SPTNet: Sparse Convolution and Transformer Network for Woody and Foliage Components Separation From Point Clouds

Highlights

• Novel Hybrid Architecture: First network to combine Sparse Convolution (SpConv) with Transformer blocks for tree component separation, balancing local feature extraction efficiency with global context modeling capabilities.
• Pure Geometry-Based Approach: Uses only XYZ coordinates without relying on radiometric features (LRI), ensuring sensor-agnostic generalization across different LiDAR devices and wavelengths.
• Adaptive Geometric Features: Introduces two innovative feature blocks—Morphological Detection Coefficient (MDC) and Normal Difference Operator (NDO)—with adaptive radius determination strategies (minimum entropy and Otsu methods) to eliminate manual parameter tuning.
• State-of-the-Art Performance: Achieves 94.69% OA and 89.96% mIoU on large tropical tree dataset (15 species), outperforming FWCNN by 0.43% OA and 0.72% mIoU.
• High Efficiency: Training requires only 8 seconds per epoch with 48 epochs to convergence (vs. 15-45s/epoch for competing methods), and model size is less than 1/3 of PointNet++.

Methodology

The proposed SPTNet consists of three main components:

1. Geometric Feature Extraction (Preprocessing):

   • MDC (Morphological Detection Coefficient): Characterizes local point distribution patterns using eigenvalues of covariance matrix. Values > 0.5 indicate linear distribution (woody), < 0.5 indicate planar distribution (foliage). Adaptive radius determined via minimum entropy principle.
   • NDO (Normal Difference Operator): Measures normal vector variations within neighborhoods. Woody components show larger normal differences (cylindrical structure) vs. foliage (planar structure). Adaptive radius determined via Otsu thresholding.
   • Input Representation: 7-dimensional vector (XYZ + MDC + NDO) organized in voxel-hash structure.

2. Network Architecture (UNet-based Encoder-Decoder):

   • Encoder: 5 SpConv blocks (local feature extraction) + 5 Transformer blocks (global context aggregation)
     • SpConv Block: 3 SubMConv layers + 1 SpConv downsampling layer with skip connections
     • Transformer Block: Multi-head self-attention (8 heads) with skip connections from SpConv features
   • Decoder: 4 inverse SpConv blocks for upsampling and dimensionality reduction
   • Output: Linear layer for binary classification (woody vs. foliage)

3. Data Organization:

   • Voxelization: 0.02m voxel size with average pooling for point-to-voxel mapping
   • Sparse Convolution: Exploits data sparsity to reduce computational cost while preserving geometric details
   • Augmentation: Random sampling (density variation) + rotation (< 45°) for training robustness

Datasets

Comprehensive validation across diverse forest environments:

• Data Sources: Canada, Finland, Cameroon (real); OnyxTree simulation (synthetic)
• Sensors: Riegl VZ-1000, various TLS devices
• Annotation: Manual labeling for woody/foliage components

Experimental Results

Quantitative Performance (Eight Datasets):

Comparison with State-of-the-Art Methods:

Efficiency Comparison:

Key Findings

1. Ablation Studies:

   • SpConv vs. Transformer: Removing SpConv blocks causes larger accuracy drop than removing Transformer blocks, indicating local geometric features are crucial for woody/foliage discrimination.
   • Feature Importance: NDO features contribute more than MDC features; combining both yields optimal results (OA improvement of 2-3% over using coordinates only).
   • Backbone Comparison: SPTNet outperforms PointNet++, DGCNN, Point Transformer, and PointBert on all test datasets while maintaining 3-6× faster training speed.

2. Robustness Analysis:

   • Occlusion Resistance: Maintains >90% OA even with 40% random point masking; degrades gracefully at 60-80% masking (OA drops to ~92% and ~77% mIoU at 80% masking).
   • Cross-Species Generalization: Successfully handles both coniferous (Pine: 99.64% WA) and broadleaf species, as well as complex tropical trees with irregular structures.

3. Adaptive vs. Fixed Radius:

   • Adaptive radius strategies (minimum entropy for MDC, Otsu for NDO) eliminate manual parameter tuning while achieving comparable or better accuracy than manually optimized fixed radii (0.1-0.5m range tested).

4. Sensor Independence:

   • Pure geometry-based approach (no radiometric features) enables seamless application to data from different LiDAR sensors without recalibration or retraining for specific wavelength characteristics.

Limitations and Future Work

• Voxel Resolution Sensitivity: Performance depends on voxel size selection (0.02m used); smaller voxels improve accuracy but increase computational cost.
• Scale Constraints: Currently tested at single-tree level; extension to plot-scale processing with multiple trees and ground points requires further development.
• Occlusion Handling: While robust to moderate occlusion (up to 60%), severe occlusion (>80%) significantly degrades performance due to loss of neighborhood geometric information.
• Future Directions: Extension to airborne LiDAR data, semantic segmentation of urban scenes, and integration with instance segmentation for complete forest inventory workflows.