Finding solutions to improve turtle re-identification and support machine learning projects across Africa
Protecting the ecosystems around us is essential to safeguarding the future of our planet and all its living citizens. Fortunately, new artificial intelligence (AI) systems are advancing conservation efforts around the world, helping to solve complex problems on a large scale – from study the behavior of animal communities in the Serengeti to help conserve a declining ecosystem, spot poachers and their injured prey to prevent the extinction of species.
As part of our mission to help humanity through the technologies we develop, it is important that we ensure that diverse groups of people build the AI systems of the future so that they are fair and just. This involves expanding the machine learning (ML) community and engaging with a wider audience to solve important problems using AI.
Through our investigation, we discovered Zindi – a dedicated partner with complementary goals – building the largest community of African data scientists and organizing competitions focused on solving Africa's most pressing problems.
OUR Scientific teamThe Diversity, Equity, and Inclusion (DE&I) team worked with Zindi to identify a scientific challenge that could help advance conservation efforts and increase participation in AI. Inspired by Zindi bounding box turtle challengewe landed on a project with real impact potential: facial recognition of turtles.
Biologists consider turtles an indicator species. These are classes of organisms whose behavior helps scientists understand the underlying well-being of their ecosystem. For example, the presence of otters in rivers is considered a sign of a clean and healthy river, since the ban on chlorinated pesticides in the 1970s brought the species to the brink of extinction. .
Turtles are another such species. By grazing on seagrass, they cultivate the ecosystem, providing habitat for many fish and crustaceans. Traditionally, individual turtles were identified and tracked by biologists using physical tags, although the frequent loss or erosion of these tags in seawater makes this method unreliable. To help solve some of these challenges, we launched an ML challenge called Turtle recall.
Given the added challenge of keeping a turtle still enough to locate its tag, the Turtle Recall challenge aimed to circumvent these issues through turtle facial recognition. This is possible because the pattern of scales on a turtle's face is unique to each individual and remains the same throughout its life, which spans several decades.
The challenge aimed to increase the reliability and speed of turtle re-identification, and potentially offer a way to completely replace the use of uncomfortable physical tags. To make this possible, we needed a dataset to work with. Luckily, after Zindi's previous turtle challenge with a Kenya-based charity Local ocean conservationthe teams were kindly able to share a dataset of labeled images of turtle faces.
The competition began in November 2021 and lasted five months. To encourage competitor participation, the team set up a Colab notebookan in-browser programming environment, which introduced two common programming tools: JAX And Haiku.
Participants were tasked with uploading challenge data and training models to predict the identity of a turtle, as accurately as possible, from a photograph taken from a specific angle. After submitting their predictions on the data retained from the model, they were able to consult a public ranking tracking the progress of each participant.
The community engagement was incredibly positive, as was the technical innovation demonstrated by the teams during the challenge. During the competition, we received applications from a wide range of AI enthusiasts from 13 different African countries, including countries that are not traditionally well represented at the largest ML conferences, such as Ghana and The Benin.
Our turtle conservation partners have indicated that the level of accuracy of participants' predictions will be immediately useful for identifying turtles in the field, meaning these models can have a real and immediate impact on wildlife conservation.
As part of Zindi's ongoing efforts to support climate positive challenges, they are also working on Swahili audio classification in Kenya to assist translation and emergency services, and air quality prediction in Uganda to improve social protection.
We are grateful to Zindi for their partnership, as well as to everyone who donated their time to the Turtle Recall challenge and the growing field of AI for conservation. And we look forward to seeing how people around the world continue to find ways to apply AI technologies to build a healthy and sustainable future for the planet.
Learn more about turtle recall at Zindi's blog and discover Zindi on https://zindi.africa/