DamCrackNet: A Novel Deep Learning Approach for Crack Severity Assessment on Concrete Dam Surfaces

Abstract

Concrete dams play a key role in flood protection and freshwater supply; however, their condition must be monitored reliably and precisely to prevent consequential damage. As concrete dams represent a highly unique field with unique damage patterns, various specialized deep learning-based inspection approaches have been developed in the past. However, these approaches remain limited to a binary assessment of damage. To address this gap, this study proposes DamCrackNet, a novel deep learning architecture for the automated detection and severity assessment of cracks in concrete dams. The proposed architecture builds on the characteristics of cracks of varying severities in such structures, such as simultaneous local texture changes and cracks visible in the global context, as well as highly heterogeneous materials, texture, and crack patterns. DamCrackNet is evaluated using five-fold cross-validation on the currently most comprehensive and diverse peer-reviewed dataset comprising 1,500 images across three crack-severity classes on concrete dam surfaces. Under this evaluation procedure, DamCrackNet achieves a mean balanced accuracy of 87% while completely avoiding catastrophic errors, specifically the misclassification of true high-severity cracks as easy cracks. Compared to a state-of-the-art ConvNeXt baseline, the architecture achieves a highly significant performance gain, quantitatively corresponding to an improvement in balanced accuracy ( $|d|=1.075, p\lt 0.01$ ). From a conceptual perspective, this demonstrates that by leveraging problem-specific characteristics, specifically designed deep learning systems are able to perform fine-grained assessments of complex quality control tasks in critical infrastructure. Practically, DamCrackNet contributes towards more automated and reliable dam monitoring.