Research map / Label-efficient and robust learning / Recognition and long-tail
Estimating Canopy Height at Scale
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.
Builds a global 10 m canopy height map from Sentinel-1/2 composites and GEDI LiDAR labels using a U-Net with a shift-resilient loss for GEDI geolocation errors and SRTM slope filtering.
Problem
Global-scale canopy height estimation from satellite imagery is lower quality than regional products, and sparse GEDI LiDAR labels suffer from geolocation shifts and slope-induced height errors in mountainous terrain.
Approach
A shift-resilient loss that, per GEDI track, minimizes the pixel-wise loss over allowed systematic shifts within a radius, combined with SRTM slope filtering of labels and cloud-filtered Sentinel-2 composites, trained with a ResNet50-backbone U-Net.
Key results
- Overall MAE / RMSE of 2.43 m / 4.73 m versus 6.47 m / 8.62 m (Lang et al.) and 6.92 m / 9.25 m (Potapov et al.).
- For GEDI measurements above 5 m, MAE / RMSE of 4.45 m / 6.72 m, compared with 8.80 m / 11.02 m and 10.01 m / 12.43 m for the two global maps.
- On the Needle Mountains ALS ablation, SRTM filtering reduces MAE from 9.77 m to 7.33 m.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed canopy-height framework instead of existing global maps | Global canopy height estimation · Test area, no filter | MAE: 6.47 -> 2.43; 6.92 -> 2.43 | Helped |
| Use the proposed canopy-height framework instead of existing global maps | Global canopy height estimation · Test area, GEDI height >5 m | MAE: 8.80 -> 4.45; 10.01 -> 4.45 | Helped |
| Add SRTM slope filtering to exclude unreliable GEDI labels (ablation) | Canopy height model · Needle Mountains ALS | MAE (m): 9.77 -> 7.33 | Helped |
| Add shifted loss to the pixel-wise Huber training loss (ablation) | U-Net, ResNet50 · Validation set | Huber loss: 1.04 ± 0.01 -> 0.99 ± 0.01 | Helped |
| Use ImageNet-pretrained instead of randomly initialized backbones (ablation) | Examined segmentation architectures · Validation set | Shifted Huber loss: Across architectures, pretrained values are higher; e.g., U-Net 1.01 -> 1.10 | Hurt |
| Replace U-Net with alternative segmentation architectures (ablation) | Canopy height model · Validation set | Shifted Huber loss: Not provided | Mixed |
| Use cyclical learning-rate schedulers instead of the linear scheduler (ablation) | U-Net, ResNet50 · Validation set | Shifted Huber loss: 1.014 (linear with AD); 1.022, 1.012, 1.010, 1.004 (without AD); 1.016, 1.060, 1.026, 1.053 (with AD) | Mixed |
| Ensemble six models instead of using a single model (ablation) | Canopy height model · Test set | MAE: 4.03 -> 3.94 | Helped |
| Replace ResNet50 with ResNet18, trained from scratch (ablation) | U-Net · Validation set | Shifted Huber loss: 1.017 -> 1.051 | Hurt |
| Replace ResNet50 with ResNet101, trained from scratch (ablation) | U-Net · Validation set | Shifted Huber loss: 1.017 -> 1.020 | No clear effect |
| Replace the pixel-wise training loss among L1, L2, and Huber (ablation) | Canopy height model · Validation set | Huber loss: Not provided | Mixed |
Surprising findings
- Using ImageNet-pretrained backbones decreased validation performance across all tested architectures, so models were trained from scratch.
- The standard U-Net outperformed more complex segmentation architectures such as DeepLabV3 and PSPNet for canopy height.
Limitations
- SRTM captures surface rather than terrain elevation, so the slope filter is an imperfect fix in mountainous regions.
- Underestimation of high canopy heights (above 30 m) persists, and for heights above 10 m the map falls 5-10 m short of the other global maps.
- Northern regions are not covered by the GEDI training data due to the ISS orbit.
Experimental setup
Datasets: Sentinel-1 GRD Sentinel-2 BOA GEDI RH100 SRTM ALS data (Needle Mountains, ablation)
Benchmarks: GEDI held-out test set (non-shifted)
Metrics: MAE MSE RMSE RRMSE MAPE
Baselines: Potapov et al. global height map Lang et al. global height map regional France height map
Builds on: U-Net GEDI AdamW
Closest papers by meaning
- Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation (ICML 2025 · similarity 0.91)
- Open-Canopy: Towards Very High Resolution Forest Monitoring (CVPR 2025 · similarity 0.89)
- FoMo: Multi-Modal, Multi-Scale and Multi-Task Remote Sensing Foundation Models for Forest Monitoring (AAAI 2025 · similarity 0.84)
- OAM-TCD: A globally diverse dataset of high-resolution tree cover maps (NeurIPS 2024 · similarity 0.81)
- Feedforward Few-shot Species Range Estimation (ICML 2025 · similarity 0.80)
- DUNIA: Pixel-Sized Embeddings via Cross-Modal Alignment for Earth Observation Applications (ICML 2025 · similarity 0.80)
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.