Multi-temporal Crack Segmentation in Concrete Structures using Deep Learning Approaches

“This paper addresses the importance of structural health monitoring for civil concrete structures, which are vital to modern society. Cracks are one of the earliest indicators of structural decay, therefore, early detection is important. Using deep learning, the study focuses on pixel-level image segmentation to identify cracks. For this, the temporal domain of crack progression is addressed in this work. I manually created and annotated a multi-temporal crack propagation dataset for semantic segmentation, consisting of 1356 samples, each illustrating crack growth on concrete over a series of 32 sequential images. A Swin Transformer adapted for three-dimensional input (Swin UNETR) was trained on the multi-temporal dataset. Its performance was compared to a mono-temporal model (U-Net) that was trained on exactly the same, but deserialized, data. Results demonstrate that the multi-temporal approach surpasses the mono-temporal model in terms of IoU and F1-score significantly, alongside superior visual predictions, all while using half the parameters of the mono-temporal model. The multi-temporal approach holds potential for long-term structural health monitoring of concrete structures, even with limited sequential data.”

Recommended citation: Harb, S., Achanccaray Diaz, P., Maboudi, M., and Gerke, M.: Multi-temporal crack segmentation in concrete structures using deep learning approaches, ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., X-G-2025, 341–348, https://doi.org/10.5194/isprs-annals-X-G-2025-341-2025, 2025.
Download Paper