Presentation Information
[Data_Collection-05]Deep Learning–Based Assessment of Unpaved Road Surface Condition from Smartphone Imagery
*Gerald Obalim1, Mitsuharu Tokunaga1 (1. Kanazawa Institute of Technology (Japan))
Keywords:
Unpaved Roads,Smartphone Imagery,Deep Learning,International Roughness Index,Developing Countries
Unpaved roads form a substantial share of road networks in developing countries and are vital for rural accessibility, yet their surface condition is difficult to monitor due to limited budgets, large network extents, and the high cost of conventional roughness measurement equipment. Although smartphone imagery and deep learning have been explored for road damage detection, their relevance to roughness-oriented asset management of unpaved roads remains insufficiently understood. This study evaluates the feasibility of using smartphone imagery and deep learning to derive surface condition indicators for unpaved roads in Uganda, where more than 94% of the nearly 150,000 km road network is unpaved. Geotagged smartphone images were collected along selected unpaved road sections, resulting in a curated dataset of 1000 Roboflow annotated images. A YOLOv11-based deep learning model was trained to detect the road surface and key surface defects including, water and mud, vegetation, and general unevenness. Detected defects were quantified, aggregated at the road-section level, and visualized using GIS tools. To assess engineering relevance, image-based outputs were compared with conventional field inspections and complemented by a structured survey of road engineers and practitioners. The initial results show that defect-based rankings derived from smartphone imagery align closely with expert condition ratings. While limitations related to slow learning rate were observed, the findings demonstrate that deep learning–based analysis of smartphone imagery can provide scalable, decision-relevant surface condition indicators closely associated with roughness on unpaved roads. The study establishes a practical foundation for integrating image-derived indicators with roughness metrics such as the International Roughness Index (IRI) and highlights future research directions including calibration with measured IRI, data fusion with inertial sensors to enhance operational reliability.
