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

Accepted to NeurIPS 2025 Datasets & Benchmarks Track

• Paper is available on OpenReview: https://openreview.net/pdf?id=BH2miB6Smc.
• Project page is available on NeurIPS: https://neurips.cc/virtual/2025/loc/san-diego/poster/121794.
• Datasets are available on Kaggle: https://www.kaggle.com/datasets/zhihaow/tree-finder.

🧩 Overview

TreeFinder provides a flexible PyTorch-based pipeline for training and evaluating semantic segmentation models on hand-labeled dead-tree imagery.
It supports custom dataset splits, config-driven experiments, and benchmarking under multiple generalization scenarios, aligned with the official TreeFinder dataset.

🚀 Features

• Multiple dataset splits — random (80–10–10), state-based, and scenario-driven (climate / forest type)
• Consistent tile loading — RGB, NIR, NDVI, and no-data mask handling
• Augmentations — flips and 90° rotations for per-tile diversity
• Config-driven experiments — all controlled via a single YAML file
• Model zoo — U-Net, DeepLabV3+, ViT-Seg, SegFormer, Mask2Former, and DOFA
• Robust training loop — masked BCE loss, early stopping, checkpointing, and LR scheduling
• Comprehensive evaluation — IoU, F1, precision, recall, accuracy (micro/macro, all vs. positive-only)
• Automated logging — loss curves, config dumps, and YAML summaries for every run

🧠 Benchmark Models

Model	Type	Description
U-Net	CNN	📄Paper Classic encoder–decoder architecture with skip connections. Trained from scratch as a baseline.
DeepLabV3+	CNN	📄Paper Employs atrous spatial pyramid pooling with a ResNet-50 backbone.
ViT-Seg	Vision Transformer	📄Paper Patch-based transformer with a transposed-convolution decoder.
SegFormer	Vision Transformer	📄Paper Hierarchical transformer with efficient multi-scale feature fusion.
Mask2Former	Vision Transformer	📄Paper Set prediction model using masked attention (Swin-T backbone).
DOFA	Foundation Model	📄Paper Multimodal foundation model pretrained on multi-sensor remote-sensing imagery.

All models are trained with consistent hyperparameters and evaluated under cross-region, cross-climate, and cross-forest-type generalization setups from the TreeFinder benchmark.

📁 Repository Structure

├── main.py                   # Entrypoint: parse args, load config, run train & eval  
├── configs/                  # YAML experiment configs  
│   ├── debug.yaml  
│   └── benchmark.yaml  
├── data_loader/              # Tile loading & split implementations  
│   ├── __init__.py           # Dispatch get_dataloader(cfg)  
│   ├── utils.py              # load_tile(), augment_tile()  
│   ├── random_split.py  
│   ├── by_state_split.py  
│   ├── by_climate_split.py  
│   └── by_tree_split.py  
├── models/                   # Model factory & builders  
│   ├── __init__.py  
│   ├── unet.py  
│   ├── deeplab.py  
│   ├── vit.py  
│   ├── segformer.py  
│   ├── mask2former.py  
│   └── dofa/                 
├── exps/                     # Training & evaluation loops  
│   ├── train.py  
│   └── evaluate.py  
└── utils/                    # Misc helpers (logging, config I/O, seed control)  
    └── tools.py  

⚙️ Configuration and ▶️ Quickstart

All experiment parameters (model, optimizer, paths, augmentations, split type) are defined in YAML under configs/.
For example:

python main.py --config configs/debug.yaml

📚 Citation

If you use this repository or the TreeFinder dataset, please cite:

Zhihao Wang, Cooper Li, Ruichen Wang, Lei Ma, George Hurtt, Xiaowei Jia, Gengchen Mai, Zhili Li, Yiqun Xie.
TreeFinder: A US-Scale Benchmark Dataset for Individual Tree Mortality Monitoring Using High-Resolution Aerial Imagery.
In Proceedings of the 39th Conference on Neural Information Processing Systems (NeurIPS), 2025.

📬 Contact

For questions or feedback, feel free to reach out:

• Zhihao Wang — zhwang1@umd.edu
• Yiqun Xie — xie@umd.edu