Real-Time Mosquito Breeding Site Detection Using a Quantized YOLOv8-Tiny Model on an FPGA-Based Drone Platform

Abstract

The increasing prevalence of mosquito-borne diseases demands scalable, accurate, and energy-efficient surveillance systems capable of real-time operation in resource-constrained environments. This extended study presents an FPGA-optimized, drone-based mosquito breeding site detection framework leveraging quantized deep learning models for edge deployment. Building upon prior work, this paper introduces an enhanced implementation of YOLOv8-Tiny and YOLOv9-Small architectures, optimized through INT8 quantization, batch normalization folding, and FPGA-aware architectural refinements for execution on the PYNQ-Z2 platform. High-resolution aerial imagery acquired from unmanned aerial vehicles (UAVs) is processed in real time using the proposed system to identify and classify potential mosquito breeding sites such as stagnant water bodies and container habitats. Experimental evaluations carried out on six state-of-the-art object detection architectures showed that the proposed quantized YOLOv8-Tiny variant offers the optimal compromise among accuracy, speed, and energy efficiency by achieving 90.2% accuracy in the field, 20 FPS performance, and consuming 7.8 W of power. It is also proved that the proposed system reduces DSP and BRAM resources by more than 20% over the floating-point processor. The results obtained in the real-world deployment scenario proved the robustness of the system under different environmental situations. It is also found that the system can cover 10 km² of area per hour while providing a 44% reduction in operational costs over traditional methods. The proposed framework using FPGA acceleration and quantization awareness is useful for developing efficient vector surveillance systems.