Research map / Detection and segmentation / Open-set and few-shot detection
OAM-TCD: A globally diverse dataset of high-resolution tree cover maps
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Releases OAM-TCD, a globally diverse, open-access dataset of 5072 10 cm/px aerial tiles with instance masks for trees and tree groups, plus baseline UNet, SegFormer and Mask-RCNN tree segmentation models.
Problem
Individual tree crown delineation and tree cover mapping at sub-meter resolution lacks open-access, geographically diverse instance-level annotated datasets.
Approach
Builds a biome-stratified, globally sampled dataset from OpenAerialMap imagery with human-labelled individual tree and canopy-group polygons, and trains reference semantic (UNet, SegFormer) and instance (Mask-RCNN) models that are evaluated via k-fold CV, a holdout split, and out-of-domain data from Zurich and Tonga.
Key results
- SegFormer mit-b5 reached 0.887 IoU in 5-fold cross-validation and 0.876 IoU on the holdout set, with larger SegFormer variants giving only limited gains.
- Mask-RCNN R50 achieved mAP50 of 41.79 in cross-validation and 43.22 on the holdout set.
- On Zurich, SegFormer mit-b5 reached IoU 0.791 against the LIDAR CHM, and instance recall was 0.36 for tree keypoints and 0.73 including canopy regions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SegFormer instead of UNet for semantic segmentation | OAM-TCD semantic segmentation · OAM-TCD | IoU | Helped |
| Increase SegFormer model size from mit-b0 through mit-b5 | OAM-TCD semantic segmentation · OAM-TCD | IoU: 0.858$\pm$0.011 to \textbf{0.887$\pm$0.003}; holdout 0.865 to \textbf{0.876} | Mixed |
| Use transformer-based successors to Mask-RCNN (DETR-family or Mask2Former) instead of Mask-RCNN | OAM-TCD instance segmentation · OAM-TCD | mAP50 | No clear effect |
Surprising findings
- Transformer-based Mask2Former and DETR were not reliably better than Mask-RCNN and showed poorer convergence and RAM issues with many objects.
- Increasing SegFormer size beyond mit-b1/b2 gave limited performance gains, suggesting dataset annotation consistency and diversity limits.
Limitations
- Annotations are inconsistent (non-ecologist annotators, individual vs. group decisions) and biomes such as boreal, tundra and flooded grasslands are under-represented.
- Dense closed-canopy forests are not individually annotated, so instance models underperform there.
- No extensive validation against in-situ field data was performed.
Experimental setup
Datasets: OAM-TCD MS-COCO ImageNet Swisstopo SWISSIMAGE 2022 (Zurich) WeRobotics Open AI Challenge (Tonga)
Benchmarks: OAM-TCD holdout Zurich urban tree inventory and LIDAR CHM WeRobotics Open AI Challenge (Tonga)
Metrics: IoU accuracy F1 mAP50 recall
Baselines: DeepForest Mask2Former DETR
Builds on: Mask-RCNN SegFormer UNet Detectron2
Closest papers by meaning
- Bringing SAM to new heights: leveraging elevation data for tree crown segmentation from drone imagery (NeurIPS 2025 · similarity 0.88)
- Open-Canopy: Towards Very High Resolution Forest Monitoring (CVPR 2025 · similarity 0.86)
- FoMo: Multi-Modal, Multi-Scale and Multi-Task Remote Sensing Foundation Models for Forest Monitoring (AAAI 2025 · similarity 0.86)
- DUNIA: Pixel-Sized Embeddings via Cross-Modal Alignment for Earth Observation Applications (ICML 2025 · similarity 0.86)
- GTPBD: A Fine-Grained Global Terraced Parcel and Boundary Dataset (NeurIPS 2025 · similarity 0.85)
- Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.