Abstract
Vector-borne diseases pose a persistent public health challenge in tropical regions such as Vietnam, where traditional ground-based surveillance methods struggle with scale and accuracy. This study presents a framework that integrates Unmanned Aerial Vehicle (UAV) multispectral imagery with deep learning techniques to detect small-to-medium-sized water bodies, important habitats for arbovirus vectors. High-resolution multispectral images were captured with the DJI Phantom 4 (P4M) Multispectral UAV in rural and peri-urban areas of Binh Duong province in Vietnam. A curated dataset of 982 annotated images was created, comprising RGB, near-infrared (NIR), and normalized difference water index (NDWI) bands. Six state-of-the-art object detection and segmentation models were evaluated, including YOLOv7, YOLOv7x, DocF, U-Net, MSNet, and RTFNet. Among them, segmentation models (U-Net and MSNet) using RGB + Green + NIR + NDWI achieved the best performance with dice scores above 0.92. The results show that the combination of UAV multispectral imagery with deep learning significantly improves the detection accuracy of water bodies in complex tropical conditions. This approach provides a scalable, cost-effective solution for mapping small water bodies and contributes to targeted vector control and disease prevention measures in arbovirus-prone regions.
Introduction
Vector-borne diseases like dengue, malaria, and Zika remain a massive public health burden in tropical regions. Effective "vector control" starts with identifying where mosquitoes breed: small-to-medium water bodies like ponds, drainage ditches, and even roadside puddles. However, mapping these sites in a tropical environment like Vietnam is a significant challenge for traditional methods. Ground-based surveillance is too slow to cover large areas, and satellite imagery lacks the resolution to see small puddles. Furthermore, dense vegetation and frequent cloud cover often hide water from standard cameras. This research addresses these challenges by creating an automated, high-resolution system that can identify these tiny habitats within complex tropical landscapes.
Methodology
The research team implemented a comprehensive five-phase workflow to transition from raw aerial data to actionable predictions. This process encompasses study design, data collection, image preprocessing (orthomosaic generation), feature engineering (NDWI calculation), model training, and rigorous evaluation.
The workflow benchmarks two primary deep learning architectures: object detectors (YOLOv7, YOLOv7x, and DocF), which use bounding boxes, and semantic segmentors (U-Net, MSNet, and RTFNet), which classify every pixel to delineate exact boundaries.
Data Collection and Study Area
The study focused on rural and peri-urban areas within the Binh Duong province of Vietnam. Using a DJI Phantom 4 Multispectral (P4M) drone, the team captured high-resolution imagery across five spectral bands: Red, Green, Blue, Near-Infrared (NIR), and Red Edge.
A curated dataset of 982 annotated images was developed, featuring challenging classes such as "invisible water" (obscured by vegetation) and "irrigated water" (mixed with crops). The sub-decimeter resolution (approximately 2.6–3.5 cm per pixel) allows for the detection of very small breeding sites that satellite imagery would miss. The dataset and implementation are available on GitHub at WaterMAI.
Key Insights and Results
The study revealed that multispectral data is critical for high accuracy. While standard RGB imagery often confuses shadows or dark soil with water, the NIR band and the Normalized Difference Water Index (NDWI) provide the necessary contrast for reliable detection. The models were even able to learn to identify "invisible water" by recognizing the environmental patterns surrounding obscured water bodies.
Semantic segmentation models significantly outperformed bounding-box-based object detectors. The U-Net and MSNet models, utilizing the "cognirndwi" configuration (RGB + Green + NIR + NDWI), achieved Dice scores exceeding 0.92. In contrast, the best object detection model (YOLOv7x) reached a weighted F1 score of only 0.74, highlighting why pixel-level segmentation is the superior approach for capturing the irregular and fluid boundaries of small water bodies.
Conclusion and Future Work
This framework provides a scalable and cost-effective solution for mapping mosquito breeding grounds in complex tropical landscapes. By integrating multispectral UAV imagery with advanced deep learning, public health officials can better target vector control efforts. While the results are promising, future work will focus on expanding the dataset to include more diverse geographies and developing "lightweight" versions of these algorithms optimized for real-time deployment on edge devices.
Summary updated