Estimation of NATM excavation-support classes at tunnel excavation faces using a deep learning algorithm
The study proposes a deep learning model to classify NATM excavation-support classes from tunnel face photographs. Using MobileNetV2 and MobileNetV3‑Small on 245 photos from the BT‑05 tunnel in Osmaniye, Turkey, the models achieved macro‑F1 scores of 0.893 and 0.847 respectively, with MobileNetV2 maintaining 0.844 accuracy on an independent test. Grad‑CAM analysis confirmed predictions were based on geologically relevant features, demonstrating a fast, objective decision‑support tool for on‑site geotechnical assessment.
The study proposes a deep learning model to classify NATM excavation-support classes from tunnel face photographs. Using MobileNetV2 and MobileNetV3‑Small on 245 photos from the BT‑05 tunnel in Osmaniye, Turkey, the models achieved macro‑F1 scores of 0.893 and 0.847 respectively, with MobileNetV2 maintaining 0.844 accuracy on an independent test. Grad‑CAM analysis confirmed predictions were based on geologically relevant features, demonstrating a fast, objective decision‑support tool for on‑site geotechnical assessment.
The approach demonstrates how lightweight deep learning models can be applied to geotechnical data in tunnel engineering, offering a scalable solution that aligns with Swiss AI initiatives focused on digital transformation and efficient resource use in civil infrastructure projects.
Original source record. Open the original record to verify the underlying announcement.