A.I. birder does what a human never could
Scientists are poised to better understand the inner lives of birds.
An immense frustration ecologists encounter is prompted by the attempt to keep track of individual animals in a study. This task only becomes more difficult when trying to pinpoint small, mobile animals like songbirds.
While intelligent computer algorithms can help scientists better complete this task, training these systems to recognize different species — let alone individuals in a species — can take thousands of data points, time, and money.
However, French and Portuguese researchers recently devised a way to streamline this process. They designed a deep-learning network that can identify individual birds with up to 92 percent accuracy in three different species.
This tech can not only save scientists resources but can help them collect important data about the lives of birds — and better understand what may be leading to their decline in North America.
The study team describes their approach to solving this historic problem using strategically timed photos, RFID sensors, and deep learning in a paper published Monday in the journal Methods in Ecology and Evolution.
"A major challenge for the application of individual recognition using deep learning methods is the need for collecting extensive training data," the researchers write. "Acquiring training data typically involves labeling, [instead] we provide an efficient pipeline for collecting training data, both in captivity and in the wild, and we train CNNs for individual re-identification."
Instead of chasing down birds to take their photos in a number of different environments, the researchers used RFID chips and motion-sensor cameras in feeding areas to capture and automatically label images of birds from different angles, providing a robust image of what an individual bird looks like.
By collecting photo and RFID data of birds the researchers were able to train an A.I. to recognize individuals in a species.
How does it work — The researchers tested out their new approach using three common species of birds: the sociable weaver (Philetairus socius,) the Great tit (Parus major,) and the Zebra finch (Taeniopygia guttata.)
In both wild and captive environments, the researchers used RFID sensors on the birds and autonomous cameras to capture a thousand images per individual of the species.
Labeling these photos with information such as what type of bird was in the image is a process that would typically take hundreds of hours of tedious work. Combining these two technologies actually made the process seamless. It also helped the researchers form a large dataset to feed to their A.I. algorithm.
The more data provided — the better the A.I. is.
To make sure the algorithm wasn't simply learning to memorize the different markings on these birds (a process called "overfitting") the researchers used part of their dataset to train the algorithm and set the other part aside to use as a validation set, or essentially a way to check the A.I.'s work.
What were the results — After setting the A.I. loose to check just how well it had learned about these birds, the researchers found that it demonstrated 92 percent accuracy when identifying sociable weavers, 90 percent accuracy when identifying Great tits, and 87 percent accuracy when identifying Zebra finches.
The researchers do write that introducing new individual birds to the party did make it harder for the A.I. to accurately identify these individuals. The authors report that there was a 17 percent chance of misidentification when new birds were introduced.
Scientists used motion sensor cameras to take a photo of a bird, like this Great tit, every two seconds to form a robust dataset.
What's next — The researchers write that the success of their model represents a new era of technological opportunity for ecologists.
"[T]he ability to move beyond visual marks and manual video coding will revolutionize our approach to addressing biological questions," write the authors. "Importantly, it will allow researchers to expand their sample sizes, thereby providing more power to test hypotheses. Finally, it will open up opportunities to address questions that previously were not tractable."
But, that said, there are several elements that researchers say need to be refined with their algorithm. While the algorithm showed high accuracy identifying individuals in a given location during a given season the authors write that more research needs to be done to see how this level of accuracy will stand the test of time. Similarly, a more robust dataset in the future should also include photos of these birds at different stages of their life, like when molting.
Being able to identify individuals through time and space could be an incredibly powerful tool for understanding long-term animal behavior.
Abstract: Individual identification is a crucial step to answer many questions in evolutionary biology and is mostly performed by marking animals with tags. Such methods are well-established, but often make data collection and analyses time-consuming, or limit the contexts in which data can be collected. Recent computational advances, specifically deep learning, can help overcome the limitations of collecting large-scale data across contexts. However, one of the bottlenecks preventing the application of deep learning for individual identification is the need to collect and identify hundreds to thousands of individually labelled pictures to train convolutional neural networks (CNNs). Here we describe procedures for automating the collection of training data, generating training datasets, and training CNNs to allow identification of individual birds. We apply our procedures to three small bird species, the sociable weaver Philetairus socius, the great tit Parus major and the zebra finch Taeniopygia guttata, representing both wild and captive contexts. We first show how the collection of individually labelled images can be automated, allowing the construction of training datasets consisting of hundreds of images per individual. Second, we describe how to train a CNN to uniquely re-identify each individual in new images. Third, we illustrate the general applicability of CNNs for studies in animal biology by showing that trained CNNs can re-identify individual birds in images collected in contexts that differ from the ones originally used to train the CNNs. Finally, we present a potential solution to solve the issues of new incoming individuals. Overall, our work demonstrates the feasibility of applying state-of-the-art deep learning tools for individual identification of birds, both in the laboratory and in the wild. These techniques are made possible by our approaches that allow efficient collection of training data. The ability to conduct individual recognition of birds without requiring external markers that can be visually identified by human observers represents a major advance over current methods.
