Automatic Gravel Road Dust Monitoring Analysis Using Deep Learning Models and Small Uncrewed Aircraft Systems

Abstract Gravel, unsurfaced, or unpaved road networks account for 35% of the total roadways in the US, with some states such as Iowa having about 60% of their roadways unpaved, and regular monitoring is required to maintain and extend their service life. Among the major types of gravel road distress, dust clouds from passing traffic can cause significant environmental problems, impair visibility, and pose health risks to human respiratory systems. Popular technical manual for unsurfaced road maintenance management typically recommend assessment of dust-severity levels based on the severity of a dust cloud produced by a vehicle driving at 40 km / h ( 25 mi / h ). Because this method exposes assessors to unsafe conditions, its use is impractical for long road networks and high-dust severity conditions, so there is a need to explore automated and remote-sensing techniques for rating dust severity. Although previous studies have employed image-processing techniques to assess dust levels from images captured by vehicle-mounted cameras, dust accumulation on windshields and unstable camera attachments have reduced the reliability of such data collection. To address these limitations, this paper investigates the application of small uncrewed aircraft systems (sUAS) in rating dust clouds generated by traffic at various speeds. For this experiment, we collected dust data using sUAS from three gravel roads in Iowa and applied deep-learning models, specifically You Only Look Once version 8 large (YOLOv8l), to detect the truck driven for dust activation and residual network (ResNet) models to classify the level of dust severity in each sUAS video frame. The ResNet-152 classification model achieved the best result on the test data set with a weighted F1-score of 0.9867, highlighting a promising outcome for adoption of sUAS for dust monitoring.