Speakers
Description
Mosquito-borne diseases remain a major global health challenge, particularly in regions where surveillance infrastructure is limited. Acoustic monitoring provides a non-invasive, low-cost method for tracking mosquito populations, yet automated recognition remains challenging. Most existing approaches rely on small datasets and focus on binary detection or genus-level classification, with limited attention to finer traits such as sex and age that affect vector competence. This study introduces a deep learning system capable of classifying mosquito wingbeat sounds by genus (Aedes, Culex, Anopheles), sex (male, female), and age group (young, old), while also discriminating mosquito from non-mosquito sounds for robust field applicability. We collected 34,153 mosquito and 34,111 non-mosquito recordings in Ghana, which, to our knowledge, is substantially larger than prior African mosquito bioacoustic datasets. Wingbeat signals were converted into spectrograms and classified using a lightweight CNN with convolutional, pooling, dropout, and dense layers. The model was trained on 13 output classes using Adam optimization and achieved 92.2% accuracy on a test set of 5,866 samples. Across all mosquito classes (genus, sex, and age group), F1-scores ranged from 0.92 to 0.98. The dedicated non-mosquito class had an F1-score of 0.70. The inclusion of a dedicated non-mosquito class improved system robustness against environmental noise. By integrating large-scale data with deep learning, this framework advances automated mosquito surveillance beyond binary detection, offering finer epidemiological insights at scale. While the dataset was collected under controlled conditions, the system lays the groundwork for real-time, scalable, and cost-effective vector monitoring with strong potential for adaptation to field deployments in Africa and other endemic regions.