SELVABOX, the largest open-access dataset for tropical tree crown detection, spans three countries and includes over 83,000 manually labeled crowns. This extensive dataset enhances ecological research and machine learning applications, offering crucial insights into biodiversity and climate regulation in tropical forests.
Tackling Tropical Forest Monitoring
Tropical forests, covering just 10% of the Earth’s land area, are vital ecosystems that store the majority of the planet’s biomass and biodiversity. The largest trees within these forests play a crucial role in carbon storage, influencing global climate regulation. However, these ecosystems face significant threats from climate change and human interventions, leading to altered tree demographics and increased mortality rates. Monitoring these changes is essential for understanding the potential of tropical forests to mitigate climate change.
Traditionally, monitoring tropical trees involves labor-intensive ground surveys, which are slow, costly, and sometimes dangerous. These surveys, often covering tens of hectares, can take years to complete. Remote sensing technologies offer a promising alternative, enabling large-scale forest cartography and individual tree detection. However, existing satellite imagery lacks the resolution needed to accurately distinguish individual trees in dense tropical canopies, and cloudy conditions further complicate satellite sensing in these regions.
Drones, or unoccupied aerial vehicles (UAVs), provide a solution with their ability to capture high-resolution imagery at the centimeter level. Despite this advantage, the high cost of LiDAR technology used in drones limits its widespread adoption in financially constrained tropical regions. Therefore, there is a pressing need for high-quality, high-resolution RGB datasets specifically tailored for tropical forests. Until now, such datasets have been scarce, hindering the development of robust models for tree crown detection in these critical ecosystems.
SELVABOX: A Breakthrough Dataset
Addressing these challenges, the SELVABOX dataset represents a significant advancement in tropical tree crown detection. Spanning three neotropical countries—Brazil, Ecuador, and Panama—SELVABOX comprises over 83,000 manually labeled tree crowns. This dataset is an order of magnitude larger than all previous tropical forest datasets combined, providing an unprecedented resource for researchers and practitioners.
The research team conducted extensive benchmarks on SELVABOX, yielding two pivotal findings: first, higher-resolution inputs consistently enhance detection accuracy; second, models trained exclusively on SELVABOX demonstrate competitive zero-shot detection performance on unseen datasets, often matching or surpassing existing methods. The dataset’s high-resolution imagery, ranging from 1.2 to 5.1 cm per pixel, enables detailed tree characterization at the pixel level, crucial for accurate detection and delineation of individual tree crowns.
Furthermore, the researchers developed a unified multi-resolution pipeline, training models on SELVABOX alongside three other datasets with resolutions from 3 to 10 cm per pixel. This approach resulted in a detector that ranks first or second across all evaluated datasets, showcasing the dataset’s versatility and the effectiveness of the developed models.
Setting New Standards
The SELVABOX dataset has set a new benchmark for tropical tree crown detection, offering a comprehensive resource that significantly outperforms existing datasets in size and resolution. The models trained on SELVABOX not only excel in in-distribution settings but also exhibit strong generalization capabilities on out-of-distribution samples, a critical feature for real-world applications.
The research highlights the importance of high-resolution inputs in improving detection accuracy, a finding that could influence future dataset development and model training strategies. By making the dataset, code, and pre-trained weights publicly available, the researchers have provided a valuable tool for advancing both tropical forest monitoring and the application of machine learning to environmental challenges.
Future Directions and Impact
SELVABOX has the potential to transform how researchers monitor and understand tropical forests, offering insights into biodiversity, carbon storage, and the impacts of climate change. The dataset’s scale and quality enable the development of more accurate and generalizable models, bridging the gap between ecology and computer vision.
As the field of remote sensing continues to evolve, SELVABOX’s contributions will likely inspire further research into multi-resolution and multi-scale analysis, fostering collaboration between ecologists and computer scientists. Researchers and practitioners are encouraged to explore SELVABOX and contribute to its ongoing development, ensuring that this vital resource continues to support the preservation and understanding of tropical forests.
Reference: Baudchon, H., Ouaknine, A., Weiss, M., Teng, M., Walla, T. R., Caron-Guay, A., Pal, C., & Laliberté, E. (2026). SELVABOX: A High-Resolution Dataset for Tropical Tree Crown Detection. DOI: https://doi.org/10.48550/arXiv.2507.00170