Abstract
Blood-sucking insects such as mosquitoes, ticks, and sandflies are vectors for various pathogens causing diseases like arboviroses, malaria, and Lyme disease. Climate change, global economic growth, migration, and increased trade significantly affect the distribution of these insects. For instance, the expansion of Aedes albopictus, a vector for Zika, chikungunya, and dengue, highlights the need for accurate species identification during surveys to mitigate health risks. Traditional morphological methods are often inadequate for damaged samples or extensive surveys, while biological protocols are costly and unsuitable for field analysis. Wing Interferential Patterns (WIPs) have emerged as a promising method for species identification, as these colored patterns on insect wings vary significantly among species.This study utilized the ARIM collection from the Institut de Recherche pour le Développement (IRD), consisting of over 100,000 insect specimens. Insect wings were dissected, placed on glass slides, and imaged using a Keyence™ VHX 1000 microscope. The resulting database contains 5,516 images of seven families (Culicidae, Calliphoridae, Muscidae, Glossinidae, Tabanidae, Ceratopogonidae, Psychodidae) and 21 genera. A Convolutional Neural Network (CNN) approach was developed for Diptera species classification. The CNN extracts hierarchical features through convolutional layers and classifies them using fully-connected and softmax layers. The confusion matrix showed minimal misclassification, with an overall accuracy of 95.5%, demonstrating the high taxonomic value of WIP imaging. Specific accuracies were: Anopheles 97.0%, Glossina 97.5%, Culex 97.0%, Aedes 97.1%, and Phlebotomus 96.2%. This study confirms the efficiency and semantic power of CNNs in identifying insect species using WIP imaging, providing a robust and rapid method for species identification essential for addressing the threats posed by emerging vector-borne diseases.