Transfer learning for undrained shear strength prediction: A case study with instance-based transfer

• Domain shift between source and target soils confirmed with statistical tests. • Machine learning model showing need for adaptation in geotechnical prediction. • Instance-weighted KRR achieved stable, accurate USS prediction. • SHAP analysis identified density and moisture as dominant predictors of USS. Accurate estimation of undrained shear strength (USS) is critical for safe and cost-effective geotechnical design. However, limited availability of high-quality site-specific data often constrains traditional machine learning models, especially when applied across regions with heterogeneous soil characteristics. This study investigates instance-based transfer learning (TL) as a strategy to improve USS prediction in a cross-regional case study under domain shift conditions. A source dataset from Finland (217 samples) and a smaller target dataset from Sylhet, Bangladesh (134 samples) were used in this study. Statistical measures confirmed significant distributional differences between regions. Three TL methods, TrAdaBoost.R2, Two-Stage TrAdaBoost.R2, and instance weighted kernel ridge regression (IWKRR) were compared against a baseline AdaBoost.R2 trained solely on the target domain. IWKRR achieved the highest predictive accuracy ( R ² = 0.876, RMSE = 6.89 kPa, MAE = 5.46 kPa) and maintained stable performance even with low target data availability. SHapley Additive exPlanation (SHAP)-based interpretability analysis identified density and moisture content as the dominant features influencing predictions across both domains. This case demonstrates that instance-based TL such as IWKRR can effectively mitigate domain shift, reduce data dependency, and provide reliable USS predictions in low-resource geological settings.