The processing of hundreds of thousands images from our camera traps mobilizes the expertise of researchers, machine learning and volunteer observers. Throughout the summer, extensive work was carried out to optimize the method. Next season of citizen science program “Wild Mont-Blanc” looks promising!
Is it really a chamois? The answer is far from obvious when one observes hundreds of images, taken from different angle, by the 40 camera traps set up in the Mont Blanc Massif. How, then, can we be sure to correctly identify species? Without it being too time-consuming.
For the past 3 years, CREA Mont-Blanc researchers have been experimenting Machine Learning to help them in their work. Their partner Devoteam Revolve is providing valuable expertise in this area. Last year, the solution tested made it possible to sort out images without animals. That’s 2/3 of the images captured. “Our learning software solution is open source. Implemented from the Megadetector solution, it allows researchers to have a simple review of the detection and learning phases.” explains Michaël Thibon, consultant in charge of the project at Devoteam Revolve.
Automated Machine Learning
Throughout the summer, significant work was carried out on the automation of Machine Learning (ML). Devoteam compared its progress with the CNRS DeepFaune model, already used by CREA Mont-Blanc. “DeepFaune is generalist and based on French fauna, so it is very complete. We are specific to the Alps and our scope is more limited. Technically, we don’t overlearn the model and we try to minimize the “noise” around the animals, to refine its reliability as much as possible,” explains Michaël.
The results are promising. “Our model gives very satisfactory answers for large herbivores. We need to improve further and make our machine learn about small species.”
Devoteam Revolve’s tool has another singularity. It allows renting the time of use strictly necessary for processing and analysis, from AWS (Amazon Web Service). It does not require CREA Mont-Blanc to purchase servers, a real advantage regarding the overall cost of the project.
AI and Citizen Science
At the same time, the citizen science component has also been improved. Matthieu Vercaemer spent 6 months of his internship as an environmental engineer (AgroParisTech) on the subject. “I studied the performance of citizen science in the identification of images via the Zooniverse portal. It was a question of objectifying the ratio of observers needed on the Wild Mont-Blanc program” he explains.
During the first two seasons, 50 identifications were needed to validate only one specie. Thanks to Matthieu’s statistical analysis, it is now possible to limit this group to 30 people, without any loss of quality.
Matthieu then compared his results with those from machine learning. “The identification of species by a human is clearly better, but the machine learning solution is already 90% reliable. It will make further progress and will be a real help for “discreet” species, such as the ermine.”
Wild Mont-Blanc season 3, coming soon
The 3rd season of Wild Mont-Blanc, online citizen science program, will begin early November, with thousands of volunteers gathered on the Zooniverse platform. “This year, we are switching to a production configuration, with the uploading of all sequences without prior observation by CREA Mont-Blanc. Thanks to the work of Matthieu and Devoteam Revolve, we are changing the scope of the project. All the sequences are screened by machine learning, given a score and then given to volunteer observers. It is up to them to determine which animal it is, among the twenty or so species monitored,” explains Colin Van Reeth, head of citizen science at CREA Mont-Blanc.
Researchers and naturalist experts will only assist in the process as a back-up, in the case of a high-stakes species (wolf, ptarmigan, etc.) or a lack of consensus among volunteer observers. This year, a new species will be identified (European wild cat) and two others have been gathered (mountain hare and the European hare).
Leave a Reply