ESO: an Evolutionary Algorithm for Efficient Model Reduction Via Automated Frequency Band Selection in Passive Acoustic Monitoring

Abstract Passive acoustic monitoring (PAM) is an important tool for wildlife monitoring. Deep learning, particularly convolutional neural networks (CNNs), has become the standard approach for developing bioacoustic classifiers. However, real‐time classification remains challenging due to the high computational complexity of these models and the need for resource‐intensive preprocessing steps. One way to reduce model size and computational complexity is to decrease the input dimensions, specifically by reducing the mel‐spectrogram size—the visual representation of sound. We introduce Evolutionary Spectrogram Optimisation (ESO), an evolutionary algorithm that automatically identifies and extracts narrow frequency bands from mel‐spectrograms that are informative enough to allow CNN‐based species detection. ESO is designed to optimise the selection and size of each of these bands, while maximising classification score, and minimising the number of CNN parameters. It eliminates preprocessing steps such as low‐pass filtering and downsampling during mel‐spectrogram creation, optimising for real‐time processing. We evaluated ESO on three distinct PAM case studies targeting the detection of Hainan gibbon, Thyolo Alethe and Pin‐tailed Whydah vocalisations across diverse natural soundscapes. Compared to a baseline CNN trained on full mel‐spectrograms with preprocessing, ESO reduces mel‐spectrogram size by up to 57% while simultaneously decreasing CNN parameters by up to 72% and improving the F1‐score by up to 6%. Our findings revealed that ESO increases inference speed by 45%, reduces peak memory usage by 47% and lowers energy consumption by 56% for the Pin‐tailed Whydah dataset. To facilitate its use, we provide ESO as a Python package and a user‐friendly graphical user interface. A comprehensive user guide is available on the project's GitHub page https://github.com/ufuk‐cakir/ESO. This study serves as stepping stones towards facilitating the deployment of onboard models for PAM and advancing research in this direction. By reducing model complexity while maintaining high classification accuracy, ESO is a promising solution for real‐time PAM applications. It enhances computational efficiency by reducing model inference time, memory usage and power consumption, making the method well suited for onboard deployment. While the evolutionary algorithm requires computational resources to identify the optimal solution, its reusability and intuitive design make it a valuable tool for further research and practical applications.