A comprehensive model optimisation for YOLO model-based recognition of pavement structural damage in radar images

Ground Penetrating Radar (GPR) and object detection models are increasingly utilised for pavement assessment, yet the specific impact of training configuration optimisation on GPR-based damage detection remains under-explored. This study systematically investigates the engineering effects of dataset size, model scale (YOLOv5s, YOLOv5m, YOLOv5l) and hyperparameter configurations on the detection of pavement subsurface defects (e.g. cracks and looseness). Using a field-validated dataset of 1000 GPR images, the results demonstrate that for the sparse electrical features typical of radar waveforms, replacing the default SGD optimiser with the Adam optimiser and optimising the initial learning rate to 0.001 leads to a 20% increase in accuracy. The final empirically optimised YOLOv5l model achieved a mean Average Precision (mAP) of 88.2%. This study provides a verified, deployable configuration framework for automated GPR data analysis, advancing the precision and intelligence of highway maintenance operation systems.