Automatic acoustic classification of insect species based on directed acyclic graphs

Area of Science:

• Bioacoustics
• Machine Learning
• Entomology

Background:

• Accurate insect species identification is crucial for ecological monitoring and pest control.
• Traditional methods often struggle with large datasets and complex acoustic signatures.
• Wingbeat sound analysis offers a non-invasive approach to insect classification.

Purpose of the Study:

• To develop an interpretable and accurate insect classification system using acoustic data.
• To leverage directed acyclic graphs (DAGs) and hidden Markov models (HMMs) for enhanced classification performance.
• To provide a publicly available dataset and benchmark for insect bioacoustics research.

Main Methods:

• A directed acyclic graph (DAG) scheme was designed, with nodes employing hidden Markov models (HMMs).
• Mel-scaled spectrograms were extracted from insect wingbeat sounds to capture temporal acoustic features.
• The classification interpretability was achieved by analyzing the activated paths within the DAG.

Main Results:

• The proposed DAG-HMM approach demonstrated superior performance compared to state-of-the-art methods in insect species classification.
• Experiments were conducted on a dataset of 50,000 wingbeat sounds from six distinct insect species.
• The system successfully classified various mosquito species (Ae. aegypti, Cx. quinquefasciatus, Cx. stigmatosoma, Cx. tarsalis) and other insects (Musca domestica, Drosophila simulans).

Conclusions:

• The developed DAG-HMM framework provides an effective and interpretable solution for insect classification based on wingbeat acoustics.
• This approach advances the field of bioacoustics by offering a robust method for analyzing complex sound patterns.
• The findings highlight the potential of machine learning in ecological studies and pest management strategies.