Efficient Urban Tree Species Classification via Multirepresentation Fusion of Mobile Laser Scanning Data

Urban tree species identification is crucial for forest management and ecosystem assessment. Mobile laser scanning (MLS) provides significant advantages for this task through its flexibility in navigating complex urban environments with spatial constraints. However, MLS-based classification faces challenges, such as intricate canopy structures, incomplete point clouds from urban occlusions, and intraspecies variations. This study presents tree morphology multirepresentation fusion network, integrating 3-D point cloud data with 2-D projections for enhanced tree species classification. The framework employs a backpack-mounted MLS system to capture high-quality point cloud data. The core architecture features an adaptive hierarchical sampling module extracting multiscale geometric features, followed by a cross-view fusion module that implements stagewise fusion of 3-D structural information with 2-D representations. This fusion strategy not only leverages established 2-D feature extraction pipelines, but also addresses sparsity issues in point cloud projections. The method was validated on a diverse dataset of eight urban tree species (seven broadleaf and one coniferous species). Quantitative assessment yielded 98.57% F1-score and 98.77% overall accuracy with moderate computational resources (2.25M parameters, 1.11G FLOPs), demonstrating significant improvements over existing methods. The proposed workflow achieves a balance between classification accuracy and processing efficiency, making it suitable for large-scale urban tree inventory applications.