Automated classification of albatross acoustic behaviour at sea: A free and open-source classifier for seabird sounds

Advancements in acoustic data collection technologies have greatly increased their use in wildlife monitoring, but produce large volumes of data that are challenging to analyse manually. Recent developments in machine learning, particularly convolutional neural networks (CNNs), have transformed audio data analysis, enabling efficient and accurate sound classification. This study aimed to develop a method for automatic classification of behaviour (on-water activity, flight, vocalisation and preening) from sounds recorded by free-ranging albatrosses of two species equipped with audio recorders during foraging trips at sea. Using a manually labelled seabird audio dataset, a general-purpose CNN model was created and trained in Google Colab. The model development followed a structured workflow, including audio data preparation, pre-processing, model architecture and training, and performance evaluation. The model achieved a global accuracy and precision of 95 % during testing. Despite high overall accuracy, performance varied across sound categories due to the inherent complexity of distinguishing behaviours, leading to differences in prediction errors. This study primarily focused on developing and validating an accessible, high-performance workflow for automated acoustic classification, with the goal of enabling future ecological and conservation applications. It demonstrated that a generic web-based CNN model can effectively classify seabird sounds into different behaviours with high accuracy. The approach provides a foundation for future ecological and conservation applications, enabling detailed exploration of activities, interactions and environmental context of seabird behaviour using acoustic data. By leveraging open-source platforms and accessible tools, this work provides a foundation for future advancements in automated acoustic monitoring, making it accessible to a diverse range of researchers. • Bird-borne audio recordings reveal seabird behaviour and environmental context at sea. • General-purpose CNN model achieves 95 % accuracy in classifying seabird sounds. • Web-based framework simplifies seabird sound analysis for diverse researchers. • Study highlights open-access resources' role in advancing ecological informatics.