Impact of Under-Sampling Techniques on the Performance of the Raveling Severity Classification Model Using ResNet50

Pavement distress datasets collected from field surveys often exhibit severe class imbalance, which degrades deep learning performance. To balance the class distribution, this study investigates two guided under-sampling strategies, cluster-based under-sampling (K-means) and density-based under-sampling (DBSCAN), applied to a real-world imbalanced raveling dataset collected from provincial highways in Khyber Pakhtunkhwa, Pakistan, using a ResNet50-based classification framework. The results indicate that under-sampling substantially improves minority-class recognition, with recall for Medium Severity raveling increasing by up to 25% and High Severity raveling by up to 9% compared to the original imbalanced model. DBSCAN-based under-sampling provides more stable class-wise performance and achieves the improvement in F1-score (8% medium severity and 5% for high severity raveling), demonstrating a more favorable precision–recall trade-off. These findings demonstrate that structured under-sampling, particularly density-based methods, enhances the reliability of pavement distress severity classification.