Deep Learning Approach for Identifying Dengue Mosquito Larvae in Potential Breeding Sites Using Multispectral Images

Abstract

Abstract Dengue fever, transmitted by Aedes aegypti and Aedes albopictus mosquitoes, is a major public health concern, especially in developing countries like Sri Lanka. Traditional methods for identifying dengue mosquito breeding sites are often labor-intensive and need more coverage. This dissertation presents advanced deep-learning models that utilize multispectral imaging captured by drones to improve the detection and classification of dengue mosquito larvae, providing a more efficient and scalable approach. The research developed two models: Model 1 is designed to detect the presence of dengue mosquito larvae, achieving a validation accuracy of 95% and a testing accuracy of 86% in real-world conditions. Model 2 classifies the developmental stages of dengue mosquito larvae, reaching a validation accuracy of 73%, with a mean squared error (MSE) of 0.085 and an average absolute difference of 0.1491 in stage classification, demonstrating high precision. Additionally, the study investigates how varying water depths affect detection accuracy. Results show that shallower waters significantly boost model performance due to better visibility of larvae features. However, performance decreases in deeper waters, indicating that additional enhancements are needed to maintain accuracy in these conditions. This research not only progresses the field of vector control but also lays the groundwork for automated dengue mosquito surveillance systems that could potentially lower the incidence of dengue through timely and cost-effective prevention strategies.