Sex and age determination in European lobsters using AI-Enhanced bioacoustics

Abstract

Monitoring aquatic species presents considerable challenges due to their elusive nature and complex habitats. Consequently, the development and application of innovative, non-invasive approaches, such as Passive Acoustic Monitoring (PAM), are paramount for effective ecological assessment and management. The present study addresses these challenges by focusing on the acoustic emissions of Homarus gammarus (European lobster), a key representative species of rocky benthic environments that underpins valuable local fisheries and aquaculture ventures. A comprehensive understanding of lobster habitats, welfare, reproduction, sex, and age is critical for robust aquaculture management, ecological research, conservation strategies, and sustainable fisheries. While bioacoustic emissions have been successfully employed to classify various aquatic species using Artificial Intelligence (AI) models, such as fish, the present research specifically leverages lobster bioacoustics to classify European lobster by age group (juvenile and adult) and sex (male and female). Despite lacking vocal cords, different lobster species produce characteristic sounds, including stridulation (European spiny lobster, Panulirus elephas; Caribbean spiny lobster, Panulirus argus), buzzing or carapace vibrations (European lobster, Homarus gammarus; American lobster, Homarus americanus), rattling (tropical spiny lobster, Panulirus ornatus), and clicking or snapping sounds. These acoustic signals are amenable to classification using advanced computational AI models. The dataset was collected at Johnshaven in Scotland, at a local lobster facility operated by Murray McBay and Company. Hydrophones were installed underwater in concrete tanks to record lobster sounds. We evaluate the performance of Deep Learning (DL) models, specifically One-Dimensional Convolutional Neural Networks (1D-CNN) and One-Dimensional Deep Convolutional Neural Networks (1D-DCNN), and six commonly used Machine Learning (ML) models (Support Vector Machine, k-Nearest Neighbours, Naive Bayes, Random Forest, Extreme Gradient Boosting, and Multi-Layer Perceptron) for age and sex classification. Mel-Frequency Cepstral Coefficients (MFCCs) were used as baseline features for all models, while a Multi-Feature Fusion (MFF) approach was employed to confirm the consistency of classification performance. MFCCs are well established for robust audio feature extraction, and both MFCC and MFF yielded consistent classification results. Most models achieved classification accuracies exceeding 97% for adult versus juvenile differentiation, with the exception of Naive Bayes (91.31%). For sex classification, all models except Naive Bayes exceeded 93.23% accuracy with MFCC features. These results highlight the strong potential of supervised ML and DL approaches to extract age- and sex-related information from lobster sounds. Overall, this research demonstrates a promising non-invasive approach for lobster monitoring, conservation, and management in aquaculture and fisheries, supporting the development of real-world edge-computing applications for PAM of underwater species.