Vision-based automated road construction progress monitoring: Improved u-net segmentation approach

Manual road construction progress monitoring (CPM) in road projects often leads to inefficiencies, delays, and cost overruns. While Unmanned Aerial Vehicle (UAV)-based photogrammetry has improved automated data collection, manual identification of as-built components and progress measurement still limits project control. This study introduces a UAV-based photogrammetry approach using deep learning segmentation to automate the detection and classification of road layers for automated road construction progress measurement. A road layer classifier, based on a pre-trained improved U-Net architecture with ResNet-34, was developed to classify and segment road layers, including the surface layer, base course, subbase course, and subgrade. To address the class imbalance, the classifier employs an inverse class frequency weighted cross-entropy loss, improving segmentation accuracy for underrepresented layers. The classifier achieved a validation accuracy of 97.87% with a loss of 8.08%. The method was validated with two real-world case studies through a web-based prototype, achieving high accuracies of 93.63 and 96.79% in automated progress measurement, confirming the feasibility of detecting pavement layers and measuring progress in real pavement construction projects. By automating critical aspects of road progress measurement, this approach significantly improves project control, enhances decision-making capabilities, and addresses the critical need for more efficient and reliable construction progress monitoring systems in the industry.