Research map / Time series / Causal and generative time series
Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation
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 3D U-Net that ingests 12-month Sentinel-2 monthly stacks plus a Sentinel-1 composite, trained on sparse GEDI LiDAR labels, to produce a 10 m temporal canopy height map of Europe for 2019–2022.
Problem
Large-scale, high-resolution canopy height mapping over multiple years is lacking; prior maps are single-year or coarse, and median composites discard seasonal and geolocation-shift information needed for accurate and temporally consistent height estimates.
Approach
Replace temporal median composites with a full 12-month Sentinel-2 monthly stack processed by a 3D U-Net (temporal dimension collapsed through the network), trained on multi-year GEDI labels with a shift-tolerant Huber loss, and smoothed with a quadratic spline for temporal consistency.
Key results
- The final model reaches MAE 4.76 m and MSE 74.28 m^2 for 2020 with R^2_7 0.591, versus 5.46 m MAE for the next-best existing map (a 13% improvement).
- Multi-year 3D-Stack model gives average MAE 4.95 m across 2019-2022 versus 6.02 m for the 2D-Composite-2020 single-year baseline.
- Against Tolan et al. across 2019-2021, the model reduces MAE by about 61% (average 4.69 m vs 12.05 m).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a 12-month Sentinel-2 stack instead of a median composite | Canopy-height estimation model · Europe GEDI validation, 2020 | MAE [m]: 5.66 -> 5.51 | Helped |
| Train on multiple years rather than 2020 only | 3D-Stack canopy-height model · Europe GEDI validation, 2020 | MAE [m]: 5.48 -> 5.05 | Helped |
| Use 3D temporal convolutions rather than a 2D monthly stack | Multi-year stack canopy-height model · Europe GEDI validation, 2020 | MAE [m]: 5.17 -> 5.05 | Helped |
| Use the full-data 3D-Stack-MultiYear-L model instead of the next-best map baseline | European canopy-height mapping · Europe GEDI validation, 2020 | MAE [m]: 5.46 -> 4.76 | Helped |
| Use the proposed model instead of the multi-year map baseline | European canopy-height mapping · Europe GEDI validation, 2019–2021 | MAE [m]: 12.38 -> 4.77; 12.39 -> 4.76; 11.37 -> 4.53 | Helped |
| Remove Sentinel-2 bands B01 and B09 (ablation) | 3D-Stack-MultiYear-L · Europe GEDI validation | L1 (m): 1.992 ± 0.003 -> 1.991 ± 0.002 | No clear effect |
| Use only winter months (Nov–Feb) instead of mixed seasons (ablation) | Canopy-height model · Europe GEDI validation | Huber Loss (m): 1.122 ± 0.002 -> 1.169 ± 0.003 | Hurt |
| Use only summer months (Jun–Sep) instead of mixed seasons (ablation) | Canopy-height model · Europe GEDI validation | Huber Loss (m): 1.122 ± 0.002 -> 1.130 ± 0.002 | Hurt |
| Combine winter and summer months (Jan–Feb, Aug–Sep) rather than using either season alone (ablation) | Canopy-height model · Europe GEDI validation | Huber Loss (m): 1.169 ± 0.003 -> 1.122 ± 0.002; 1.130 ± 0.002 -> 1.122 ± 0.002 | Helped |
Surprising findings
- The best single-year model (3D-Stack-2020) only matches the weakest multi-year model (2D-Composite-MultiYear) in average MAE.
- Ablation shows removing Sentinel-2 bands B01 and B09 has negligible effect on performance.
Limitations
- Sparse GEDI labels (about 0.15% of pixels per patch) and no labels above 51.6N limit coverage and supervision.
- Minor tree growth is hard to detect over the short 2019-2022 window; the model mainly captures deforestation.
- Lower spatial detail than very high-resolution maps (e.g., Maxar/Planet-based) in some visual comparisons.
Experimental setup
Datasets: Sentinel-1 Sentinel-2 GEDI L2A Copernicus Land Monitoring Service Forest Type Map
Benchmarks: GEDI-labeled validation locations across Europe (1,500 sites) ALS data for tall-tree patches
Metrics: MAE MSE R^2 R^2_7
Baselines: Lang et al. Liu et al. Pauls et al. Tolan et al. Turubanova et al. 2D-Stack-2020 2D-Composite-2020 3D-Stack-2020
Builds on: U-Net Huber loss Adam
Closest papers by meaning
- Estimating Canopy Height at Scale (ICML 2024 · similarity 0.91)
- Open-Canopy: Towards Very High Resolution Forest Monitoring (CVPR 2025 · similarity 0.91)
- 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 0.83)
- DUNIA: Pixel-Sized Embeddings via Cross-Modal Alignment for Earth Observation Applications (ICML 2025 · similarity 0.81)
- GeoPlant: Spatial Plant Species Prediction Dataset (NeurIPS 2024 · 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.