Research map / Detection and segmentation / Open-set and few-shot detection
Open-Canopy: Towards Very High Resolution Forest Monitoring
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.
Introduces Open-Canopy, an open-access country-scale benchmark pairing 1.5 m SPOT 6-7 imagery with LiDAR-HD canopy height over France, plus Open-Canopy-∆ for canopy height reduction detection, and benchmarks vision backbones on both.
Problem
Canopy height estimation and change monitoring at meter-level resolution from satellite imagery lacks open-access datasets, hindering reproducibility and fair evaluation of computer vision models.
Approach
Builds a fully open-access, country-scale VHR canopy height dataset from French national SPOT 6-7 imagery (1.5 m) and LiDAR-HD ALS-derived heights, and benchmarks many modern backbones adapted to 4-band input, plus a consecutive-year height-loss segmentation subset.
Key results
- PVTv2 trained on Open-Canopy reaches MAE 2.52 m, nMAE 22.9%, and tree cover IoU 90.5%, outperforming existing maps such as Liu et al. (MAE 4.83 m) and Tolan et al. (MAE 5.07 m).
- On Open-Canopy-∆, PVTv2 achieves F1 54.1% and IoU 37.0% for canopy height loss masks, versus 6.0% F1 for Schwartz et al. and 1.7% for Global Forest Change.
- Hierarchical ViTs (SWIN, PCPVT, PVTv2) and UNet/HViT outperform plain ViT-B; ImageNet-pretrained models beat CLIP and DINOv2 pretraining.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use hierarchical PVTv2 instead of ViT-B | Canopy height estimation model · Open-Canopy test | MAE: 4.26 -> 2.52 | Helped |
| Use PVTv2 instead of UNet | Open-Canopy canopy height estimation · Open-Canopy test | MAE: 2.67 -> 2.52 | Helped |
| Use ImageNet pretraining instead of DINOv2 pretraining | ViT-B canopy height model · Open-Canopy test | MAE: 4.84 -> 2.87 | Helped |
| Train on Open-Canopy and apply to SPOT 6-7 rather than NAIP imagery | PVTv2 canopy height model · Utah out-of-domain | MAE: 4.38 -> 2.08 | Helped |
| Use PVTv2 instead of Schwartz height maps | Canopy height reduction detection · Open-Canopy-∆ Chantilly | F1 score: 6.0 -> 54.1 | Helped |
| Use PVTv2 instead of GFC | Canopy height reduction detection · Open-Canopy-∆ Chantilly | F1 score: 1.7 -> 54.1 | Helped |
| Add the proposed initialization scheme instead of fully random initialization (ablation) | PVTv2 with ImageNet1k pretraining · Open-Canopy test | MAE: 11.17 -> 2.52 | Helped |
| Add the proposed initialization scheme instead of LoRA rank 4 (ablation) | PVTv2 with ImageNet1k pretraining · Open-Canopy test | MAE: 4.54 -> 2.52 | Helped |
| Add the proposed initialization scheme instead of random first-layer initialization (ablation) | PVTv2 with ImageNet1k pretraining · Open-Canopy test | MAE: 2.87 -> 2.52 | Helped |
Surprising findings
- Foundation models pretrained on large natural-image datasets (CLIP, DINOv2) and satellite-pretrained ScaleMAE generalize poorly to canopy height estimation, underperforming ImageNet-pretrained CNNs and hierarchical ViTs.
- Plain ViT-B underperforms convolutional UNet and hierarchical transformers, contrary to trends in natural image analysis.
Limitations
- Open-access constraints limit spatial and temporal resolution; reproducing the dataset with commercial data would cost millions of dollars.
- Coverage is limited to metropolitan France and lacks forest types such as rainforests, which may hurt generalization.
- ALS ground truth has errors (e.g., multiple echoes) and LiDAR spans all seasons while SPOT images are from spring and summer.
Experimental setup
Datasets: Open-Canopy LiDAR-HD ALS point clouds SPOT 6-7 imagery (DINAMIS) NAIP NEON REDB ALS
Benchmarks: Open-Canopy Open-Canopy-∆
Metrics: MAE nMAE RMSE Bias tree cover IoU Precision Recall F1 score IoU
Baselines: UNet DeepLabv3 ViT-B HViT PCPVT Swin ScaleMAE DINOv2 ViT-B CLIP ViT-B Tolan et al. height map Liu et al. height map Schwartz et al. height map
Builds on: UNet Vision Transformer PVTv2 Swin Transformer Tolan et al. canopy height model
Closest papers by meaning
- Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation (ICML 2025 · similarity 0.91)
- Estimating Canopy Height at Scale (ICML 2024 · similarity 0.89)
- OAM-TCD: A globally diverse dataset of high-resolution tree cover maps (NeurIPS 2024 · similarity 0.86)
- FoMo: Multi-Modal, Multi-Scale and Multi-Task Remote Sensing Foundation Models for Forest Monitoring (AAAI 2025 · similarity 0.86)
- ForestLPR: LiDAR Place Recognition in Forests Attentioning Multiple BEV Density Images (CVPR 2025 · similarity 0.84)
- Bringing SAM to new heights: leveraging elevation data for tree crown segmentation from drone imagery (NeurIPS 2025 · similarity 0.83)
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.