great tit identified by feeder system

AI model trained to distinguish between individual birds

Image credit: André Ferreira

An international team of researchers has trained an AI model to recognise individual birds: a skill which eludes humans.

Distinguishing between individual animals is important for long-term monitoring of populations and protecting species from pressures such as climate change. However, it is also one of the most expensive, troublesome, and time-consuming aspects of animal behaviour research.

While some creatures such as leopards have unique markings which allow humans to recognise individuals by eye, most species require additional visual identifiers such as coloured bands to be distinguished. Attaching bands to birds’ legs can be stressful and disruptive to the animals, limiting the scope of research.

Seeking an alternative method for distinguishing between individual birds, researchers from institutes in France, Germany, Portugal, and South Africa developed the first AI bird identification tool of its kind.

They began by collecting thousands of labelled images of great tits, sociable weavers, and zebra finches (some of the most commonly studied birds in behavioural ecology).

Rather than manually labelling the thousands of birds in the dataset – which would have been an unfeasibly time-consuming task – the researchers built bird feeders with camera traps and sensors. Most of the birds in the study populations already carried a passive integrated transponder tag (similar to a pet microchip) which allowed antennas on the feeders to “read” the identity of the bird and use this information to tag the corresponding photograph.

The data was used to train a convolutional neural network: a type of neural network suitable for tasks like image classification. While this family of methods has previously been used to identify species of primates, pigs, and elephants, it has never been used to identify smaller animals like birds.

The AI models were tested with images of individual birds that were part of the training set. This method was able to identify birds with an accuracy of over 90 per cent for great tits and sociable weavers (wild populations) and 87 per cent for the zebra finches (captive population). According to lead author Dr André Ferreira, the tool is able to recognise dozens of individual birds from an image, even though humans cannot do this reliably.

“The development of methods for automatic, non-invasive identification of animals completely unmarked and unmanipulated by researchers represents a major breakthrough in this research field,” said Ferreira. “Ultimately, there is plenty of room to find new applications for this system and answer questions that seemed unreachable in the past."

Ferreira cautioned that at this stage the AI model is only capable of re-identifying individuals it has been shown before and would not be able to identify new birds. It is not known how changes in the appearance of an individual bird – such as moulting – would affect the performance of the model. The researchers are trying to collect much larger datasets containing images of individuals over long periods of time, which could overcome these limitations.

Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.

Recent articles