Acoustic Identification of Aedes aegypti: Experimental Assessment of Different Machine Learning Techniques

Mosquito-borne diseases affect over 700 million people annually, demanding effective and scalable monitoring solutions. Traditional traps, while reliable, are costly and have limited coverage. Recent studies have shown that mosquito wingbeat audio can enable species identification using convolutional neural networks, but alternative machine learning (ML) models remain underexplored. This study compares five ML architectures – Residual Network, Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), Conformer, and Audio Spectrogram Transformer (AST) – for classifying Aedes aegypti from audio data. Using public datasets and standardized preprocessing, we evaluate performance through cross-validation and multiple classification metrics. All architectures analyzed achieved accuracy and F1-scores above 88%, with particular emphasis on MLP and LSTM which, despite their simpler structures, showed competitive performance compared to other networks. These findings support the development of low-cost, audio-based mosquito monitoring systems.