Introduction
Urban forest management is vital for ecological health, yet manual species identification is prohibitively expensive and slow. Mobile LiDAR (Light Detection and Ranging) systems, such as the RIEGL VMX-450, provide a high-resolution 3D solution by capturing urban environments at driving speeds. However, the resulting point clouds are massive and complex, often featuring overlapping tree crowns and structural similarities between species. This paper proposes an automated, two-stage workflow using a 3D Convolutional Neural Network (3D-CNN) to classify tree species with high precision.
Methodology
The proposed approach consists of two primary phases: individual tree segmentation and species classification.
Tree Segmentation
To isolate individual trees from the raw mobile laser scanning (MLS) data, the researchers first apply a voxel-based upward growing filtering technique to remove terrain points. Following ground removal, trees are extracted using a combination of Euclidean distance clustering and a Voxel-based Normalized Cut (VNCut) algorithm. VNCut is particularly effective at separating intermingled tree crowns by treating voxels as nodes in a graph and determining optimal boundaries based on spatial similarity.
Tree Species Classification
Once segmented, trees are converted into a 3D voxel grid. To enhance the geometric data, Inverse Distance Weighted (IDW) interpolation is used to assign laser intensity values to each voxel, providing a "pseudo-spectral" dimension that reflects leaf and bark characteristics.
The classification is performed by a custom 3D-CNN architecture. As shown in the diagram below, the model utilizes four 3D convolutional layers designed for horizontal-rotation invariance, followed by a max-pooling layer and three fully connected (FC) layers. This architecture allows the network to learn complex 3D branching patterns and foliage densities directly from the volumetric data.
Experimental Results
The model was evaluated using a dataset from Huandao Road in Xiamen, China, encompassing ten common urban tree species. The training set included 50,000 samples, while the test set comprised a 4-kilometer road segment with 2,013 trees.
The 3D-CNN achieved an overall classification accuracy of 96.4% with a Kappa coefficient of 0.96. This performance significantly outperformed traditional deep learning methods like Deep Boltzmann Machines (86.1%) and Deep Belief Networks (95.6%). Producer’s and user’s accuracies remained consistently high (above 93%) across all species, demonstrating the robustness of integrating intensity data with 3D spatial features.
Conclusion
By combining advanced voxel-based segmentation with a 3D-CNN that leverages both geometry and laser intensity, this research provides a highly accurate tool for urban forest inventory. While structural mimicry between certain species remains a minor challenge, future integration of multi-spectral LiDAR data promises to further refine classification capabilities in even more complex environments.
Summary updated