A comparative analysis of individual tree and area-based approaches for above-ground biomass estimation and uncertainty analysis using LiDAR and hyperspectral data and machine learning

Accurate uncertainty decomposition in above-ground biomass (AGB) estimation is indispensable for the quantification of the carbon sequestration potential. Although parametric models benefit from established uncertainty propagation techniques (e.g. Taylor series expansion), rapid adoption of nonparametric machine learning models necessitates a unified and systematic framework that comprehensively decomposes uncertainty sources across machine learning-based AGB estimation approaches. To address this gap, we propose a hybrid uncertainty quantification approach that synergistically embeds Monte Carlo simulations within machine learning methods. By utilizing unmanned aerial vehicle-mounted light detection and ranging and hyperspectral sensors, we compared the accuracy of four machine learning models and performed comparative uncertainty analysis across two AGB estimation frameworks: individual-tree approach (ITA) and area-based approach (ABA). Two key findings were revealed: 1) model residuals dominated the total uncertainty (contributed 58.8% and 79.6% of the uncertainty in the ITA and ABA, respectively) across all algorithms; 2) the uncertainty in the ITA was 33.9% higher than that in the ABA (16.2% vs. 12.1%), primarily due to cumulative errors in individual-tree detection. Our approach establishes a new paradigm for quantifying errors in machine learning-based carbon stock mapping.