Comparison of shallow and deep learning methods for LiDAR-based aboveground biomass estimation of eucalyptus plantations in Ibu Kota Nusantara, East Kalimantan Province, Indonesia

Forests store substantial carbon in aboveground biomass (AGB), making accurate estimation essential for carbon accounting and climate change mitigation. This study integrates LiDAR-derived metrics, including canopy height metrics (CHM), statistical height metrics (SHM), intensity metrics (IM) and L-moment metrics (LMOM), with machine learning to improve individual-tree AGB prediction. We evaluate tree-based ensemble methods and deep learning architectures, including convolutional neural networks (CNNs) enhanced with multilayer perceptrons (MLP), attention mechanisms, squeeze-and-excitation (SE) blocks and Kolmogorov–Arnold Networks (KAN). Tree-based ensembles achieved robust predictive performance, with LightGBM obtaining the best shallow learning result (test R2=0.63; RMSE = 43.83 kg/tree). Conventional CNNs exhibited lower predictive accuracy, whereas KAN consistently improved nonlinear feature learning, with increasing test R2 from 0.19–0.46 to 0.44–0.59 while reducing RMSE to 45.02–45.78 kg/tree. The best KAN model achieved a test R2 of 0.65 and RMSE of 42.13 kg/tree, albeit with reduced training stability. Furthermore, efficient CNN2D-KAN architectures achieved competitive performance with training times comparable to LightGBM, demonstrating that KAN can improve representation learning. These findings highlight that while ensemble models remain superior in capturing overall variability, the KAN integration plays critical role in enabling deep learning models to reach competitive performance, positioning them as a robust alternative for accurate biomass estimation.