Deep learning-assisted automatic road crack detection from UAV-based photogrammetric data

DÖŞ M. E., Yigit A. Y., Uysal M.

PAMUKKALE UNIVERSITY JOURNAL OF ENGINEERING SCIENCES-PAMUKKALE UNIVERSITESI MUHENDISLIK BILIMLERI DERGISI, cilt.32, sa.4, ss.751-763, 2026 (ESCI, TRDizin)

Özet

Surface cracks in pavement structure, if not addressed in a timely manner, can lead to structural deterioration and increased maintenance costs. Early detection of cracks, particularly on critical components such as bridge surfaces, is essential for preserving structural integrity. However, traditional inspection methods are timeconsuming, costly, and often rely on subjective assessments. This study proposes an integrated, next-generation approach for crack detection that combines Unmanned Aerial Vehicle (UAV) imaging, photogrammetric modeling, and deep learning techniques to ensure both visual and positional accuracy. High-resolution images collected along a rural road section were evaluated using a photogrammetry workflow. In the accuracy analysis performed, the internal orientation (camera calibration) error was determined to be 1.83 mm after balancing. External accuracy (three-dimensional (3D) positional accuracy) was verified with a Root Mean Square Error (RMSE) value of 2.14 mm. On the resulting orthomosaic images, automatic crack detection was performed using Transformer-based CT-CrackSeg. The predicted crack masks were validated against field observations and reference measurements, yielding 92.5% Precision, 88.3% Recall, 90.3% F1-Score, and 87.6% Intersection over Union (IoU). The results demonstrate that the proposed method provides a reliable, repeatable, and practical solution for fast and accurate crack detection, particularly in low-traffic and structurally sensitive environments. The integration of high-resolution, georeferenced imagery with deep learning enables both morphological and spatial analysis of cracks, offering a powerful use case in current literature. This approach is well-suited for integration into road maintenance management systems and can support the development of proactive, data-driven decision support mechanisms.