Exploring Deep Learning Algorithms for Aedes Mosquito Detection from Smartphone Images

The outbreak of dengue fever in recent years has become a grave public health concern as it has spread to 20 countries in South America and South Asia. As vectors of the flavivirus, several mosquito species belonging to the Aedes genus are responsible for transmitting dengue fever. Effective vector surveillance and control are essential in reducing dengue outbreaks. However, due to their minute variations in anatomical structure, it is challenging to identify Aedes mosquitos without expert entomologists using a microscope. In this regard, deep learning algorithms can play a vital role in identifying mosquitoes using smartphone-captured images and pave the way for deskilling automated vector surveillance, provided that sufficient training examples are available. In this study, we develop the "Aedes Mosquito Image Dataset (AMID)" consisting of smartphone-captured mosquito images consisting of 8 species: Aedes aegyptie, Aedes koreicus, Aedes albopictus, Culex pipiens, Armigeres subalbatus, Culex quinquifasciatus, Aedes japonicus and some other unknown species. The images are collected by trapping mosquitos in several locations in Dhaka, followed by image capturing and expert annotations. Additional image data is collected from open-access online repositories. A 5-fold cross-validation experiment is set up, and the sample size is increased through data augmentation. Later on, a number of pre-trained deep learning models were used to classify the various mosquito species: EfficientNetB0, DenseNet121, MobileNet, and ResNet101. The results show that EfficientNetB0 has achieved the best overall accuracy of 92.06 (±0.34)% while DenseNet121, MobileNet and ResNet101 have achieved an accuracy of 90.67 (±0.36)%, 89.64 (±0.38)% and 85.67 (±0.44)%. We have also developed a prototype smartphone web application that can identify mosquito species in real time from uploaded mosquito images.