Abstract: Accurate estimation of aboveground biomass (AGB) is essential for forest monitoring and carbon stock assessment. Airborne laser scanning (ALS) is widely used for large-scale AGB estimation, yet acquiring reference biomass from field measurements for training biomass regression models remains time-consuming and labour-intensive. Here we explore the potential of synthetic ALS data to enhance forest biomass estimation. Two virtual forest plots were generated using a voxel-based forest reconstruction approach to simulate ALS data. We compared the model performances under varying amount and proportion of simulated and real samples in the training set. We find that models trained exclusively on simulated samples underperform models trained solely on real samples. When real samples are scare, incorporation of synthetic samples substantially improves the model performance, with coefficient of determination (R²) increased by 0.001–0.73 and the root mean square error (RMSE) decreased by 0.07–2.26 Mg ha^-1. When sufficient real samples are available, adding a small number of simulated samples further improves model performance, with RMSE decreased by 0.12–1.46 Mg ha^-1. The optimal performance (R² = 0.852, RMSE = 33.47 Mg ha^-1) is obtained when real samples comprise about 83% of the training samples. These findings demonstrate that synthetic ALS data can effectively complement real datasets in AGB modelling, improving accuracy under diverse data availability conditions.

Key words: airborne laser scanning (ALS), aboveground biomass (AGB), LiDAR simulation, virtual forest, random forest