Research using AI to track Amazon rainforest species produces landmark results

Thanks to the tenacity of a George Mason University biology professor, animals in the Amazon have a lot less privacy these days, but that’s good for scientists who to want know what’s going on in the Brazilian rainforests.

David Luther recognizes the importance of playing the long game in research, but even he couldn’t have imagined the hurdles and roadblocks ahead of him following the approval of his 2019 4-VA Collaborative Research Grant proposal, “Species richness resilience to habitat fragmentation and restoration in tropical rainforests.”

His plan was to install recording equipment at 50 sites as part of the Biological Dynamics of Forest Fragments Project (BDFFP) in the Amazon rainforest of Brazil. These audio and video devices were meant to document and measure differences in animal community composition and the rate of animal recovery in secondary forest and forest fragments.

The proposed budget was devoted entirely to purchasing the wide array of materials necessary for the effort — cameras, acoustic recorders, and batteries—along with the international travel needed to bring the project to fruition.

In March 2020, all of Luther’s efforts came to a halt due to the pandemic. Additionally, one of the key members of the planning team, Tom Lovejoy, passed away in December 2021. Lovejoy was recognized as one the world’s leading conservation biologists and often referred to as the “godfather of biodiversity.” With Lovejoy’s passing, Luther lost a mentor and a critical member of the team.

However, Luther stayed the course, revamping his team and rewriting the schedule. Finally, in June 2022, he received the green light to move ahead. Between June and October 2022, 136 cameras and 81 acoustic devices were installed across 50 sites at BDFFP.

Today, to Luther’s great delight, the results have proved far more successful than he could have ever anticipated. Tens of thousands of animal images from camera traps and audio recordings have already been collected.

To analyze the data, Luther built a team of 15 George Mason undergraduate researchers, artificial intelligence (AI) experts, and members of Arbimon, a nonprofit organization that specializes in analyzing acoustic recordings from the tropics to identify animals.

The student volunteers on this project included Alexis Lembke, Amanda Jones, Adriana Em, Madison Cheung, Morgan Ellingsworth, and Grace Carriero. Aline Medeiros, a PhD student in environmental science and policy, helped manage the undergraduate researchers working on the audio files. Medeiros will also use the captured data as the basis of her PhD research.

Another set of students helped identify animals in the camera images and entered that information into a large database. Hibo Hassan, Jordan Seidmeyer, Katie Russell, Carolian Sanabria, Adrian Em, Alix Upchurch, Piper Robinson, Tristan Silva-Montoya, and Estefany Umana spent hours creating this treasure trove of records. Emilia Roberts, a master’s student in environmental science and policy, managed these undergraduate researchers.

Discussing the project, Luther explained that, for the acoustic recordings, they built templates for 250 bird species and trained AI models to automatically detect and classify songs for each.

“The model performed very well in our evaluations. We have already detected 201 of the 250 species,” said Luther, who teaches in the Biology Department of George Mason’s College of Science. “Thanks to our model, new recordings can be passed through it to automatically detect species calls, facilitating long-term monitoring and efficient analyses moving forward. We are now working with local experts in Manaus, Brazil, to apply the same platform for frogs at our study sites in the Amazon rainforest.”

The biodiversity data is being used to assess how each species responds to variations in forest structure and recovery from forest fragmentation. Luther brought on Konrad Wessels from George Mason’s Geography and Geoinformation Science Department to assist with satellite information from the Global Ecosystem Dynamics Investigation instrument, which uses high-resolution lasers to provide detail in three-dimensional forest structure. These results will build predictive models to study how the forest structure can forecast mammal and bird diversity and individual species occurrence in tropical rainforests.

In an important finding, the team has determined that the three-dimensional complexity of the forest structure, as measured by foliage height diversity, is the biggest predictor of mammal and bird diversity at this research site, and potentially others, in the Amazon rainforest.

The project continues to gain traction. The team has created a website featuring the results of the acoustic portion of the research, which has been very well received. Some of the acoustic training models have been used by teams competing for the X-Prize, a competition designed to encourage technological developments supporting “radical breakthroughs for the benefit of humanity.”

Building on the 4-VA funded study, Luther also received a $200,000 National Science Foundation grant to continue both the camera and acoustic research. Luther and Wessels recently submitted a grant to NASA to expand on the research findings and apply them to the entirety of the Amazon basin.

“Through 4-VA@Mason, this project is up, running, and delivering fantastic information that will help scientists worldwide better design monitoring schemes for biodiversity in remote tropical forests, as well as those interested the relationship between habitat structure and degradation and species resilience to disturbance,” said Luther.