Spatial Generalization of Machine Learning Models for CPTU-Based Stratigraphic Identification in Heterogeneous Marine Sediments

Accurate stratigraphic characterization is essential for offshore geotechnical engineering, where nearshore marine sediments commonly exhibit strong spatial variability and complex stratigraphic transitions. Although machine learning (ML) methods have shown promise for CPTU-based stratigraphic identification, random sample-based validation may overlook spatial dependence among CPTU measurements and consequently overestimate model generalization. This study develops a spatially validated ML framework using an engineering-specific CPTU–borehole database from the Jintang Subsea Tunnel project. Six ML algorithms were evaluated using five-fold stratified cross-validation and a leakage-controlled nested leave-one-borehole-out (Nested LOBO) framework to compare internal predictive performance with within-site spatial transferability. Results show that random validation consistently yielded higher scores than borehole-wise validation, underscoring the need for spatially independent assessment in heterogeneous marine deposits. Under Nested LOBO validation, XGBoost achieved the highest mean Macro-F1 (0.859 ± 0.058), while RF obtained the highest Accuracy (0.918 ± 0.034) and Cohen’s kappa (0.855 ± 0.071). Compared with these ML models, the Robertson soil behavior type method yielded a Macro-F1 of 0.521 ± 0.069, an Accuracy of 0.762 ± 0.121, and a Cohen’s kappa of 0.614 ± 0.133. SHAP analysis identified normalized cone resistance and friction ratio as the dominant contributors to model predictions, with contribution patterns broadly consistent with established CPT-based soil behavior interpretation. The proposed framework provides a practical basis for more reliable evaluation of CPTU-based ML models in marine stratigraphic identification.