Monte Carlo cross-validation driven benchmarking of CNN and transformer-based segmentation models for airfield sealed and unsealed cracks analysis

Recent applications of deep learning (DL) for automated crack detection and length quantification are important for pavement maintenance. However, most studies rely on a single training–validation split, limiting assessment of model robustness and performance variability. This study addresses this gap using Monte Carlo cross-validation (MC-CV) for segmentation and crack length quantification. The need for MC-CV-based robustness analysis was first demonstrated on the DeepCrack dataset, where U-Net showed ±1.06% variation in F1 score across fifteen replicates compared with previously reported single-split results. Five DL models, YOLOv8m, YOLOv9c, YOLOv11m, SegFormer-B1, and U-Net, were then evaluated on an in-house airfield dataset containing predominantly sealed cracks from Custer Airport (TTF) and unsealed cracks from Osage Municipal Airport (D02). An 85/15 training–validation split provided the best-balanced performance. Optimal confidence thresholds were 0.15 for YOLOv8m and 0.20 for YOLOv9c and YOLOv11m. YOLOv8m achieved mean intersection of union (IOU) and Dice scores of 59.5% and 73.5%, with crack-length errors of 15.5% and 8.8% for TTF and D02, respectively. Statistical reliability analysis showed that 98 MC-CV replicates achieved a ±0.5% margin of error at 95% confidence, with an average crack-length error of 11.6%. Overall, the findings highlight the importance of MC-CV for consistent evaluation of DL models for crack detection and length quantification, supporting its broader adoption in pavement inspection research.