State-run Agricultural Research Corporation EMBRAPA will launch algorithms that help create sustainable management projects

04/19/2024

---

Data collection in the field: with EMBRAPA’s help, the cost of a typical survey has dropped from R$140 to as low as R$4 per hectare — Foto: Felipe Sá/Divulgação

An artificial intelligence tool is aiding in the identification of commercially valuable trees and pinpointing their exact location in the forest. Species such as Brazil nut, cumaru-ferro, açaí, and cedar are recognized with an accuracy rate of 95%, reducing production costs and promoting more sustainable forest management practices in the Amazon.

Netflora, a methodology developed by the state-run Agricultural Research Corporation (EMBRAPA), encompasses a set of algorithms trained with artificial intelligence to recognize forest species based on botanical characteristics available in a database. The methodology is targeted towards companies in the forestry sector, academic professionals, agro-extractivist associations, and environmental agencies.

Evandro Orfanó, an EMBRAPA researcher who co-heads these studies, Netflora automates forest activity planning and enhances the precision and efficiency of management plans. “Once trained and specialized, the algorithm also provides metrics such as diameter and canopy area, which enable the estimation of each tree’s wood volume through allometric equations (which relate to shapes and sizes),” he said.

EMBRAPA’s research to enable the use of AI in the forestry sector began in 2015 and covers several aspects of the activity. In the current phase, studies are conducted under the Geoflora project, implemented in the states of Acre, Rondônia, Roraima, Amapá, Pará, and Amazonas, in collaboration with the JBS Fund for the Amazon.

The adoption of these technologies requires investments in computers, drones, batteries, and suitable office infrastructure. According to Mr. Orfanó, the drastic reduction in production costs, especially in the forest inventory stage, offsets these expenses.

The cost of a traditional species survey with field teams ranges between R$100 and R$140 per hectare of mapped forest. In the Netflora methodology, the expense is between R$4 and R$6 per hectare.

This significant difference is due to the agility in obtaining and processing information. “A forestry company using traditional management can map up to 10,000 hectares of forest per year. With the use of AI, operational capacity can increase to up to 1 million hectares in the same period,” Mr. Orfanó added.

To build the algorithm’s “training” database, drones mapped over 40,000 hectares of forest in 37 areas of Acre, Rondônia, and the southern Amazon. Over two years of study, researchers conducted around a thousand flight plans, each generating about 300 aerial images. These images underwent processing and were transformed into orthophotos (georeferenced and high-resolution images). The information from orthophotos was used to train nine algorithms, each with different purposes and performance levels.

“We have algorithms that recognize a single forest species, while others can identify different groups or the main timber and non-timber trees of Acre and other locations in the Amazon. Some algorithms have already achieved high performance, but this learning will be continuous,” said Mr. Orfanó. EMBRAPA aims to map 80,000 hectares and include new areas of commercial interest in the Amazon to expand the database.

The state-run company will launch the first two algorithms that have undergone the refinement phase next Wednesday. One of them can recognize “açaí solteiro” palms in productive (with clusters) and non-productive phases in Acre. The other, in addition to single-stem açaí, can recognize nine other palm species from the Amazon (paxiúba, buriti, jaci, ouricuri, murmuru, tucumã, inajá, patauá, and bacaba).

The launch of the other seven algorithms is expected to take place by February 2025. The list includes algorithms that identify timber and non-timber species in different Amazonian locations, as well as versions intended for environmental monitoring and species recognition in agroforestry systems.

Forest Engineer Mauro Alessandro Karasinski, a doctoral student at the Federal University of Paraná (UFPR) and a member of the Netflora creation team, said that the algorithm learns tree canopy patterns and organizes this information to recognize features in images of newly mapped areas. “As a result of learning, a ‘shapefile’ [file with identification and location of each species and indication of certainty level] allows for the preparation of the forest inventory, providing information on the number of existing trees, divided by class or genus, and other data on the species and the mapped area,” he said.

*Por Marcelo Beledeli — Porto Alegre

Source: Valor International

https://valorinternational.globo.com/