Pothole detection using multispectral sensor and unmanned aerial vehicle imagery / Muhammad Hafiz Aizuddin Mohd Zaidi

Today, it is truly stimulating for the road management department to rapidly obtain large-scale technical insights into road pavement conditions, particularly with the rapid expansion of road networks, especially highways. In earlier times, conventional methods such as field investigations and manual measurements were utilized to collect data and assess pavement distresses. The aim of this study is to evaluate the effects of multispectral sensors on pothole detection using very high-resolution images captured by Unmanned Aerial Vehicles (UAVs) based on the structure-from-motion photogrammetry approach. The study has three objectives: to evaluate the accuracy of 3D pothole estimations from UAV images compared to actual pothole data, to investigate the impact of multispectral band combinations on pothole edge detection, and to assess different algorithms for pothole area extraction using multispectral and visible images. Aerial photos were acquired using the Mavic 2 Pro quadcopter UAV, which conducted flight missions at varying altitudes for RGB imagery data. Additionally, the DJI Phantom 4, equipped with a multispectral sensor (Parrot Sequoia), collected multispectral imagery data during flights at a 10-meter altitude. The flight missions were conducted in two study areas with asphalt surfaces affected by potholes, where measurements and assessments were carried out to gather distress data. Pothole dimension data were obtained from manual on-site measurements and compared with automated measurements using 3D models processed in Agisoft Modeller software, revealing higher accuracy at a low altitude of 2 meters. The optimal band combination for pothole detection involved utilizing two or more bands, including the green and red bands, resulting in the highest accuracy. Furthermore, this study demonstrates that Support Vector Machine (SVM) consistently outperformed the Maximum Likelihood Classifier (MLC) in pothole classification, achieving an overall accuracy of 95.77% and 99.1% compared to MLC. The findings of this study can contribute to improve guidelines for local authorities, such as Jabatan Kerja Raya (JKR), and professionals in performing systematic pothole maintenance, enhancing existing methods such as the IKRAM Road Scanner (IRS), a specialized vehicle equipped with a wide array of survey products for scanning pavement distress.