In recent decades, LiDAR technology has become an indispensable tool for collecting extensive 3D data in the field of forest inventory. Often known as laser scanning, this technology facilitates the acquisition of point cloud data, providing detailed insights into canopy structure. In particular, LiDAR provides the opportunity to map forest leaf area with unprecedented accuracy, while leaf area has remained an important source of uncertainty affecting models of gas exchanges between the vegetation and the atmosphere. The vigilant monitoring of leaf area contributes significantly to comprehending the seasonal fluxes in tropical forests, thereby refining the precision of climate models in predicting the repercussions of global warming. Various vehicles are used for data collection, with terrestrial laser scanning (TLS) providing detailed but labor-intensive 3D data. Airborne laser scanning (ALS) covers larger areas but with lower point density, posing challenges for observing understory vegetation due to canopy occlusions. The rise of UAV laser scanning (ULS) addresses these challenges, offering dense data collection without on-site intervention. In fact, forest monitoring requires accurate semantic segmentation to distinguish leaves from wood, which is crucial for monitoring foliage density variations with applications in carbon sequestration, disease monitoring, and harvest planning. Existing methods for TLS data face challenges when applied to ULS due to class imbalance and reliance on unreliable intensity information. To address this, we propose an end-to-end approach named SOUL (Semantic segmentation On ULs) utilizing only point coordinates as input of the neural network for versatility across locations and sensors. The contributions of this work are three-fold. First, SOUL is the first approach designed for semantic segmentation on ULS point clouds in tropical forests, showcasing superior wood point classification. Second, the GVD (Geodesic Voxelization Decomposition) preprocessing method addresses the challenge of training neural networks from sparse point clouds in tropical forest environments. Third, the proposed rebalanced loss function provides a versatile solution for addressing class imbalance in various deep learning architectures. Our research offers promising insights into the application of ULS for forest monitoring, bridging gaps in existing methodologies and laying the foundation for improved semantic segmentation in challenging tropical forest environments.