Condition Assessment of Low-Volume Roads in Norway: A Deep Learning Approach for Automatic Assessment at Network Level

Abstract Pavement condition monitoring is an important aspect of an efficient Pavement Management System (PMS). However, the monitoring of road sections is a tedious work. Therefore, research communities have come up with deep learning based automatic methods for detecting and classifying road damages. But for a highway authority overlooking a large network of roads, an aggregate measure of road condition is more useful. Therefore, in this work, the performance of different methods for quantifying road condition has been tested with respect to subjective rating at network level. The road condition was quantified using Damage Count (DC), Damaged Area (DA) and Weighted Damage Count (WDC) approaches using algorithms based on YOLOv8 and DeepLabV3 models for detection and segmentation, respectively. An experiment was followed by expert ratings to determine Mean Panel Rating (MPR) for 36 sections on a selected road stretch in Norway. Proposed models to determine Section Condition Rating (SCR) based on the damage count and weighted damage count approaches were able to predict MPR with an R 2 value of 0.93. Further, the SCR WDC was found to have a reasonable correlation with International Roughness Index (R 2 equal to 0.79).