Shila Fallahy, Nima Rezazadeh
Detecting cracks in historic masonry from real-world images is challenging because models trained on curated datasets often perform poorly when applied to new sites. This study introduces MARBLE-DA, an explainable and label-free approach that adapts deep learning models for masonry crack detection without requiring new annotations. The method standardises image inputs using Sobel-edge preprocessing and updates a pretrained classifier through shallow adaptive batch normalisation combined with entropy-minimisation and confidence-guided refinement. Predictions are calibrated to maintain balanced sensitivity and precision, and self-refinement is selectively activated under strict confidence conditions. The framework was evaluated across 4 cross-domain scenarios, consistently achieving strong accuracy and reliability whilst remaining robust under class imbalance. Explainability analyses using Local Interpretable Model-Agnostic Explanations, Gradient-weighted Class Activation Mapping, and occlusion sensitivity confirmed that the model's decisions relied on genuine crack structures rather than surface textures, improving transparency and practitioner confidence. MARBLE-DA also demonstrated computational efficiency on standard hardware, with runtime scaling proportionally to dataset size. These findings establish MARBLE-DA as a reproducible and interpretable pathway for applying deep learning to heritage masonry inspection. • Introduced MARBLE-DA, an explainable label-free framework for unsupervised domain adaptation in masonry crack detection. • Achieved superior cross-domain performance with balanced precision and recall, surpassing established adaptation baselines. • Integrated calibrated prediction and post-hoc visual explanations (LIME, Grad-CAM, occlusion) to enhance transparency. • Delivered a reproducible, efficient pipeline aligned with heritage workflows and deployable on standard hardware platforms.