-based reference enhances model performance for small plot sizes.