Intelligent recognition of boulders in shield tunnels based on GPR dataset optimization using SE-EMD-WTD

Coastal urban strata containing numerous boulders that significantly affects shield-tunneling construction. In addition, urban environments are characterized by high noise levels , making the interpretation of ground-penetrating radar (GPR) images challenging. This paper proposes a combined denoising method based on the Shannon entropy (SE) theory, integrating empirical-mode decomposition (EMD) and wavelet-Threshold denoising (WTD). The feasibility of using deep-learning methods to identify solitary rocks was validated using a shallow convolutional neural network (CNN). Subsequently, transfer-learning and data-augmentation strategies were employed to optimize the Faster Region-based (R-)CNN. Finally, validation was performed using onsite GPR data . The results showed that the signal-to-noise ratio (SNR) of the GPR images denoised using the SE-EMD-WTD algorithm increased by 45.43 % compared with EMD and by 27.84 % compared with WTD. The root-mean-square error (RMSE) decreased by 35.32 % compared with EMD and by 19.37 % compared with WTD. Shallow CNN achieved a final recognition accuracy of 92.6 % on the GPR image-validation dataset. In actual engineering practice, boulder-recognition results based on the Faster R-CNN were consistent with the drilling results. • A new denoising algorithm is proposed. • Faster R-CNN is employed for intelligent identification of boulders in shield tunnel sections. • The accuracy of Faster R-CNN is validated by field drilling results.