Investigation of the Transferability of Measured Data for Application of YOLOv8s in the Identification of Road Defects: An SA-Indian Case Study

This study investigates the transferability of measured road-damage data between distinct geographic domains using the YOLOv8s deep-learning framework. A comparative evaluation was performed on two datasets: the locally developed RDD2024_SA (South Africa) and the publicly available RDD2022_India (India). Five training–testing scenarios were designed to analyze intra- and inter-dataset generalization, emphasizing the influence of dataset scale, annotation consistency, and class structure on detection performance. When trained and tested within the same domain, YOLOv8s achieved high accuracy (mAP@0.5 > 0.95), confirming the strength of localized feature learning. However, performance degraded substantially under cross-domain testing, revealing a sensitivity to differences in road texture, illumination, and labeling style. Reducing the number of classes from six to four dominant types improved stability (mAP@0.5 ≈ 0.78) by mitigating annotation noise and class imbalance. Furthermore, a transfer-learning configuration, in which the India-trained model was fine-tuned on 20% of the South-African dataset, achieved mAP@0.5 = 0.86, demonstrating effective recovery of cross-domain detection performance. These findings highlight the importance of domain-aligned data preparation, targeted fine-tuning, and balanced class representation in building robust and transferable AI systems for sustainable, data-driven road maintenance.