Robustness enhancement of CNN-based mosquito larva species classification with out of distribution-aware augmentation and ensemble learning

Abstract

Accurate identification of mosquito larvae is important for vector surveillance and early control of mosquito-borne diseases, yet classification performance often degrades under blur, acquisition variability, and cross-domain shift. This thesis presents a robustness-oriented deep learning framework for mosquito-larva species classification. A benchmark dataset, MosquitoLarvae-7400, was developed with 7,463 high-resolution images of four medically important species captured using laboratory microscopes and smartphone-based macro imaging under varied visual conditions. A comparative benchmark across multiple CNN and YOLO based classifiers was conducted, followed by the development of Dilated YOLOv8 as the principal backbone. To further improve robustness, the framework incorporated OOD-aware augmentation, a three-branch specialized design with clean, standard augmented, and OOD-aware branches, and late-fusion ensemble strategies for internal and external evaluation. The results showed that Dilated YOLOv8 achieved the strongest overall balance of accuracy, robustness, and computational efficiency among the selected models. Under internal degraded conditions, linear stacking was the most effective ensemble strategy, whereas under external out-of-distribution conditions, the OOD-specialized branch and safe ensemble fusion provided the most reliable performance. These findings demonstrate the value of combining robust backbone design, broader training distributions, and deployment-aware ensemble strategies for mosquito-larva classification.