ign-lidar-hd 2.5.0

IGN LiDAR HD Dataset Processing Library for Building LOD Classification

IGN LiDAR HD Processing Library

Version 2.5.0 | 📚 Full Documentation

Transform IGN LiDAR HD point clouds into ML-ready datasets for building classification

Quick Start • What's New • Features • Documentation • Examples

---

📊 Overview

A comprehensive Python library for processing French IGN LiDAR HD data into machine learning-ready datasets. Features include GPU acceleration, rich geometric features, RGB/NIR augmentation, and flexible YAML-based configuration.

Key Capabilities:

• 🚀 GPU Acceleration: 6-20x speedup with RAPIDS cuML
• 🎨 Multi-modal Data: Geometry + RGB + Infrared (NDVI-ready)
• 🏗️ Building Classification: LOD2/LOD3 schemas with 15-30+ classes
• 📦 Flexible Output: NPZ, HDF5, PyTorch, LAZ formats
• ⚙️ YAML Configuration: Reproducible workflows with example configs

---

✨ What's New in v2.5.0

🎉 Major Release: System Consolidation & Modernization

v2.5.0 represents a complete internal modernization while maintaining 100% backward compatibility!

Unified Feature System ✨

• FeatureOrchestrator: New unified class replaces FeatureManager + FeatureComputer
• Simpler API: One class handles all feature computation with automatic strategy selection
• Better organized: Clear separation of concerns with strategy pattern
• Fully compatible: All existing code works without changes

Improved Code Quality

• 67% reduction in feature orchestration code complexity
• Enhanced error messages and validation throughout
• Complete type hints for better IDE support
• Modular architecture for easier maintenance and extension

Migration Made Easy

• Zero breaking changes: Your v1.x code continues to work
• Deprecation warnings: Clear guidance for future-proofing your code
• Migration guide: Step-by-step instructions in MIGRATION_GUIDE.md
• Backward compatible: Legacy APIs will be maintained through v2.x series

# NEW (v2.0) - Recommended unified API
from ign_lidar import LiDARProcessor

processor = LiDARProcessor(
    config_path="config.yaml",
    feature_mode="lod3"  # Clearer mode specification
)

# Access unified orchestrator
orchestrator = processor.feature_orchestrator
print(f"Feature mode: {orchestrator.mode}")
print(f"Has RGB: {orchestrator.has_rgb}")
print(f"Available features: {orchestrator.get_feature_list('lod3')}")

# OLD (v1.x) - Still works with deprecation warnings
# feature_manager = processor.feature_manager  # Deprecated but functional
# feature_computer = processor.feature_computer  # Deprecated but functional

Why upgrade?

• Future-proof your code for v3.0
• Access to new features and improvements
• Better performance and error handling
• Professional, maintainable codebase

📖 See MIGRATION_GUIDE.md for complete upgrade instructions
📖 Full Release History

---

🚀 Quick Start

Installation

# Standard installation (CPU)
pip install ign-lidar-hd

# Optional: GPU acceleration (6-20x speedup)
./install_cuml.sh  # or follow GPU_SETUP.md

Basic Usage

# Download sample data
ign-lidar-hd download --bbox 2.3,48.8,2.4,48.9 --output data/ --max-tiles 5

# Enrich with features (GPU accelerated if available)
ign-lidar-hd enrich --input-dir data/ --output enriched/ --use-gpu

# Create training patches
ign-lidar-hd patch --input-dir enriched/ --output patches/ --lod-level LOD2

Python API

from ign_lidar import LiDARProcessor

# Initialize and process
processor = LiDARProcessor(lod_level="LOD2")
patches = processor.process_tile("data.laz", "output/")

---

📋 Key Features

Core Processing

• 🎯 Complete Feature Export - All 35-45 computed geometric features saved to disk (v2.4.2+)
• 🏗️ Multi-level Classification - LOD2 (12 features), LOD3 (38 features), Full (43+ features) modes
• 📊 Rich Geometry - Normals, curvature, eigenvalues, shape descriptors, architectural features, building scores
• 🎨 Optional Augmentation - RGB from orthophotos, NIR, NDVI for vegetation analysis
• ⚙️ Auto-parameters - Intelligent tile analysis for optimal settings
• 📝 Feature Tracking - Metadata includes feature names and counts for reproducibility

Performance

• 🚀 GPU Acceleration - RAPIDS cuML support (6-20x faster)
• ⚡ Parallel Processing - Multi-worker with automatic CPU detection
• 🧠 Memory Optimized - Chunked processing, 50-60% reduction
• 💾 Smart Skip - Resume interrupted workflows automatically (~1800x faster)

Flexibility

• 📁 Processing Modes - Three clear modes: patches only, both, or LAZ only
• 📋 YAML Configs - Declarative workflows with example templates
• 📦 Multiple Formats - NPZ, HDF5, PyTorch, LAZ (single or multi-format)
• 🔧 CLI & API - Command-line tool and Python library

---

💡 Usage Examples

Mode 1: Create Training Patches (Default)

# Using example config
ign-lidar-hd process \
  --config-file examples/config_training_dataset.yaml \
  input_dir=data/raw \
  output_dir=data/patches

# Or with CLI parameters
ign-lidar-hd process \
  input_dir=data/raw \
  output_dir=data/patches \
  output.processing_mode=patches_only

Mode 2: Both Patches & Enriched LAZ

ign-lidar-hd process \
  --config-file examples/config_complete.yaml \
  input_dir=data/raw \
  output_dir=data/both

Mode 3: LAZ Enrichment Only

ign-lidar-hd process \
  --config-file examples/config_quick_enrich.yaml \
  input_dir=data/raw \
  output_dir=data/enriched

⚠️ Note on Enriched LAZ Files: When generating enriched LAZ tile files, geometric features (normals, curvature, planarity, etc.) may show artifacts at tile boundaries due to the nature of the source data. These artifacts are inherent to tile-based processing and do not appear in patch exports, which provide the best results for machine learning applications. For optimal quality, use patches_only or both modes.

GPU-Accelerated Processing

ign-lidar-hd process \
  --config-file examples/config_gpu_processing.yaml \
  input_dir=data/raw \
  output_dir=data/output

Preview Configuration

ign-lidar-hd process \
  --config-file examples/config_training_dataset.yaml \
  --show-config \
  input_dir=data/raw

Python API Examples

from ign_lidar import LiDARProcessor, IGNLiDARDownloader

# Download tiles
downloader = IGNLiDARDownloader("downloads/")
tiles = downloader.download_by_bbox(bbox=(2.3, 48.8, 2.4, 48.9), max_tiles=5)

# Process with custom config
processor = LiDARProcessor(
    lod_level="LOD3",
    patch_size=150.0,
    num_points=16384,
    use_gpu=True
)

# Single tile
patches = processor.process_tile("input.laz", "output/")

# Batch processing
patches = processor.process_directory("input_dir/", "output_dir/", num_workers=4)

# PyTorch integration
from torch.utils.data import DataLoader
dataset = LiDARPatchDataset("patches/")
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

---

🎓 Feature Modes (LOD2 vs LOD3 vs Full)

LOD2 Mode (12 features) - Fast Training

Best for: Basic building classification, quick prototyping, baseline models

Features: XYZ, normal_z, planarity, linearity, height, verticality, RGB, NDVI

Performance: ~15s per 1M points (CPU), fast convergence

LOD3 Mode (38 features) - Detailed Modeling

Best for: Architectural modeling, fine structure detection, research

Additional Features: Complete normals (3), eigenvalues (5), curvature (2), shape descriptors (6), height features (3), building scores (3), density features (5), architectural features (4)

Performance: ~45s per 1M points (CPU), best accuracy

Full Mode (43+ features) - Complete Feature Set

Best for: Research, feature analysis, maximum information extraction

All Features: Everything from LOD3 plus additional height variants (z_absolute, z_from_ground, z_from_median), distance_to_center, local_roughness, horizontality

Performance: ~50s per 1M points (CPU), complete geometric description

Output Format:

• NPZ/HDF5/PyTorch: Full feature matrix with all features
• LAZ: All features as extra dimensions for GIS tools
• Metadata: feature_names and num_features for tracking

📖 See Feature Modes Documentation for complete details.

---

📦 Output Format

NPZ Structure

Each patch is saved as NPZ with:

{
    'points': np.ndarray,        # [N, 3] XYZ coordinates
    'normals': np.ndarray,       # [N, 3] surface normals
    'curvature': np.ndarray,     # [N] principal curvature
    'intensity': np.ndarray,     # [N] normalized intensity
    'planarity': np.ndarray,     # [N] planarity measure
    'verticality': np.ndarray,   # [N] verticality measure
    'density': np.ndarray,       # [N] local point density
    'labels': np.ndarray,        # [N] building class labels
    # Facultative features:
    'wall_score': np.ndarray,    # [N] wall likelihood (planarity * verticality)
    'roof_score': np.ndarray,    # [N] roof likelihood (planarity * horizontality)
    # Optional with augmentation:
    'red': np.ndarray,           # [N] RGB red
    'green': np.ndarray,         # [N] RGB green
    'blue': np.ndarray,          # [N] RGB blue
    'infrared': np.ndarray,      # [N] NIR values
}

Available Formats

• NPZ - Default NumPy format (recommended for ML)
• HDF5 - Hierarchical data format
• PyTorch - .pt files for PyTorch
• LAZ - Point cloud format for visualization (may show boundary artifacts in tile mode)
• Multi-format - Save in multiple formats: hdf5,laz, npz,torch

💡 Tip: For machine learning applications, NPZ/HDF5/PyTorch patch formats provide cleaner geometric features than enriched LAZ tiles.

---

📚 Documentation

Quick Links

• 📖 Full Documentation
• 🚀 Installation Guide
• ⚡ GPU Setup
• 🎯 Quick Reference
• 🗺️ QGIS Integration

Examples & Workflows

• examples/ - Python usage examples and configuration templates
• examples/config_lod2_simplified_features.yaml - Fast LOD2 training (12 features)
• examples/config_lod3_full_features.yaml - Detailed LOD3 modeling (38 features)
• examples/config_complete.yaml - Full mode with all 43+ features
• examples/config_multiscale_hybrid.yaml - Multi-scale adaptive features
• PyTorch Integration
• Parallel Processing

Architecture & API

• System Architecture
• Geometric Features Reference
• Feature Modes Guide
• CLI Reference
• Python API
• Configuration Schema

---

🛠️ Development

# Clone and install in development mode
git clone https://github.com/sducournau/IGN_LIDAR_HD_DATASET
cd IGN_LIDAR_HD_DATASET
pip install -e ".[dev]"

# Run tests
pytest tests/

# Format code
black ign_lidar/

---

📋 Requirements

Core:

• Python 3.8+
• NumPy >= 1.21.0
• laspy >= 2.3.0
• scikit-learn >= 1.0.0

Optional GPU Acceleration:

• CUDA >= 12.0
• CuPy >= 12.0.0
• RAPIDS cuML >= 24.10 (recommended)

---

📄 License

MIT License - see LICENSE file for details.

---

🤝 Support & Contributing

• 🐛 Report Issues
• 💡 Feature Requests
• 📖 Contributing Guide

---

📝 Cite Me

If you use this library in your research or projects, please cite:

@software{ign_lidar_hd,
  author       = {Ducournau, Simon},
  title        = {IGN LiDAR HD Processing Library},
  year         = {2024},
  publisher    = {GitHub},
  url          = {https://github.com/sducournau/IGN_LIDAR_HD_DATASET},
  version      = {2.5.0}
}

Project maintained by: ImagoData

---

Made with ❤️ for the LiDAR and Machine Learning communities

⬆ Back to top