Deep Learning-Based Classification of Anopheles stephensi Adult Mosquitoes with Enhanced Solutions for Data Imbalance.

Recently, a number of studies have emerged that demonstrate the power of integrating AI and citizen science for enhancing mosquito surveillance. Typical R&D has focused on detecting multiple species of mosquitoes from image datasets that are balanced. In our paper, we report findings on using AI methods and citizen-science images to detect a particularly dangerous and invasive mosquito - Anopheles stephensi - which is now a critical vector of malaria in both Asia and Africa. Given the practical significance of detecting this one species among many other mosquitoes in nature, we focused on class imbalance. Specifically, our imbalanced dataset consisted of 1195 images of An. stephensi mosquitoes and 6021 images of non-An. stephensi mosquitoes, all of which were taken via modern smartphones in varying backgrounds. We experimented with multiple state-of-the-art AI models and techniques to address class imbalances. Results reveal that the EfficientNet-B1 architecture performs best for detecting the An. stephensi class. Findings also demonstrate that downsampling the majority class results in very poor accuracies. The best classification results were obtained when the original (non-downsampled) dataset was used, and when the model was trained with focal loss, a loss function specifically designed to handle class imbalances. Grad-CAM visualizations further confirm that the model learned anatomically relevant features for species identification.