1666094 - Airfield Distress Detection Using YOLOV8 on Small Uncrewed Aircraft System Collected Data

Routine airfield inspection and maintenance are necessary to ensure the pavement's serviceable condition over its lifetime. Longitudinal, transverse, and diagonal (LTD) cracks, corner breaks, shattered slabs in Portland Cement Concrete (PCC) pavement; and longitudinal and transverse (L&T) cracks of Asphalt Concrete (AC) pavement consist of most of the airfield pavement distresses. Currently, an experienced inspector visually inspects and notes the location and details of these distresses, which are later used to provide a rating to the pavement. The process is laborious, time-consuming, and depends on the experience of the pavement inspector. In this research, an attempt has been made to automatically identify those distresses in Red-Green-Blue images using three variants of deep learning model YOLO8; YOLOv8-nano, YOLOv8-medium, and YOLOv8-large. YOLOV8 is a widely used off-the-shelf deep learning object detection model that allows rapid training and easy execution. A training dataset of 3,000 images collected using small Uncrewed Aircraft Systems (sUAS), also known as a drones, from four airports in Iowa and Michigan was developed. The transfer learning technique was used to train the models with the goal of reaching a loss value lower than 1. The training dataset passed through each model 300 times for adequate training. Each model reported an F1 score, a metric for deep learning model performance evaluation, of more than 0.85. This accuracy showed that crack-related distress detection using deep learning models could complement the traditional method of airfield pavement inspection.