A Multifeature Fusion Network for Tree Species Classification Based on Ground-Based LiDAR Data

Light Detection and Ranging (LiDAR) holds considerable promise for tree species classification. Existing networks that utilize point clouds of individual trees have shown promising results. However, challenges such as incomplete point cloud data, uneven point density across different components of the tree, and complex tree morphologies can hinder classification accuracy. To overcome these limitations, we introduced the Multi-Feature Fusion Tree Classifier Network (MFFTC-Net). This network leverages a novel boundary-driven point sampling (BDPS) method that preserves more canopy points and mitigates the effects of uneven point density. We also utilize the umbrella-repSurf (REP-U) module, which captures local geometric features and enhances the model's responsiveness to tree structural nuances. The backbone of MFFTC-Net integrates these innovations through a multi-feature fusion approach, utilizing set abstraction (SA) for local information capture and transformer-based feature interaction (TFI) for robust multi-scale feature integration. Our results demonstrate that MFFTC-Net significantly outperforms other state-of-the-art methods in LiDAR-based tree species classification, achieving the highest overall accuracy and kappa coefficients on both a self-built dataset of four species and a public dataset of seven species.