Explainable classification of goat vocalizations using convolutional neural networks

Efficient precision livestock farming relies on having timely access to data and information that accurately describes both the animals and their surrounding environment. This paper advances classification of goat vocalizations leveraging a publicly available dataset recorded at diverse farms breeding different species. We developed a Convolutional Neural Network (CNN) architecture tailored for classifying goat vocalizations, yielding an average classification rate of 95.8% in discriminating various goat emotional states. To this end, we suitably augmented the existing dataset using pitch shifting and time stretching techniques boosting the robustness of the trained model. After thoroughly demonstrating the superiority of the designed architecture over the contrasting approaches, we provide insights into the underlying mechanisms governing the proposed CNN by carrying out an extensive interpretation study. More specifically, we conducted an explainability analysis to identify the time-frequency content within goat vocalisations that significantly impacts the classification process. Such an XAI-driven validation not only provides transparency in the decision-making process of the CNN model but also sheds light on the acoustic features crucial for distinguishing the considered classes. Last but not least, the proposed solution encompasses an interactive scheme able to provide valuable information to animal scientists regarding the analysis performed by the model highlighting the distinctive components of the considered goat vocalizations. Our findings underline the effectiveness of data augmentation techniques in bolstering classification accuracy and highlight the significance of leveraging XAI methodologies for validating and interpreting complex machine learning models applied to animal vocalizations.