Enhancing Unpaved Road Condition Monitoring in Uganda Using Smartphone Imagery and Deep Learning Gerald Obalim (a*), Mitsuharu Tokunaga (b)
a) Research Student, Department of Civil and Environmental Engineering, Graduate School of Engineering, Kanazawa Institute of Technology, Japan
*geraldobalim[at]gmail.com
b) Professor, Department of Civil and Environmental Engineering, Graduate School of Engineering, Kanazawa Institute of Technology, Japan
Abstract
Developing countries such as Uganda face significant challenges in maintaining vast networks of unpaved roads due to resource constraints, lack of real-time data, and manual inspection inefficiencies. Of Uganda^s estimated 150,000 km road network, less than 15% is paved, making the need for management solutions especially urgent. This study explores a low-cost, scalable framework for unpaved road condition monitoring using smartphone imagery and deep learning. Focusing on district roads in Northern Uganda, geo-tagged images were captured along selected road sections and annotated using the Visual Geometry Group (VGG) image annotator. A total of 360 images were sorted and divided into two sets for training and validation. The images were used to train and evaluate an Ultralytics YOLOv8 object detection model, capable of identifying visible defects such as potholes and surface erosion. Detected damages were quantified and visualized using QGIS, with road segments classified according to Uganda^s established road distress rating guidelines. Initial results demonstrate a strong correlation between the model^s output and conventional field assessments by experienced road engineers, thereby offering a faster and more objective alternative to manual inspection. This not only enhances consistency and decision-making but also enables more strategic deployment of maintenance resources. Shadow interference from roadside vegetation occasionally affected detection accuracy, suggesting potential benefits from future shadow-removal processing. The study introduces a foundational dataset for Uganda^s unpaved roads and establishes a transferable framework for similar applications in other low-resources contexts.
Keywords: Deep learning, road condition monitoring, smartphone imagery, unpaved roads
Topic: Topic C: Emerging Technologies in Remote Sensing
