Acoustic remote sensing with deep learning enables non-invasive estimation of seabird nest density

Abstract

Passive Acoustic Monitoring (PAM) is increasingly used as a non-invasive tool for biodiversity monitoring, particularly for species breeding in habitats where visual surveys are difficult or biased. The Critically Endangered African Penguin ( Spheniscus demersus ) is a highly vocal, burrow-nesting colonial seabird, making it an ideal species for evaluating whether soundscape data can be translated into reliable estimates of nest density. We deployed Autonomous Recording Units at multiple sampling points at the Stony Point Penguin Colony, South Africa, over two consecutive breeding seasons (2024 and 2025), capturing soundscapes spanning different nest densities and environmental conditions. We developed an automated detector for Ecstatic Display Songs (EDS), the species' territorial song, using a Convolutional Neural Network trained on a multi-source dataset from both in situ and ex situ recordings. The detector achieved high recall and precision, supporting the use of heterogeneous training datasets for reliable detection. To link vocal activity to nest density, we derived, for each sampling point, an index based on the maximum number of EDS detections across 30-min recordings and modelled its relationship with active nest counts using Generalized Additive Models. The index reliably predicted nest density, but the relationship was nonlinear: nest density rose as vocal activity increased before plateauing at higher levels. The model trained on 2024 data generalized well to independent 2025 data. Together, these results show that, up to intermediate densities, nest density can be reliably estimated from passive acoustic recordings alone. This approach requires no observer presence and is applicable to large-scale, long-term seabird monitoring.