TreeFinder: A US-Scale Benchmark Dataset for Individual Tree Mortality Monitoring Using High-Resolution Aerial Imagery

TreeFinder is a large-scale, high-resolution benchmark dataset for mapping individual dead trees across the contiguous United States (CONUS). The dataset is built from 0.6 m National Agriculture Imagery Program (NAIP) aerial imagery and provides pixel-level annotations of individual dead trees, ecological metadata, and benchmark evaluation settings for robust machine learning model development.

TreeFinder was accepted to the NeurIPS 2025 Datasets & Benchmarks Track.

Project page: https://github.com/zhwang0/treefinder

Dataset Description

Tree mortality is an important indicator of forest health, carbon dynamics, ecosystem disturbance, and wildfire risk. However, existing large-scale forest monitoring products often operate at moderate spatial resolution and may miss individual tree-level mortality signals.

TreeFinder addresses this gap by providing a machine-learning-ready benchmark for individual dead tree segmentation using high-resolution aerial imagery. It is designed to support research at the intersection of computer vision, remote sensing, ecological monitoring, and carbon assessment.

Dataset Summary

TreeFinder includes:

• 1,000 sites across the 48 contiguous U.S. states
• 23,000+ hectares of high-resolution NAIP imagery
• 0.6 m spatial resolution aerial imagery with 4 spectral channels: RGB + NIR
• 20,000+ manually annotated individual dead trees
• Pixel-level segmentation masks for dead tree crowns
• ML-ready image patches tiled into 224 × 224 samples
• Ecological metadata including geographic location, climate zone, and forest type
• Benchmark settings for evaluating in-domain performance and domain generalization

Supported Tasks

TreeFinder supports the following machine learning tasks:

• Individual dead tree segmentation
• Binary semantic segmentation
• Remote sensing image segmentation
• Domain generalization across ecological and geographic conditions
• Benchmarking foundation models and task-specific segmentation models for environmental monitoring

Dataset Structure

Each sample contains a high-resolution NAIP image patch and a corresponding binary segmentation mask.

The imagery contains four channels:

Red, Green, Blue, Near-Infrared

The segmentation mask is binary:

0 = background
1 = dead tree

The associated metadata is available in the .csv file.

Data Sources

TreeFinder is derived from high-resolution NAIP aerial imagery covering forested regions across CONUS. Dead tree annotations were manually created and validated using expert interpretation and multi-temporal image comparison.

The dataset is intended to provide a reproducible benchmark for evaluating whether machine learning models can detect fine-scale tree mortality patterns from aerial imagery.

Dataset Splits

TreeFinder provides ML-ready splits for model training and evaluation. The dataset supports both standard supervised learning and domain generalization evaluation.

Recommended split types include:

• Random split: standard train/validation/test evaluation
• Cross-region split: training and testing across different U.S. regions
• Cross-climate split: evaluation under Köppen–Geiger climate domain shifts
• Cross-forest-type split: evaluation across dominant forest type groups

Please refer to the accompanying benchmark configuration scripts for the exact split definitions in our codebase.

Metadata

Each image patch is enriched with ecological and geographic metadata, including:

• Geographic coordinates
• U.S. state
• Site identifier
• Köppen–Geiger climate zone
• Primary tree type derived from USDA Forest Service maps

These metadata enable controlled evaluation under realistic ecological domain shifts, such as:

• East-to-West transfer
• Humid-to-arid climate transfer
• Conifer-to-broadleaf forest transfer
• Cross-biome generalization

Benchmark Models

TreeFinder includes benchmark results from representative semantic segmentation models, including both task-specific architectures and foundation-model-based approaches.

Evaluated models include:

• U-Net
• DeepLabV3+
• ViT-based segmentation model
• SegFormer
• Mask2Former
• DOFA, a multimodal foundation model trained on satellite data

These models are evaluated under both in-domain and out-of-domain settings to assess robustness across geographic, climatic, and forest-type shifts.

Evaluation

TreeFinder is designed for binary semantic segmentation. Recommended evaluation metrics include:

• Intersection over Union (IoU)
• F1 score / Dice coefficient
• Precision
• Recall
• Pixel accuracy

Because dead trees are sparse relative to background pixels, users are encouraged to report class-sensitive metrics such as IoU, F1, precision, and recall rather than relying only on overall accuracy.

Intended Uses

TreeFinder is intended for research and benchmarking in:

• Individual tree mortality mapping
• Forest health monitoring
• Remote sensing image segmentation
• Ecological disturbance assessment
• Carbon cycle and forest carbon monitoring
• Wildfire risk-related vegetation analysis
• Domain generalization for environmental computer vision
• Evaluation of remote sensing foundation models

Ethical and Environmental Considerations

TreeFinder does not contain personal or sensitive human information. The dataset is based on aerial imagery of forested landscapes and is intended for environmental monitoring and scientific research.

Potential positive impacts include improved forest health assessment, better ecological disturbance monitoring, and support for carbon and wildfire-related research. Potential risks include misuse of model outputs for unsupported operational decisions if uncertainty, domain shift, or model limitations are ignored.

How to Use

Example loading workflow:

from datasets import load_dataset

dataset = load_dataset("zhwang1/TreeFinder")
print(dataset)
print(dataset["train"][0])

The GitHub repository is available: https://github.com/zhwang0/treefinder.

Citation

If you use TreeFinder in your research, please cite:

@inproceedings{wang2025treefinder,
  title     = {TreeFinder: A US-Scale Benchmark Dataset for Individual Tree Mortality Monitoring Using High-Resolution Aerial Imagery},
  author    = {Wang, Zhihao and Li, Cooper and Wang, Ruichen and Ma, Lei and Hurtt, George and Jia, Xiaowei and Mai, Gengchen and Li, Zhili and Xie, Yiqun},
  booktitle = {Proceedings of the 39th Conference on Neural Information Processing Systems (NeurIPS 2025), Datasets and Benchmarks Track},
  year      = {2025}
}

Dataset Contact

For questions, issues, or collaboration inquiries, please open an issue in the associated GitHub repository or contact the dataset authors.

Acknowledgements

TreeFinder was developed to support scalable, reproducible, and ecologically grounded machine learning research for forest mortality monitoring. We thank the contributors, annotators, and collaborators who supported dataset development, validation, and benchmarking.