Stable

Individual Tree Segmentation

Segment individual tree crowns from LiDAR point clouds using mean-shift clustering with adaptive bandwidth estimation.

This tool uses a kernel-based mean-shift algorithm to identify and delineate individual tree crowns in airborne LiDAR data. Each tree is segmented as a 3D point cluster, with adaptive bandwidth (kernel smoothing radius) automatically estimated from local crown geometry. This approach is particularly effective for dense forests where trees overlap in 2D projection, as the algorithm operates in full 3D space.

Key Features

• Adaptive Bandwidth Estimation: Automatically scales the search radius for each tree seed based on local point density and crown size, adapting to variable forest structure (e.g., understory vs. canopy trees).
• Vegetation Class Filtering: Optionally restricts segmentation to vegetation-classified points (ASPRS classes 3-5: low/medium/high vegetation), reducing noise from ground or structural outliers.
• Height-Based Filtering: Excludes points below a minimum height threshold (default 2.0 m) to focus on tree crowns above understory vegetation.
• Flexible Output Encoding: Encodes tree segment IDs in LiDAR RGB or classification fields, with support for multi-channel encoding (e.g., RGB+user_data for >16M unique trees).
• Grid-Accelerated Mode: Optional spatial binning for faster convergence on very large point clouds, with configurable refinement iterations.
• Tile-Based Processing: Supports memory-efficient processing of large datasets via optional tiling with configurable overlap.

Algorithm

The tool implements a mean-shift clustering workflow: (1) seeds are initiated at points above the height threshold, (2) each seed iteratively shifts toward the local density mode (3D Euclidean space, weighted by kernel), (3) converged modes are merged if they fall within a distance threshold, (4) points are assigned to their nearest converged mode, and (5) clusters below the minimum point count are rejected as noise. Parallel processing across seeds accelerates computation on multi-core systems.

Use Cases

• Forest inventory and tree counting in dense multi-story stands.
• Canopy structure analysis for ecological or fire-risk modeling.
• Individual tree height and crown diameter extraction (via downstream post-processing).
• Change detection between multi-temporal LiDAR acquisitions.

Performance Tuning

• For Speed: Enable grid_acceleration with larger grid_cell_size (e.g., 1.0), and reduce max_iterations or adaptive_neighbors.
• For Accuracy: Disable grid acceleration or increase grid_refine_iterations, reduce bandwidth_min for smaller trees, or lower convergence_tol for stricter convergence.
• For Large Files: Enable tile_size (e.g., 100 m × 100 m tiles with 5 m overlap) to cap per-tile memory and enable parallel tile processing.

Parameters

Project Links