Point cloud feature optimization of dual-mechanism LiDAR with reflection and fluorescence spectra for plant classification

• A LiDAR system that can acquire both reflection and fluorescence information. • Constructing and optimizing spatial and spectral features suitable for classification of complex plant scenario. • Using MPA to optimize features and using label smoothing to improve plant classification accuracy. LiDAR lacks spectral information related to plant properties, limiting its ability to identify plant in complex environments using only three-dimensional spatial information and single-band spectral data. We designed a dual-mechanism LiDAR system that combines reflection and fluorescence, proposing a refined plant classification method based on the optimization of "spatial-reflection-fluorescence" five-dimensional features (3D spatial + dual spectral dimension). We validated the advantages of this system and method in refined plant classification research. First, based on previous studies, we designed a coaxial single-source reflection-fluorescence dual-mechanism LiDAR system, which can accurately extract key reflection and fluorescence band information of target plant. By leveraging the intrinsic coupling relationship of the "spatial-reflection-fluorescence" five-dimensional data in the generated point cloud, we determined precise point cloud neighborhood ranges through dual-spectral information neighborhood optimization. Second, we constructed multiple spatial and spectral features suitable for the dual-mechanism LiDAR. Finally, we employed the MPA (Marine Predator Algorithm) for spatial-spectral feature optimization to eliminate feature redundancy and determine the optimal feature combination for refined plant classification. Additionally, we used dual-spectral point cloud neighborhood label smoothing to reduce salt-and-pepper noise errors, further improving plant classification accuracy, providing significant reference for LiDAR-based refined plant classification research. Compared with the results of only using reflection information for classification, overall classification accuracy after adding fluorescence information were improved by 5.15 % and 4.81 % in two experimental scenarios, respectively. Moreover, using the MPA and dual-spectral point cloud neighborhood label smoothing methods, we achieved classification accuracies of 98.25 % and 98.78 % for different plant categories and growth states in the two experimental scenarios.