Estimating Individual Tree Height and Species from UAV Imagery

Jannik Endres1, Etienne Laliberté2,1, David Rolnick1,3, Arthur Ouaknine1,3
1Mila – Quebec AI Institute, 2Université de Montréal, 3McGill University
Paper (Coming Soon) arXiv (Coming Soon) Code Hugging Face logo Benchmark (Coming Soon) Hugging Face logo Model Weights (Coming Soon)
Teaser figure

TL;DR: BIRCH-Trees is the first benchmark for joint individual tree height estimation and species identification from UAV images. Our model, DINOvTree, leverages a VFM to extract features and predicts the height and species of the center tree in the input image with two separate heads.

Abstract

Accurate estimation of forest biomass, a major carbon sink, relies heavily on tree-level traits such as height and species. Unoccupied Aerial Vehicles (UAVs) capturing high-resolution imagery from a single RGB camera offer a cost-effective and scalable approach for mapping and measuring individual trees.

We introduce BIRCH-Trees, the first benchmark for individual tree height and species estimation from tree-centered UAV images, spanning three datasets: temperate forests, tropical forests, and boreal plantations.

We also present DINOvTree, a unified approach using a Vision Foundation Model (VFM) backbone with task-specific heads for simultaneous height and species prediction. Through extensive evaluations on BIRCH-Trees, we compare DINOvTree against commonly used vision methods, including VFMs, as well as biological allometric equations. We find that DINOvTree achieves top overall results with accurate height predictions and competitive classification accuracy while using only 54% to 58% of the parameters of the second-best approach.

BIRCH-Trees

BIRCH-Trees (Benchmark for Individual Recognition of Class and Height of Trees) is the first benchmark for individual tree height and species estimation from tree-centered UAV images. We formulate individual tree height estimation and species identification as regression and classification tasks, respectively. We provide high-resolution RGB drone images, alongside corresponding Digital Surface Models (DSMs), of tree canopies from three distinct environments:

  • Temperate forest: Quebec Trees
  • Tropical forest: Barro Colorado Island (BCI)
  • Boreal plantation: Quebec Plantations

Quebec Trees

Representative examples from the Quebec Trees dataset test split, illustrating one example per class alongside its corresponding ground truth class and height.

BCI

Representative examples from the BCI dataset test split, illustrating one example per class alongside its corresponding ground truth class and height.

Quebec Plantations

Representative examples from the Quebec Plantations dataset test split, illustrating one example per class alongside its corresponding ground truth class and height.

DINOvTree

DINOvTree method figure
A shared VFM (blue) extracts features from an RGB image. In the height estimation head (orange), a learnable height query cross-attends to adapted patch tokens to predict \( \hat{h} \). In the classification head (yellow), a learnable classification query cross-attends to adapted patch tokens to obtain a classification token. We then concatenate it with the VFM [CLS] token to predict \( \hat{\mathbf{p}} \). The total loss \( \mathcal{L} \) is the weighted sum of the height estimation loss \( \mathcal{L}_H(h, \hat{h}) \) and the classification loss \( \mathcal{L}_S(\hat{\mathbf{p}}, c) \) supervised by ground truth height \( h \) and class \( c \).

Results

We compare our DINOvTree against competing methods in terms of parameter count, height estimation, and classification on the three datasets of BIRCH-Trees. We group models by parameter count: fewer than 300M (top) and more than 300M (bottom). For competing methods, 'Param.' denotes combined parameters of two independently-trained task-specific models. We report the mean ± standard error over 5 seeds. Bold and underline denote best and second-best. * indicates an oracle with ground truth class and segmentation; † denotes the Mask R-CNN variant from Hao et al. and Fu et al.; ❄ indicates a frozen backbone.

Quebec Trees

Method Param.
(M) ▼		 Height Estimation Classification
δ1.25 (%) ▲ MSLE (10-2) ▼ MAE (m) ▼ RMSE (m) ▼ F1 (%) ▲ Acc (%) ▲
Allometric equations* 0 48.74 10.97 2.98 3.73 - -
Mask R-CNN† 88 74.34 ± 0.32 4.09 ± 0.06 1.73 ± 0.02 2.26 ± 0.02 65.49 ± 0.45 65.86 ± 0.60
ResNet50 47 68.16 ± 0.13 5.45 ± 0.09 1.96 ± 0.01 2.48 ± 0.01 76.67 ± 0.38 74.01 ± 0.50
ResNet50 w/ DSM 47 68.11 ± 0.16 5.12 ± 0.05 1.96 ± 0.01 2.47 ± 0.01 76.36 ± 0.36 73.58 ± 0.50
ConvNext-B 175 84.82 ± 0.39 2.47 ± 0.08 1.32 ± 0.02 1.71 ± 0.02 84.87 ± 0.42 82.91 ± 0.40
SwinV2-B 174 77.39 ± 0.87 3.38 ± 0.12 1.73 ± 0.04 2.17 ± 0.05 85.78 ± 0.45 83.77 ± 0.56
MambaVision-B 193 68.47 ± 0.70 60.19 ± 4.47 2.36 ± 0.06 3.42 ± 0.10 85.37 ± 0.28 83.42 ± 0.40
AnySat 250 78.03 ± 0.28 3.35 ± 0.05 1.57 ± 0.01 2.04 ± 0.01 69.99 ± 0.61 68.98 ± 0.44
PECore-B 187 84.85 ± 1.03 2.31 ± 0.14 1.29 ± 0.04 1.68 ± 0.05 86.82 ± 0.33 84.57 ± 0.53
DINOv3-B 171 86.35 ± 0.26 2.10 ± 0.04 1.25 ± 0.01 1.62 ± 0.01 87.46 ± 0.53 85.10 ± 0.80
DINOvTree-B (Ours) 100 88.29 ± 0.32 1.85 ± 0.05 1.17 ± 0.02 1.54 ± 0.02 86.89 ± 0.31 85.03 ± 0.45
DINOv3-L-Sat (❄) 606 53.07 ± 0.04 10.71 ± 0.02 2.81 ± 0.00 3.40 ± 0.00 27.43 ± 0.12 25.03 ± 0.11
DINOv3-L (❄) 606 58.28 ± 0.06 8.42 ± 0.01 2.48 ± 0.00 3.06 ± 0.00 28.76 ± 0.07 26.64 ± 0.06
DINOv3-L 606 88.03 ± 0.38 1.86 ± 0.04 1.17 ± 0.01 1.53 ± 0.01 88.85 ± 0.37 86.92 ± 0.57
DINOvTree-L (Ours) 328 89.28 ± 0.28 1.71 ± 0.04 1.12 ± 0.02 1.46 ± 0.02 88.38 ± 0.38 87.25 ± 0.36

BCI

Method Param.
(M) ▼		 Height Estimation Classification
δ1.25 (%) ▲ MSLE (10-2) ▼ MAE (m) ▼ RMSE (m) ▼ F1 (%) ▲ Acc (%) ▲
ConvNext-B 175 85.91 ± 0.96 2.53 ± 0.07 3.68 ± 0.09 4.75 ± 0.08 45.99 ± 0.55 48.16 ± 0.63
SwinV2-B 174 86.38 ± 0.69 2.53 ± 0.09 3.51 ± 0.07 4.64 ± 0.07 45.73 ± 1.21 46.94 ± 1.42
PECore-B 187 89.22 ± 0.49 2.18 ± 0.07 3.21 ± 0.04 4.24 ± 0.06 41.92 ± 1.94 42.31 ± 1.59
DINOv3-B 171 85.74 ± 0.69 2.61 ± 0.09 3.61 ± 0.07 4.69 ± 0.07 46.44 ± 1.32 46.44 ± 1.32
DINOvTree-B (Ours) 100 91.25 ± 0.40 1.95 ± 0.05 3.11 ± 0.08 4.14 ± 0.10 48.96 ± 0.85 50.22 ± 1.00
ConvNext-L 392 86.14 ± 0.74 2.49 ± 0.07 3.63 ± 0.10 4.71 ± 0.09 46.98 ± 0.70 48.79 ± 0.71
PECore-L 634 91.88 ± 0.49 1.93 ± 0.03 3.01 ± 0.02 3.99 ± 0.03 51.93 ± 1.17 52.97 ± 0.59
DINOv3-L 606 86.78 ± 1.39 2.32 ± 0.12 3.41 ± 0.10 4.44 ± 0.11 53.49 ± 0.61 53.87 ± 1.20
DINOvTree-L (Ours) 328 94.14 ± 0.23 1.68 ± 0.01 2.75 ± 0.01 3.75 ± 0.02 54.39 ± 0.67 55.58 ± 0.63

Quebec Plantations

Method Param.
(M) ▼		 Height Estimation Classification
δ1.25 (%) ▲ MSLE (10-2) ▼ MAE (m) ▼ RMSE (m) ▼ F1 (%) ▲ Acc (%) ▲
ConvNext-B 175 84.70 ± 0.31 1.63 ± 0.04 0.39 ± 0.00 0.61 ± 0.01 83.78 ± 1.47 85.95 ± 1.47
SwinV2-B 174 80.37 ± 0.58 2.16 ± 0.11 0.47 ± 0.01 0.71 ± 0.02 79.33 ± 2.34 80.21 ± 2.22
PECore-B 187 84.57 ± 0.31 1.56 ± 0.04 0.39 ± 0.01 0.58 ± 0.01 84.34 ± 1.36 84.67 ± 1.92
DINOv3-B 171 86.71 ± 0.39 1.48 ± 0.04 0.37 ± 0.00 0.58 ± 0.01 82.56 ± 1.86 84.00 ± 1.76
DINOvTree-B (Ours) 100 85.13 ± 0.40 1.54 ± 0.03 0.38 ± 0.00 0.57 ± 0.01 81.88 ± 1.20 83.97 ± 0.45
ConvNext-L 392 84.34 ± 0.32 1.65 ± 0.02 0.40 ± 0.00 0.61 ± 0.01 84.46 ± 0.45 85.25 ± 0.87
PECore-L 634 86.74 ± 0.74 1.44 ± 0.04 0.36 ± 0.01 0.58 ± 0.01 86.61 ± 0.70 86.90 ± 1.35
DINOv3-L 606 86.26 ± 0.59 1.39 ± 0.03 0.36 ± 0.00 0.57 ± 0.01 85.94 ± 0.37 85.80 ± 0.38
DINOvTree-L (Ours) 328 85.09 ± 0.26 1.47 ± 0.02 0.37 ± 0.00 0.58 ± 0.01 87.80 ± 1.19 87.63 ± 1.33

BibTeX

@article{endres2026treeheightspecies,
  title     = {Estimating Individual Tree Height and Species from UAV Imagery},
  author    = {Endres, Jannik and Lalibert{\'e}, Etienne and Rolnick, David and Ouaknine, Arthur},
  journal   = {arxiv:ToDo [cs.CV]},
  year      = {2026}
}