ARBOR: the drone-trained AI to monitor shihuahuaco trees in the Peruvian Amazon

By Inforegion · June 24, 2026.

Peru's Agency for the Supervision of Forest Resources and Wildlife (OSINFOR) and the United Kingdom's University of Sheffield have officially unveiled ARBOR, an artificial intelligence plugin designed for forest monitoring and the assessment of specific tree species in the Peruvian Amazon. The tool was formally handed over to the Peruvian State during the event 'Dialoguemos por los bosques: tecnología para la innovación y la sostenibilidad' (Let's Talk About Forests: Technology for Innovation and Sustainability).

ARBOR is not a conventional logging detection system. What sets it apart from traditional approaches lies in broadening the scope of monitoring: instead of merely identifying forest loss, the tool can identify, segment and georeference individual species from images captured by drones in supervised management areas. This qualitative leap —from detecting absences (felled trees) to cataloging presences (living trees, classified by species)— represents a paradigm shift in Amazonian forest oversight.

In its first phase, ARBOR was trained specifically to recognize the shihuahuaco (Dipteryx micrantha), an emblematic species of the Peruvian Amazon both for its ecological value and for its commercial value on international timber markets. The shihuahuaco is included in Appendix II of the Convention on International Trade in Endangered Species of Wild Fauna and Flora (CITES), meaning that its international trade is regulated and requires documentation of legal origin.

To train the algorithms, the team used a geospatial database belonging to OSINFOR itself, made up of 176 orthomosaics —high-resolution aerial images generated through drone photogrammetry— obtained in supervised forest management areas. That database includes records of 1,883 trees, of which more than 700 are shihuahuacos. This is a relevant feature: the model was trained with data collected directly in Peruvian Amazon forests, which reduces the risk of mismatch between the training environment and the real application environment —one of the most frequent problems in computer vision models applied to tropical ecosystems.

Williams Arellano Olano, head of OSINFOR, stressed at the launch the value of the collaboration between the regulatory body and academia: 'This joint work between OSINFOR and academia has allowed us to develop concrete solutions to improve the sustainable management of our forests and strengthen forest monitoring through the use of artificial intelligence.' For his part, Jefferson dos Santos, a University of Sheffield researcher who took part in the development, highlighted the tool's export potential: 'The information generated by OSINFOR has been key to developing a technology with the potential to make an impact not only in Peru's forests, but in other countries.'

From the Peruvian regulatory standpoint, the application of ARBOR has a direct legal justification. Article 46 of the Forestry and Wildlife Law (Law No. 29763) establishes that oversight of shihuahuaco harvesting must be carried out at 100%. Fulfilling that mandate with traditional human and logistical resources —inspectors physically traveling to remote areas of the Amazon— is, in practice, extremely costly and slow. ARBOR offers a way to scale that legal obligation: by automating the identification and georeferencing of trees through drone images, OSINFOR can cover larger areas, more frequently and at lower operating cost.

The tool is part of the digital innovation ecosystem that OSINFOR has been building since 2015, when it began incorporating digital platforms, drones, satellite imagery and artificial intelligence into its oversight processes. ARBOR was formally born within OSINFORLAB, the agency's innovation and digital transformation laboratory, where other systems have also been developed, such as the Edge AI-based forest fire detector and the Certificate of Legal Origin.

The same event saw the presentation of ADETOP v2 Web, an updated and optimized version of the artificial intelligence algorithm that OSINFOR released in July 2025 to determine whether logging was legal or illegal. According to the figures provided at the event, ADETOP v2 has already been used in the analysis of 181 cases. This indicates that OSINFOR is not building laboratory tools disconnected from real operations, but rather is applying AI models to concrete oversight case files, with direct legal implications.

The event was attended by senior officials from Peru's environmental sector: Erasmo Otárola Acevedo (head of SERFOR), Deyvis Huamán Mendoza (director of Territorial Management of Natural Protected Areas at Sernanp), Carmen Vegas Guerrero (deputy minister for Strategic Development of Natural Resources at the Ministry of the Environment) and Vilma Vilcas Melchor (regional manager for Forestry and Wildlife of Ucayali). The presence of representatives from different levels of government —central and regional— suggests that ARBOR is being positioned not only as an OSINFOR tool, but as a resource potentially shared among several institutions with forestry responsibilities.

As sector context, the pressure on Peru's Amazon forests is significant. According to data cited by the outlet itself, the Amazon lost more than 736,000 hectares to deforestation in 2025, with Peru among the most affected countries, especially due to gold mining. In that context, tools that make it possible to verify the legal origin of timber and to precisely track the status of CITES-regulated species take on strategic relevance for both conservation and international trade: timber buyers in European and North American markets are increasingly demanding traceability and legality guarantees, backed by legislation such as the European Union Deforestation Regulation (EUDR).

From the perspective of agentic AI and computer vision systems applied to the environment, ARBOR illustrates a growing trend: the use of models trained with proprietary geospatial data —orthomosaics, multispectral images, field records— for tasks of species identification and classification in high-biodiversity ecosystems. The ability to georeference each detected tree turns the model's output not only into a classification label, but into a spatial datum that can be integrated into geographic information systems (GIS), forest inventory databases and chain-of-custody traceability platforms.

The choice of the shihuahuaco as the pilot species is no accident. As a high-commercial-value timber, protected by CITES and subject to mandatory 100% oversight, it concentrates the incentives for both illegal exploitation and rigorous supervision. If ARBOR proves its effectiveness on this species, extending the model to other CITES or high-value species presupposes additional labeling and retraining work, but on an architecture already validated under Amazon field conditions.