High-precision identification of groundwater states in tunnel: An image recognition-driven two-stage neural network approach

● A two-stage neural network model enables accurate identification and quantitative analysis of groundwater states on tunnel face. ● Two groundwater databases provide high-quality datasets for intelligent recognition. ● The classification model achieves 96.4% accuracy, while the segmentation model performs well in identifying different groundwater states. ● Engineering verification confirms high consistency, proving this model as an effective intelligent solution for tunnel engineering. To address the challenges of low efficiency, strong subjectivity, and insufficient intelligent data acquisition in surrounding rock classification of transportation tunnel under complex geological states, an intelligent image recognition model is established in this study. Firstly, A classification-segmentation database containing 1923 high-resolution tunnel face images is constructed. Then, a two-stage neural network model integrating the MobileNetV2 classification model and the DeepLabV3+ semantic segmentation model is proposed, sequentially achieving groundwater existence identification and groundwater state identification. Database tests show that the MobileNetV2 classification model achieves a prediction accuracy of 96.4% on 192 classification image samples, while the DeepLabV3+ semantic segmentation model attains a mean pixel accuracy (MPA) of 92.6% on 90 segmentation image samples. Finally, the superiority of the proposed model in this paper is confirmed through comparative model analysis and engineering case applications. This intelligent analysis model provides reliable technical support for risk assessment and decision-making in underground engineering.