Sound-based Mosquito Classification via Featurization and Machine Learning

Mosquito-borne diseases impact over 3 billion people globally, causing more than 600,000 deaths each year. Precise mosquito species identification is crucial for outbreak prediction but is limited by manual, subjective methods requiring specialized skills and equipment, hindering scalability. Climate change exacerbates this by altering habitats. Citizen science, leveraging smartphone-captured mosquito sounds, offers an alternative option for scalable identification and tracking. Low-cost identification techniques can enable broad deployment in sensors and other inexpensive devices for non-invasive mosquito tracking in remote areas, reducing human disease exposure. Our work investigates sound featurization and machine learning for non-invasive mosquito type identification. Using a dataset of 200 smartphone-recorded mosquitoes, sound featurization, and Random Forest classification, we show that two common breeds (Anopheles and Culiseta) can be distinguished with over 98% accuracy, and that Mel Frequency Cepstral Coefficient (MFCC) Featurization outperforms other methods (spectrograms and spectrogram featurization) for this 2-class problem but underperforms in distinguishing between Anopheles variants.