InceptionFormer: A Deep Learning Framework for Individual Tree Point Cloud Completion

In dense forests, UAV laser scanning (ULS) point clouds of individual trees often suffer from structural incompleteness in the lower trunk region due to canopy occlusion and signal attenuation. This incompleteness affects the accuracy of forest structural parameter estimation and carbon stock assessment.To address this issue, we propose InceptionFormer, a deep learning network designed for individual tree point cloud structural completion. The model integrates an Inception Feature Aggregation (IFA) module to extract multi-scale geometric features and a Sparse Attention (PSA) module to capture global contextual information, enabling effective learning of structural incompleteness. To train and evaluate the model, we further construct the TreeCompletion3D dataset by collecting multiple publicly available UAV LiDAR datasets and simulating structural absence based on height, point density, and canopy base height (CBH).