CHMv2: Improvements in Global Canopy Height Mapping using DINOv3
Authors: John Brandt, Seungeun Yi, Jamie Tolan, Xinyuan Li, Peter Potapov, Jessica Ertel, Justine Spore, Huy V. Vo, Michaël Ramamonjisoa, Patrick Labatut, Piotr Bojanowski, Camille Couprie
Organizations: World Resources Institute, 10 G St NE #800, Washington, DC 20002, USA · Fundamental AI Research (FAIR), Meta, 75002 Paris, France · Meta, 1 Hacker Way, Menlo Park, CA 94025, USA · University of Maryland, Department of Geography, College Park, MD 20742, USA
Accurate canopy height information is essential for quantifying forest carbon, monitoring restoration and degradation, and assessing habitat structure, yet high-fidelity measurements from airborne laser scanning (ALS) remain unevenly available globally. Here we present CHMv2, a global, meter-resolution canopy height map derived from high-resolution optical satellite imagery using a depth-estimation model built on DINOv3 and trained against ALS canopy height models. Compared to existing products, CHMv2 substantially improves accuracy, reduces bias in tall forests, and better preserves fine-scale structure such as canopy edges and gaps. These gains are enabled by a large expansion of geographically diverse training data, automated data curation and registration, and a loss formulation and data sampling strategy tailored to canopy height distributions. We validate CHMv2 against independent ALS test sets and against tens of millions of GEDI and ICESat-2 observations, demonstrating consistent performance across major forest biomes.
Figures & tables
Figure 1: Visual improvements from CHMv1 to CHMv2 in a disturbance area in the Amazon (top), an urban forest in Central Java, Indonesia (middle) and a plantation in the Ghanaian cocoa belt (bottom).
Figure 2: Training and testing/validation data locations for NAIP-3DEP and SatLidar datasets.
Figure 3: Data registration methodology and analysis.
Figure 4: Compared to CHMv1 on NAIP-3DEP (top) and SatLidar v2 (bottom), CHMv2 exhibits greatly reduced biases, especially for trees above 30 m. The inset figures plot the 95th percentiles of CHM on 50×50 crops ( y axis) as a function of the 95th percentiles of Ground Truths ( x axis). Scales are in meters.
Figure 5: Qualitative improvements over CHMv1, in terms of sharpness and accuracy. Comparison on the NAIP-3DEP (two top lines) and SatLidar v2 (bottom lines) on 256 × 256 samples. CHMv1 is without GEDI correction.
Figure 6: Comparison of CHMv1 and CHMv2 on the Multisource dataset. CHMv1 is without GEDI correction.
Figure 7: Canopy height profiles along a single 10-m wide transect comparing ALS reference measurements to five canopy height datasets. Gray points show ALS returns within the transect corridor; the black line shows the ALS-derived canopy height profile. ALS data is from Kalimantan, Indonesia (1.6412 ° , 115.281 ° ) and is ≥200 km from any observation within the SatLidar v2 dataset.
Dataset
Resolution (m)
MAE
Pauls et al. [ 39 ]
10
7.5
Lang et al. [ 26 ]
10
8.4
Potapov et al. [ 40 ]
30
4.9
CHMv1 [ 50 ]
1
4.3
CHMv2
1
3.0
Table 1: Comparison of global canopy height products to ground truth ALS-derived CHM data in the SatLidar v2 test split.
Figure 8: Comparison with GEDI and ICESat-2.
Backbone
Setting
Loss
Sampling
Train data
SatLidar v2
NAIP-3DEP
MAE
R2
Bias
Edge
MAE
R2
Edge
CHMv1
Base
SiLog
Uniform
Neon train
4.3
0.53
2.6
0.67
2.2
0.54
0.77
DINOv3
Base
SiLog
Uniform
Neon train
3.9
0.64
2.2
0.66
2.0
0.59
0.72
DINOv3
Base
SiLog
Uniform
SatLidar v1 train
3.5
0.78
0.9
0.62
2.0
0.46
0.63
DINOv3
Base
SiLog
Uniform
SatLidar v2 train
3.5
0.76
0.5
0.58
2.1
0.52
0.63
DINOv3
Final
Cur.
Uniform
SatLidar v2 train
3.5
0.75
-0.5
0.55
2.2
0.48
0.58
Table 2: Main ablation from the CHMv1 settings to the CHMv2 settings, demonstrating improvements due to the DINOv3 backbone, training data diversity, registration, new architecture and loss. The base setting corresponds to the CHMv1 architecture and training parameters, and the final setting to the changes described in Section 2.6 .
Figure 9: Impact of our loss, a Combination of SiLog, Charbonnier and Multi-scale Patch Gradient on the NAIP-3DEP results. Compared to SiLog loss, results are sharper and more accurate.
Table 12Table 13
Category sampling
NAIP-3DEP
SatLidar v2
SatLidar v2
NAIP-3DEP
Multisource
MAE
R2
MAE
R2
MAE
R2
✗
50 %
50 %
3.07
0.83
1.37
0.93
2.98
0.77
✓
50 %
50 %
3.07
0.84
1.36
0.93
2.98
0.78
✓
30 %
70 %
3.10
0.83
1.38
0.93
2.94
0.82
✓
70 %
30 %
3.18
0.82
1.35
0.93
2.98
0.79
Table 7: Data mix ablation: using a 50-50 ratio of data source brings a good compromise. Category sampling shows slightly improved metrics. To complement this ablation, we computed that the category sampling improves the MAE of trees above 35 meters by at least two meters. Using DINOv3 viT-L backbone, Charbonnier + Gradient Loss.
Forest monitoring is critical for climate change mitigation. However, existing global tree height maps provide only static snapshots and do not capture temporal forest dynamics, which are essential for accurate carbon accounting. We introduce ECHOSAT, a global and temporally consistent tree height map at 10 m resolution spanning multiple years. To this end, we resort to multi-sensor satellite data to train a specialized vision transformer model, which performs pixel-level temporal regression. A self-supervised growth loss regularizes the predictions to follow growth curves that are in line with natural tree development, including gradual height increases over time, but also abrupt declines due to forest loss events such as fires. Our experimental evaluation shows that our model improves state-of-the-art accuracies in the context of single-year predictions. We also provide the first global-scale height map that accurately quantifies tree growth and disturbances over time. We expect ECHOSAT to advance global efforts in carbon monitoring and disturbance assessment. The maps can be accessed at https://github.com/ai4forest/echosat.
Jan Pauls, Karsten Schrödter, Sven Ligensa +7
Department of Information Systems, University of Münster, Germany · Laboratoire des Sciences du Climat et de l’Environnement (LSCE), France · Zuse Institute Berlin (ZIB), Germany +2
High-resolution mapping of canopy height is essential for forest management and biodiversity monitoring. Although recent studies have led to the advent of deep learning methods using satellite imagery to predict height maps, these approaches often face a trade-off between data accessibility and spatial resolution. To overcome these limitations, we present SERA-H, an end-to-end model combining a super-resolution module (EDSR) and temporal attention encoding (UTAE). Trained under the supervision of high-density LiDAR-derived Canopy Height Models (CHM), our model generates 2.5 m resolution height maps from freely available Sentinel-1 and Sentinel-2 (10 m) time series data. Evaluated against an open-source benchmark dataset in France, SERA-H, with a MAE of 2.6 m and R2 of 0.82, not only outperforms standard Sentinel-1/2 baselines but also approaches the accuracy of methods based on commercial very high-resolution imagery (e.g., SPOT-6/7), while relying solely on freely available data. These results demonstrate that combining high-resolution ALS supervision with the spatio-temporal information embedded in Sentinel time series enables the reconstruction of spatial detail finer than the native resolution of the input imagery. By approaching the accuracy of costly commercial imagery, SERA-H opens the possibility of mapping temperate forests using publicly available data with high revisit frequency.
Thomas Boudras, Martin Schwartz, Rasmus Fensholt +6
Laboratoire des Sciences du Climat et de l’Environnement (LSCE), CEA, CNRS, UVSQ, Université Paris-Saclay, Gif-sur-Yvette, France · Department of Geography and Geology, University of Copenhagen, Øster Voldgade 10, Copenhagen, DK-1350, Denmark · Kayros SAS, Paris, 75009, France +2
Monitoring canopy height change is essential for understanding carbon sinks and forest dynamics. Remote sensing enables consistent, large-scale observations of such changes, increasingly integrated with deep learning architectures such as Geospatial Foundation Models (GFMs). However, existing methods and datasets frame the problem as binary change detection, which overlooks both the continuous nature of change, especially for vegetation, and the inherent uncertainty in labels. We present the Canopy Height Change (CHC) dataset, providing 3 m resolution continuous canopy height differences and associated spatially resolved uncertainties across 10598 km2 of northern and western Spain. The dataset is paired with a co-located time series of PlanetScope satellite imagery. Based on the dataset, we introduce the task of uncertainty-aware change regression, associated metrics and strategies for fine-tuning GFMs. Furthermore, we evaluate state-of-the-art GFMs and highlight promising directions and remaining challenges for advancing continuous canopy height change estimation.
Max Gaber, Dimitri Gominski, Jaime C. Revenga +3
Department of Geosciences and Natural Resource Management, University of Copenhagen · Department of Mathematical Sciences, University of Copenhagen · Department of Computer Science and Electrical Engineering, University of Rostock +1