With DeepWeaver, we aim to address one of the greatest challenges in studying small, social animals (like sociable weavers) in the wild: identifying and tracking individuals and their behaviour.

Studying sociable weavers requires capturing the birds, marking them with individually identifiable colour rings for visual identification, and collecting samples for sexing and genotyping. While essential, these methods are often time-consuming, logistically demanding, and limits the scope of the study in various ways. For example, only the individuals that are captured can be studied, collecting data on their behaviours is hard because most behaviours are fast, their small colour rings are easily missed.

The alternative, which we have been doing for many years to explore the different aspects of the sociable weavers’ social lives is to record extensive videos and photographs (e.g. at the nest and during foraging). These recordings are invaluable for answering key ecological and behavioural questions, but generate huge volumes of data that are costly and very slow to analyse manually.

Recent advances in Artificial Intelligence, particularly in Deep Learning, have the potential to transform how animals are studied in the wild. These technologies enable the development of non-invasive methods that can automatically identify individual animals and recognise their behaviours—without the need for physical marking or disturbance.

With DeepWeaver, we bring together a team of scientists and technical experts from Portugal (CIBIO-BIOPOLIS), France (CEFE-CNRS), Switzerland (UNIVERSITAT ZURICH), and South Africa (UNIVERSITY OF CAPE TOWN) to develop these cutting-edge methods using the sociable weaver as a study system, creating tools and approaches that can later be applied to other species and ecological contexts.

1. 1. Individual identification

We are developing methods using deep learning for automatic identification of: individuals, their age, sex and relatedness, which are undistinguishable by the human eye. To achieve this, we use camera traps placed at artificial feeding stations, where birds equipped with PIT-tags (Passive Integrated Transponders) are automatically detected by RFID antennas. This setup allows us to collect labelled photographs, which are automatically linked to each bird’s identity through its PIT-tag, providing a powerful dataset for training and testing our deep learning models.

• 2. Automated annotation of behaviours

We are using our collection of videos of nestling provisioning, nest building, roosting and vigilance (collected over >10 years) and our manual annotations of those video recordings behaviours to train deep learning models for automated behaviour extraction.

• 3. Pipeline

We are working towards bring all these automation efforts together to establish a processing pipeline capable of handling these large volumes of video data. This integrated system will combine individual identification with behavioural recognition, creating a scalable framework that can be applied within the sociable weaver project and easily adapted for use in other species and research contexts.