IMPA and Imazon develop AI that identifies deforestation.

The IMPA Center for Projects and Innovation , in partnership with Imazon (Institute of Man and Environment of the Amazon), developed an algorithm to assist in the detection of deforestation in the Amazon Rainforest. The work was presented last Thursday (7) by IMPA doctoral student Rodrigo Schuller at the XXI SELPER International Symposium: “Beyond the Canopy – Technologies and Applications of Remote Sensing” , in Belém. Imazon analyzed images with and without deforestation of the region, obtained from the Copernicus Project (ESA) through Google, to train the models developed by IMPA, using neural network technology. The work “ Artificial intelligence to support deforestation monitoring in the Brazilian Amazon” has already resulted in the first version of the AI tool that is in use at Imazon. Schuller explained that the algorithm has a 95% accuracy rate and represents a 30% time saving for specialists. A task that previously took 1004 hours is now completed in 705 hours with the use of the tool. “The performance of the deforestation detection model is not sufficient to fully automate the process, but we can already use the model as an assistant,” he said. Read more: Pi Center uses neural networks to locate oil at sea Novello is the new visiting researcher at Google DeepMind Applications open for the 2025 IMPA Tech selection process Larissa Amorim, a researcher at Imazon, highlighted the importance of AI in the organization's daily work. "The creation of this AI algorithm helps in decision-making regarding the classification of deforestation and non-deforestation alerts, thus accelerating and bringing greater accuracy to the monitoring process. The time saved in validating alerts can be concentrated on other activities, such as focusing on partnerships for the effective use of data to combat and control deforestation." Before AI, Imazon used a simple algorithm capable of excluding areas with no evidence of deforestation. Subsequently, a team from the institute dedicated themselves to manually analyzing the remaining images – which demanded a lot of observation time. Despite the improvement, the Pi Center is studying how to reduce noise and increase the accuracy of the algorithm to make it more effective. “There is a field of machine learning called robust learning, which deals with ambiguous or noisy labels. Since some natural phenomena – such as rock outcrops – are similar to deforestation when viewed in satellite images, learning processes adjusted to handle these uncertainties can achieve higher performance,” explained Schuller. “Another approach we are investigating is using interpretable models, that is, models that offer explanations for why an area was considered deforested or not. In this way, we can quantify the importance of factors such as seasonal periodicity, proximity to rivers, or before-and-after images for the model's prediction in each region,” concluded the IMPA doctoral student. In addition to Schuller, the project team includes project scientist Francisco Ganacim and IMPA researcher Paulo Orenstein, who was awarded the Google Research Award in July of this year for his work on the project. Also read: 'We need to transform our expertise into wealth,' says Viana. IMPA has openings for researchers and postdoctoral fellows.