A Semi-Supervised Framework for Road Condition Assessment: From Minor Surface Distress to Post-Disaster Failures

Abstract

Accurate road-condition segmentation is crucial for intelligent transportation systems (ITS), enabling autonomous driving and post-disaster recovery. Yet, most prior studies focus on binary or low-class segmentation, limiting applicability to complex scenes where subtle surface degradations and large-scale obstructions coexist. To address this gap, we present a unified framework for pixel-level segmentation across 11 road-condition classes using a hybrid dataset integrating the Post-Disaster Road Dataset Japan (PDRDD-J), the Social Media Image Dataset for Disaster Road Damage Object Detection (SoDR), and the Road Damage Dataset 2022 (RDD2022-J). The defined classes include background, road, manhole cover, alligator crack, linear crack, pothole, vehicle, natural blockage, structural blockage, sinkhole, and collapsed road. To learn this label space under limited supervision, we propose F-UNet, a Feature-Level Data Extractor (FLDE)-guided extension of U-Net integrated into a standard teacher–student self-training protocol and trained through progressive pairwise specialization. Unlike conventional U-Net variants that rely solely on global multi-class training, F-UNet derives FLDE guidance through pairwise road–damage patch training and injects it during decoding to better handle class imbalance and thin, low-contrast defects. Extensive experiments demonstrate consistent gains over strong baselines, supported by high-confidence class-wise analysis, confusion-matrix analysis, and error and uncertainty-based failure-mode characterization under disaster edge cases. On a held-out test set, the teacher model achieves 0.8652 mIoU, representing a 15.5% improvement over standard U-Net and outperforming U-Net++, DeepLabV3+, SegNet, and FCN under the same supervision setting. In the semi-supervised configuration, the student model further improves to 0.8727 mIoU. The F-UNet source code is available at https://github.com/EdaTsujimoto/F-UNET