Deep learning-based interpretation of ground penetrating radar data for pavement condition monitoring

• Autores: Ahmed Elseicy
• Directores de la Tesis: Mercedes Solla Carracelas (dir. tes.), Henrique Lorenzo Cimadevila (dir. tes.)
• Lectura: En la Universidade de Vigo ( España ) en 2026
• Idioma: inglés
• Materias:
  • Ciencias tecnológicas
    • Tecnología de la construcción
      • Otras especialidades de tecnología de la construcción
• Enlaces
  • Tesis en acceso abierto en: Investigo
• Resumen

  • Road pavements are vital elements that should be in good physical condition to provide a safe and seamless flow to transport people and goods. However, they are subjected to various internal defects because of aging, environmental conditions, changing traffic load, and poor maintenance. Therefore, regular pavement inspections are performed to ensure their serviceability. Ground Penetrating Radar (GPR) is one of the nondestructive techniques that is widely used for inspecting road pavements subsurface. It is used to assess the condition of the pavement and the thickness of its layers. Furthermore, it can identify subsurface defects, such as cracks, voids, and subgrade settlements. Nevertheless, the current practice is to interpret GPR data using heuristic methods which are time-consuming and highly dependent on the user experience. Moreover, the output of the processing cannot be easily integrated with CAD and BIM software due to lake of data interoperability. Recently, deep learning-based methods have been applied to interpret GPR data and to overcome the interpretation subjectivity challenge. However, choosing a suitable network architecture and post-processing techniques for feature extraction is still undergoing investigation. This research aims to apply state-of-the-art deep learning techniques to process GPR data of the road pavements automatically. The target features are pavement layer thickness and road defects, such as cracks, delamination, voids, and moisture damages. The proposed algorithms aim to reduce the processing time and detect the hidden patterns of the data, which will help to eliminate the interpretation subjectivity. Moreover, the results are automatically exported as a georeferenced 3D dataset with semantic labels instead of representing them as labeled radargrams or tabular reports. The labeled dataset can be integrated with the existing building information modeling (BIM) and geographic information systems (GIS) applications.