SelvaBox : A high - resolution dataset for tropical tree crown detection
Hugo Baudchon, Arthur Ouaknine, Martin Weiss, Mélisande Teng, Thomas R. Walla, Antoine Caron-Guay, Christopher Pal, Etienne Laliberté · 2025
Read SelvaBox : A high - resolution dataset for tropical tree crown detection by Hugo Baudchon, Arthur Ouaknine, Martin Weiss, Mélisande Teng, Thomas R. Walla, Antoine Caron-Guay, Christopher Pal, Etienne Laliberté (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-00170.txt.
Introduction
Hugo Baudchon and co-authors release, in 2025, the largest open dataset yet for spotting individual tree crowns in tropical forests. Tropical crowns overlap and vary hugely, so training detectors needs many labeled examples, which barely existed. SelvaBox gathers drone images from three countries with more than eighty-three thousand hand-labeled crowns.
Benchmarks on the data show sharper images consistently help detection. Models trained only on SelvaBox matched or beat rivals on unseen tropical datasets, without extra tuning. Mixing SelvaBox with three other datasets across image resolutions produced a detector that ranked first or second on every test set. Data, code, and trained models are public.
Catalog record
| CREATOR | Hugo Baudchon, Arthur Ouaknine, Martin Weiss, Mélisande Teng, Thomas R. Walla, Antoine Caron-Guay, Christopher Pal, Etienne Laliberté |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 78 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-00170
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
SelvaBox : A high - resolution dataset for tropical tree crown detection Hugo Baudchon, Arthur Ouaknine, Martin Weiss, Mélisande Teng, Thomas R. Walla, Antoine Caron-Guay, Christopher Pal, Etienne Laliberté Abstract Detecting individual tree crowns in tropical forests is essential to study these complex and crucial ecosystems impacted by human interventions and climate change. However, tropical crowns vary widely in size, structure, and pattern and are largely overlapping and intertwined, requiring advanced remote sensing methods applied to high-resolution imagery. Despite growing interest in tropical tree crown detection, annotated datasets remain scarce, hindering robust model development. We introduce SelvaBox , the largest open - access dataset for tropical tree crown detection in high-resolution drone imagery. It spans three countries and contains more than 83 000 83\,000 manually labeled crowns – an order of magnitude larger than all previous tropical forest datasets combined. Extensive benchmarks on SelvaBox reveal two key findings: 1 higher-resolution inputs consistently boost detection accuracy; and 2 models trained exclusively on SelvaBox achieve competitive zero-shot detection performance on unseen tropical tree crown datasets, matching or exceeding competing methods. Furthermore, jointly training on SelvaBox and three other datasets at resolutions from 3 to 10 cm per pixel within a unified multi-resolution pipeline yields a detector ranking first or second across all evaluated datasets. Our dataset 1 1 1 SelvaBox dataset: https://huggingface.co/datasets/CanopyRS/SelvaBox , code 2 2 2 Preprocessing library ( geodataset ): https://github.com/hugobaudchon/geodataset 3 3 3 Benchmark, inference, and training ( CanopyRS ): https://github.com/hugobaudchon/CanopyRS , and pre-trained weights are made public. Figure 1 : The SelvaBox dataset. The illustrated samples are extracted from rasters recorded in Panama, Brazil and Ecuador with a spatial extent of 80\text{m}\times 80\text{m} and a resolution of 1.2 to 5.1 cm per pixel. The red square on the right highlights a zoom of the Ecuador sample with a spatial extent of 40\text{m}\times 40\text{m} at the same resolution. 1 Introduction Tropical forests cover 10% of the land area, but they store most of the biomass and biodiversity of plants on our planet pan_large_2011 ; gattiNumberTreeSpecies2022 . The largest trees that reach the upper canopy have a disproportionate influence on the functioning of tropical forests. For example, the largest 1% of trees store half of the carbon of forests worldwide lutzGlobalImportanceLargediameter2018a . However, tree demography patterns in tropical forests are being altered, with increasing tree mortality, due to climate change brienenLongtermDeclineAmazon2015 ; bonan_forests_2008 ; esquivel-muelbertCompositionalResponseAmazon2019 and human interventions harris_global_2021 . As such, monitoring of individual trees in tropical forests is essential to understand the current and future ability of these forests to regulate the global climate daviesForestGEOUnderstandingForest2021 . Monitoring tropical trees is a difficult task involving slow, costly, and dangerous ground surveys by forest technicians delimaMakingForestData2022 . Forest plots of tens of hectares are the gold standard of tropical tree monitoring to measure and map each individual, but completing a single one can take years of dedicated work by large teams of experts daviesForestGEOUnderstandingForest2021 . Remote sensing technologies considerably augment field work, facilitating forest cartography through aerial detection of individual trees across spatial extents vastly exceeding the practical limitations of ground-based inventories brandtUnexpectedlyLargeCount2020 . Satellite imagery has been used for forest monitoring ouaknine_openforest_2025 , such as for height map estimation at 1 m resolution tolan_very_2024 ; lang_high-resolution_2023 , or individual tree crown detection brandtUnexpectedlyLargeCount2020 ; tucker_sub-continental-scale_2023 using imagery at a resolution of 0.3 to 0.5 m. Such satellite imagery is considered very high resolution but is still too coarse to distinguish trees in dense tropical forest canopies. Furthermore, cloudy conditions complicate satellite remote sensing in the tropics. By contrast, unoccupied aerial vehicles (UAVs) or drones can be flown under clouds at tens of meters above the forest and therefore achieve cm-resolution ( <5 cm), albeit at the expense of spatial coverage reiersen_reforestree_2022 ; vasquez_barro_2023 ; cloutier_influence_2024 . While UAV LiDAR has been exhaustively explored for forest structure assessment, both with datasets puliti_for-instance_2023 ; puliti_benchmarking_2025 ; gaydon_pureforest_2025 and methods xi_see_2020 ; wilkes_tls2trees_2023 ; you_segmentation_2023 ; bai_semantic_2023 ; liu_study_2023 ; ma_forest-pointnet_2023 ; vermeer_lidar-based_2024 ; henrich_treelearn_2024 ; wielgosz_segmentanytree_2024 , its high cost and limited accessibility in tropical regions justify the development of RGB-only detection methods. However, the vast majority of open access high-quality, high-resolution tree detection RGB datasets represent temperate forests of the global North (Tab. 1 ). Tropical forests, particularly in the Global South, remain severely underrepresented and include relatively modest annotation counts ball_accurate_2023 ; vasquez_barro_2023 despite the critical significance of tropical forests for biodiversity and carbon storage. Tropical forests have a large tree species diversity gattiNumberTreeSpecies2022 and heterogeneity in tree crown sizes (Fig. 2 ), from massive emergent trees to small understory species, as well as in shapes and textures (Fig. 1 ). This highlights an open topic of research on computer vision applied to remote sensing rabbi_small-object_2020 ; li_cross-layer_2021 ; bashir_small_2021 where both large and small objects must be detected within the same scene. Tropical forest monitoring needs innovation in application-driven machine learning solutions rolnick_position_2024 to address challenges of detecting numerous objects with highly variable sizes. While convolutional neural networks (CNNs) remain the predominant approach for individual tree crown detection weinstein_individual_2019 ; zamboni_benchmarking_2021 ; yu_comparison_2022 ; ball_accurate_2023 ; zhao_systematic_2023 ; bountos_fomo_2025 , recent studies have begun exploring transformer-based detection architectures on satellite imagery JIANG2025103114 , motivated by their demonstrated effectiveness in multi-scale object recognition tasks (e.g., liu_swin_2021 ; liu_swin_2022 ; zhang_dino_2023 ). However, a comprehensive, resolution-aware benchmark systematically comparing these two model paradigms on UAV imagery across diverse forest ecosystems and out-of-distribution scenarios remains absent. With the growing number of UAV datasets acquired with different flight parameters, there is a need for models that can generalize across resolutions and for standardized frameworks to bridge the persistent gap between ecology and computer vision communities. We address these challenges through our contributions: 1 SelvaBox , a high-resolution drone imagery dataset spanning three neotropical countries (Brazil, Ecuador, Panama) and comprising over 83\,000 manual bounding box annotations on individual tree crowns; 2 An exhaustive benchmark of detection methods at varying resolutions and input sizes, including a standardized evaluation framework for UAV rasters and a comprehensive assessment of models’ generalization capacity on out-of-distribution (OOD) samples; 3 State-of-the-art models trained for tree crown detection out-performing competing methods on both topical and non-tropical forest datasets, in both in-distribution (ID) and OOD settings; and 4 two open-source Python libraries facilitating raster preprocessing, inference, postprocessing and standardized benchmarking. Through these contributions, we aim to simultaneously advance tropical forest monitoring and applications of machine learning to critical environmental challenges. 2 Related work Datasets. High-resolution drone imagery enables detailed tree characterization at the pixel level (see Figure 1 ). This capability has catalyzed the development of open access forest monitoring datasets ouaknine_openforest_2025 specifically designed for tree crown semantic segmentation tasks, including pixel-wise canopy mapping galuszynski_automated_2022 , woody invasive species identification kattenborn_uav_2019 , and tree species classification cloutier_influence_2024 ; kattenborn_convolutional_2020 . Name # Trees GSD Type Biome NeonTreeEval. weinstein_benchmark_2021 16k 10 natural temperate ReforesTree reiersen_reforestree_2022 4.6k 2 plantation tropical Firoze et al. firoze_tree_2023 6.5k 2-5 natural temperate Detectree2 ball_accurate_2023 3.8k 10 natural tropical BCI50ha vasquez_barro_2023 4.7k 4.5 natural tropical BAMFORESTS troles_bamforests_2024 27k 1.6-1.8 natural temperate QuebecTrees cloutier_influence_2024 23k 1.9 natural temperate Quebec Plantation lefebvre_uav_2024 19.6k 0.5 plantation temperate OAM-TCD veitch-michaelis_oam-tcd_2024 280k 10 mostly urban worldwide SelvaBox (ours) 83k 1.2-5.1 natural tropical Table