A Comparative Study of Deep Neural Networks for Aedes Mosquito Species Acoustic Wingbeat Classification

Abstract

In tropical regions, mosquito borne disease such as dengue and malaria remain major public health challenges. The common approaches to detect mosquito species are inefficient for a large scale surveillance, time consuming and have the risk of human error.  To address this issue, this study explores advanced identification approach using deep learning applied to mosquito wingbeat audio. Wingbeat signals were transformed into spectrogram images using the Short Time Fourier Transform (STFT), providing frequency representations as inputs for classification models. This paper examines several deep learning architectures which are convolutional neural networks (CNNs) based, such as ResNet, DenseNet, and MobileNet, as well as Capsule Networks (CapsNet) and Vision Transformer (ViT). Experiments were conducted using the Kaggle mosquito wingbeat audio dataset, with an 80/20 split for training and validation. Results showed that the Vision Transformer achieved the highest classification accuracy at 85%, followed by DenseNet at 83.5%, CapsNet at 80%, MobileNet at 79.5%, and ResNet at 78%. Our findings show that deep learning based analysis of wingbeat spectrograms especially ViT is an effective method for mosquito species identification. Among the tested models, it is proved most effective, highlighting its potential to enhance vector surveillance and support improved public health interventions.