Abstract:
Dengue fever is a major public health concern in Sri Lanka, where early identification of
Aedes aegypti and Aedes albopictus larvae is essential for effective vector control. Manual
larval identification is time-consuming, expertise-dependent, and often unreliable under field
conditions. This study presents an automated deep learning approach for classifying mosquito
larvae using Swin Transformer architectures. A dataset of 5,939 images was curated by
combining images captured by the researchers with publicly available collections from platforms
such as Mendeley, Kaggle, and Roboflow, and annotated into three classes: Aedes aegypti,
Aedes albopictus, and non-Aedes (comprising other common mosquito genera). Images were
pre-processed with augmentation techniques, and three Swin Transformer variants Swin-
T, Swin-S, and Swin-B were fine-tuned using a customized Linear → ReLU → Dropout
→ Linear head, AdamW optimization, StepLR scheduling, and Label Smoothing Cross-
Entropy. Dataset refinement, guided by Gradient-weighted Class Activation Mapping
(Grad-CAM), enabled the removal of ambiguous samples. Swin-B achieved the highest
accuracy of 96.86%, outperforming conventional convolutional neural network (CNN) models
reported in the literature. The hierarchical shifted-window attention mechanism of Swin
Transformers effectively captured fine-grained morphological features, including siphon shape,
head structure, and thorax bristles, supporting reliable species-level differentiation. These
findings demonstrate the potential of transformer-based models for dengue vector surveillance
