Estimation of NATM excavation-support classes at tunnel excavation faces using a deep learning algorithm

Accurate determination of the geotechnical conditions at the tunnel face is essential for ensuring safe and effective excavation and support operations during tunnel construction with New Austrian Tunnelling Method (NATM). Therefore, engineering geologists carefully and systematically collect geotechnical data from the excavation face to ensure reliable NATM excavation-support classes. The development of artificial intelligence algorithms has provided significant advantages for such applications. In this study, a deep learning model is proposed to identify NATM excavation-support classes from tunnel face photographs obtained during the construction of a tunnel using the NATM. The BT-05 tunnel located in Osmaniye province (Türkiye) was selected as the data source for the study. After screening for near-duplicate photographs, the dataset comprised 245 photographs of 209 distinct excavation faces. Because the distinction between the C3-1 and C3-1 A support classes is established from instrumentation rather than from the appearance of the face, these were combined, and three excavation-support classes (B2, C2, and C3-1) were classified. Lightweight MobileNet architectures were fine-tuned with a custom classifier head, with class imbalance addressed through a Focal Loss objective and weighted oversampling. Evaluation used a nested, face-grouped cross-validation in which all photographs of one excavation face were confined to a single fold and every model-selection decision, including Bayesian hyperparameter optimisation, was performed on validation data drawn from within the training partition; the procedure was repeated over three random seeds with bootstrap confidence intervals. Under this protocol, MobileNetV2 achieved a macro-F1 of 0.893 ± 0.019 (accuracy 0.905 ± 0.010) and MobileNetV3-Small 0.847 ± 0.051 (accuracy 0.878 ± 0.020), matching substantially larger architectures at a fraction of their parameter count and CPU latency. On an independent tunnel section withheld entirely from training and model selection, MobileNetV2 retained an accuracy of 0.844 and a macro-F1 of 0.783, while performance under partitioning by contiguous excavation block was lower, delimiting the conditions under which the model can be applied. Gradient-weighted Class Activation Mapping (Grad-CAM) analysis further confirmed that predictions are driven by geologically meaningful regions of the tunnel face rather than background artefacts. The results conclude that the proposed deep learning-based approach provides an objective, fast, and cost-effective decision-support tool for on-site geotechnical assessment in NATM tunnel construction, contributing significantly to the digital transformation of tunnel engineering.