Birdwatching is better with AI and CNN

We may have to add advanced birdwatching to the list of things computers can do that we can’t.

Researchers say they have demonstrated for the first time that a method of deep learning known as convolutional neural networks (CNN) can be used to train computers to consistently recognise dozens of individual birds that a person couldn’t tell apart.

And that “provides the means of overcoming one of the greatest limitations in the study of wild birds”, says André Ferreira from France’s Centre for Functional and Evolutionary Ecology (CEFE), lead author of a paper in the journal Methods in Ecology and Evolution.

Ferreira and colleagues from France, Germany, Portugal and South Africa trained their AI models to recognise thousands of images of great tits (Parus major) and sociable weavers (Philetairus socius) taken in the wild and a captive population of zebra finches (Taeniopygia guttata) – commonly studied birds in behavioural ecology.

After training, the models were tested with images they had not seen before and achieved accuracy of more than 90% for the wild species and 87% for the captive finches.

It’s certainly a more sophisticated, not to mention less stressful, approach than traditional banding.

Most birds in the study populations carried a passive integrated transponder (PIT) tag, similar to the microchips implanted in pet cats and dogs. Antennae on high-tech bird feeders were able to read the identity of the bird from the tags and trigger cameras.

It’s early days and there are limitations. The AI model is only able to re-identify individuals it has been shown before, and it is not clear how it will perform if the appearance of individual birds changes over time, for example when they moult. Images of the same bird taken months apart could be mistakenly identified as different individuals.

The authors believe, however, that both these limitations can be overcome with datasets containing thousands of images of thousands of individuals over long periods of time, which they are currently trying to collect.

And they are confident that their success to date underlines the potential to apply CNN to a range of research projects and a range of wildlife.

“[W]e found that our trained CNNs were generalisable, meaning that the rate of successful re-identification remained high across different recording contexts,” they write in their paper.

“This is particularly relevant as researchers are often interested in collecting data in contexts that are challenging, from parental behaviour at the nest to dominance interactions away from artificial feeders.”