PureForest: A Large-Scale Aerial Lidar and Aerial Imagery Dataset for Tree Species Classification in Monospecific Forests

Charles Gaydon, Floryne RocheWACV 2025

TL;DR

This paper introduces PureForest, the largest open-access dataset for tree species classification using aerial LiDAR and imagery. It addresses the critical need for large-scale, diverse benchmarks to advance deep learning models for forest management, which currently lack sufficient data for robust training and evaluation.

What Problem Does This Paper Address?

Accurate mapping of tree species is crucial for sustainable forest management, especially in the face of climate change. This mapping is essential for estimating biomass, assessing fire vulnerability, and promoting resilient tree species. Traditionally, this has relied on time-consuming visual identification from aerial imagery. While deep learning shows promise, existing datasets for tree species classification are scarce, small, and often lack the 3D structural information captured by LiDAR, making it difficult to distinguish between species. This data gap hinders the development of scalable, automated methods for forest monitoring.

Methods

The authors developed a semi-automatic annotation workflow to create the PureForest dataset. This involved:

  • Polygon Generation: Using existing forest inventory data (BD Forêt V2) and national forest databases to identify monospecific forest areas.
  • Expert Validation: A trained forest expert validated and corrected these polygons using aerial imagery from the Geoportail platform.
  • Semantic Grouping: Tree species were grouped into 13 semantic classes (e.g., Deciduous oak, Maritime pine) to simplify classification.
  • Patching and Splitting: Monospecific forest polygons were divided into 50m x 50m patches. These patches were then split into training, validation, and testing sets, stratified by semantic class and with a polygon-level split to ensure spatial generalization.
  • Subsampling: To manage dataset size and prevent bias from large polygons, patches in the train and validation sets were subsampled to a maximum of 1000 per polygon.

For experimental evaluation, they established baseline models:

  • Image Classification: A ResNet18 encoder, pretrained on ImageNet, with standard data augmentations.
  • LiDAR Classification: A RandLA-Net architecture, a lightweight and scalable model suitable for large-scale point cloud processing, with extensive data augmentation.

Datasets

The primary contribution is the PureForest dataset, which includes:

  • Aerial LiDAR Point Clouds: High-density ALS data from the Lidar HD program.
  • Very High Resolution (VHR) Aerial Imagery: Ortho-rectified images from the ORTHO HR database.

The dataset covers 339 km² across 449 monospecific forests in France, with verified labels for 18 tree species grouped into 13 semantic classes. This is significantly larger and more spatially diverse than existing public LiDAR datasets like pytreedb.

Key Insights and Ideas

  • LiDAR's Potential: The study strongly suggests that LiDAR data, by capturing 3D forest structure, is highly competitive with, and potentially superior to, aerial imagery for tree species classification, even with basic models.
  • Data Scale Matters: The sheer size and diversity of PureForest are highlighted as crucial for developing robust deep learning models that can generalize across different forest areas.
  • Semi-Automatic Annotation Efficiency: The workflow demonstrates an effective way to create large-scale, annotated datasets with reduced manual effort, making it more feasible to build comprehensive benchmarks.
  • Monospecific Forests as a Stepping Stone: Focusing on monospecific forests simplifies the problem, providing a necessary foundation before tackling the more complex challenge of mixed forest classification.

Key Results

  • The LiDAR baseline model achieved a weighted F1-score of 81.0% and an Overall Accuracy (OA) of 80.3% on the test set.
  • The image baseline model achieved a weighted F1-score of 74.6%.
  • The LiDAR model outperformed the image model by approximately 6 percentage points in weighted F1-score.
  • Classwise metrics show high accuracy for well-represented classes (e.g., Deciduous oak >90% accuracy), while rarer classes like Fir and Douglas showed lower performance, as expected.
  • Interclass confusions were observed between similar tree species (e.g., Oaks, Pines), indicating that even advanced models can struggle with subtle distinctions.

Limitations and Future Work

  • Temporal Synchronization: The temporal synchronization between LiDAR and aerial imagery is modest (up to 3 years), which could introduce minor discrepancies.
  • Spatial Autocorrelation: While efforts were made to mitigate it, some degree of spatial autocorrelation might still exist due to the polygon-level split.
  • Monospecific Focus: The dataset is limited to monospecific forests, which simplifies the problem. Future work should address mixed forests.
  • Single-Tree Classification: The current annotations are at the forest patch level. Future work could involve individual tree crown segmentation (ITCS) for single-tree classification.
  • Multimodal Fusion: The paper suggests that exploring 2D-3D modality fusion is a promising avenue for future research.
  • Contextual Metadata: The georeferenced nature of the dataset allows for augmentation with contextual metadata like soil types or satellite imagery.

Summary updated

BibTeX
@inproceedings{gaydon2025pureforest,
  author = {Charles Gaydon and Floryne Roche},
  title = {{PureForest: A Large-Scale Aerial Lidar and Aerial Imagery Dataset for Tree Species Classification in Monospecific Forests}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure