Computational methods for detecting insect vibrational signals in field vibroscape recordings

Abstract

The ecological significance of vibroscape has been largely overlooked, excluding an important part of the available information from ecosystem assessment. Insects rely primarily on substrate-borne vibrational signalling in their communication, which is why the majority of terrestrial insects are excluded from passive acoustic monitoring. The ability to monitor the biological component of the natural vibroscape has been limited due to a lack of data and methods to analyse the data. In this paper, we evaluate the use of deep learning models to automatically detect and classify vibrational signals from field recordings obtained with laser vibrometry. We created a dataset of annotated vibroscape recordings of meadow habitats, containing vibrational signals categorized as pulses, harmonic signals, pulse trains, and complex signals. We compared different deep neural network architectures for the detection and classification of vibrational signals, including convolutional and transformer models. The PaSST transformer architecture, which was fine-tuned from a pre-trained checkpoint demonstrated the highest performance on all tasks, achieving an average precision of 0.79 in signal detection. For signals with more than one hour of annotated data, the classification models achieved instance-based F1-scores above 0.8, enabling automatic analysis of activity patterns. In our case study, where 24-hour field recordings were analysed, the trained models (even those with lower precision) revealed interesting activity patterns of different species. The presented study, together with the dataset we publish with this paper, lays the foundation for further analysis of the vibroscape and the development of automated methods for ecotremological monitoring that complement passive acoustic monitoring and provide a comprehensive approach to ecosystem assessment.