Advancing conservation with AI-based facial recognition of turtles

Discovering options to enhance turtle reidentification and supporting machine studying initiatives throughout Africa

Defending the ecosystems round us is vital to safeguarding the way forward for our planet and all its residing residents. Luckily, new synthetic intelligence (AI) programs are making progress in conservation efforts worldwide, serving to sort out advanced issues at scale – from finding out the behaviour of animal communities within the Serengeti to assist preserve the diminishing ecosystem, to recognizing poachers and their wounded prey to stop species going extinct.

As a part of our mission to assist profit humanity with the applied sciences we develop, it is essential we guarantee various teams of individuals construct the AI programs of the longer term in order that it’s equitable and truthful. This consists of broadening the machine studying (ML) group and interesting with wider audiences on addressing essential issues utilizing AI. 

By means of investigation, we got here throughout Zindi – a devoted companion with complementary targets – who’re the most important group of African knowledge scientists and host competitions that concentrate on fixing Africa’s most urgent issues. 

Our Science workforce’s Range, Fairness, and Inclusion (DE&I) workforce labored with Zindi to determine a scientific problem that would assist advance conservation efforts and develop involvement in AI. Impressed by Zindi’s bounding field turtle problem, we landed on a undertaking with the potential for actual affect: turtle facial recognition. 

Biologists contemplate turtles to be an indicator species. These are courses of organisms whose behaviour helps scientists perceive the underlying welfare of their ecosystem. For instance, the presence of otters in rivers has been thought-about an indication of a clear, wholesome river, since a ban on chlorine pesticides within the Nineteen Seventies introduced the species again from the brink of extinction. 

Turtles are one other such species. By grazing on seagrass cowl, they domesticate the ecosystem, offering a habitat for quite a few fish and crustaceans. Historically, particular person turtles have been recognized and tracked by biologists with bodily tags, although frequent loss or erosion of those tags in seawater has made this an unreliable technique. To assist clear up a few of these challenges, we launched an ML problem known as Turtle Recall.

Instance of picture knowledge for 4 turtles taken from the tutorial colab pocket book. Variations in lighting, scale, background, pose, and similarities between turtles added to the complexity of the prediction problem. Credit score: Zindi.

Given the extra problem of conserving a turtle nonetheless sufficient to find their tag, the Turtle Recall problem aimed to bypass these issues with turtle facial recognition. That is attainable as a result of the sample of scales on a turtle’s face is exclusive to the person and stays the identical over their multi-decade lifespan. 

The problem aimed to extend the reliability and velocity of turtle reidentification, and doubtlessly supply a solution to change the usage of uncomfortable bodily tags altogether. To make this attainable, we wanted a dataset to work from. Luckily, after Zindi’s earlier turtle-based problem with Kenyan-based charity Native Ocean Conservation, the groups have been kindly in a position to share a dataset of labelled pictures of turtle faces.

Visualisation of which turtle head areas a neural community pays consideration to when making its predictions of which particular person is within the picture. Left: A turtle’s face from the dataset. Center/Proper: activations from DenseNet121 and EfficientNetB5 on the identical picture. Credit score: Zindi and Zindi dialogue board person ZFTurbo.

The competitors began in November 2021 and lasted 5 months. To encourage competitor participation, the workforce applied a colab pocket book, an in-browser programming atmosphere, which launched two widespread programming instruments: JAX and Haiku. 

Members have been tasked with downloading the problem knowledge and coaching fashions to foretell a turtle’s identification, as precisely as attainable, given {a photograph} taken from a selected angle. Having submitted their predictions on knowledge withheld from the mannequin, they have been in a position to go to a public leaderboard monitoring the progress of every participant. 

The group engagement was extremely constructive, and so was the technical innovation displayed by groups in the course of the problem. In the course of the course of the competitors, we acquired submissions from a various vary of AI fanatics from 13 totally different African nations – together with nations not historically effectively represented on the largest ML conferences, reminiscent of Ghana and Benin. 

Our turtle conservation companions have indicated that the participant’s degree of prediction accuracy might be instantly helpful for figuring out turtles within the discipline, that means that these fashions can have an actual and fast affect on wildlife conservation. 

As a part of Zindi’s continued efforts to assist climate-positive challenges, they’re additionally engaged on Swahili audio classification in Kenya to assist translation and emergency companies, and air high quality prediction in Uganda to enhance social welfare. 

We’re grateful to Zindi for his or her partnership, and all those that contributed their time to the Turtle Recall problem and the rising discipline of AI for conservation. And we stay up for seeing how individuals around the globe proceed to search out methods to use AI applied sciences in the direction of constructing a wholesome, sustainable future for the planet.

Learn extra about Turtle Recall on Zindi’s weblog and study Zindi at

Leave a Comment