Intelligent classification method for surrounding rock in subsea tunnels: multi-source parameter perception and model construction

Accurate surrounding rock grade identification is essential for the safe and efficient construction of subsea tunnels. Conventional surrounding rock classification methods are mainly developed for mountain tunnels and general underground engineering, and they do not sufficiently characterize subsea-environment-related deterioration caused by high-head groundwater, seawater intrusion, hydrochemical erosion, and weathering. Based on the Jiaozhou Bay Second Subsea Tunnel, this study proposes a multi-source parameter perception and intelligent classification method for surrounding rock grade identification in subsea tunnels. Six indicators, including point load strength, number of discontinuity sets, discontinuity orientation, weathering degree, groundwater condition, and Cl⁻ concentration, were selected to characterize rock strength, rock mass structure, and subsea environmental deterioration. Spearman correlation analysis and variance inflation factor (VIF) testing were used to verify the rationality and independence of the selected parameters. A field database containing 300 samples was established using 3D laser scanning, point load testing, geological investigation, and hydrochemical testing. BP neural network, random forest, and XGBoost models were developed and compared using random search and ten-fold stratified cross-validation. The optimized XGBoost model achieved the best classification performance, with a Macro-F1 score of 0.97, and maintained a minimum performance of 0.961 under repeated stratified random splitting. The feature-ablation experiment further showed that incorporating weathering degree, groundwater condition, and Cl⁻ concentration improved the Macro-precision, Macro-recall, and Macro-F1 score from 0.881, 0.938, and 0.905 to 0.972, respectively. The results demonstrate that subsea environmental parameters are necessary for characterizing rock mass deterioration in subsea tunnels, and the proposed method provides a field-data-driven approach for rapid and objective surrounding rock grade identification.