Surface-Domain Effects in the External Validation of a Mendeley-Trained CNN for Concrete Classification using SDNET2018

Concrete crack classification models frequently achieve high accuracy when evaluated on the same dataset used for model development; however, reliability may decrease when applied to external datasets with different surface textures, image conditions, and structural domains. This study evaluates the surface-domain effects in the external validation of a Mendeley-trained InceptionV3 model for concrete crack classification using SDNET2018. The source model was trained using the Mendelet Concrete Crack Images for Classification dataset and achieved 99.63% accuracy on a 6,000-image source domain test set. Without retraining or fine-tuning, the saved model was externally evaluated on SDNET2018 under four testing conditions: bridge deck, wall, pavement, and a prepared combined SDNET2018 test set. Classification accuracy and accuracy drop from the Mendeley-sourcedomain performance were used to assess cross-dataset generalization. The results showed clear variation across SDNET2018 surface domains. The pavement subset achieved the highest external validation accuracy at 90.01%, corresponding to the smallest accuracy drop of 9.62%. The bridge deck subset achieved 86.63% accuracy, with a 13.00% drop, while the wall subset obtained the lowest accuracy at 80.89%, corresponding to the largest accuracy drop of 18.74%. The prepared combined SDNET2018 test set achieved 86.25% accuracy, with a 13.38% drop from the Mendeley source-domain result. These findings indicate that although the Mendeley-trained InceptionV3 model retained reasonable external classification performances, its accuracy was not uniform across surface domains. The study demonstrates that evaluating SDNET2018 only as a combined dataset may obscure domain-specific weaknesses, particularly when one surface type produces substantially lower performance than others. Domain-specific external validation therefore provides a clearer understanding of cross-dataset robustness in CNN-based concrete crack classification. The findings support the need to consider surface-domain diversity when validating models intended for infrastructure monitoring applications. This work also aligns with SDG 9 by contributing to infrastructure monitoring innovation, SDG 11 by supporting safer and more sustainable built environments, and SDG 13 by encouraging data-driven maintenance practices that may help extend infrastructure service life and reduce unnecessary resource use.