A hybrid deep learning framework for identifying Aedes aegypti mosquitoes using CNN, ViT, and GNN architectures

Abstract Dengue has remained a major global health concern, increasing the need for accurate and scalable mosquito surveillance systems, particularly for identifying Aedes aegypti, the primary vector of dengue transmission. To address this challenge, a hybrid deep learning framework, MC-TransNet, has been developed to enhance mosquito species classification. The framework has integrated four complementary architectures—Big Transfer ResNet50x1 (BiT), DenseNet121, Vision Transformer (ViT), and Graph Neural Networks (GNN)—allowing both fine-grained morphological details and long-range structural information to be captured effectively.The model has been trained using a dataset of 2,640 images representing three medically significant mosquito species. MC-TransNet has achieved 98.55% accuracy, 93.10% precision, 100% recall, 96.42% F1-score, and a Matthews Correlation Coefficient (MCC) of 96.96%, demonstrating strong and balanced predictive performance across all classes. These results have surpassed the performance reported in earlier CNN-based and transformer-based studies.To support transparency and field applicability, explainable AI techniques—Grad-CAM, Grad-CAM++, and LIME—have been incorporated to highlight key anatomical regions influencing classification decisions. Overall, the findings have indicated that MC-TransNet offers a robust and interpretable solution for automated mosquito identification, with significant potential for strengthening real-time vector surveillance and dengue prevention efforts.