Combination models of random forest for predicting seismic liquefaction based on SPT, CPT, Vs databases considering sampling strategies

The sampling strategy has an important impact on the accuracy of seismic liquefaction discrimination models. In addition, different models may produce contradictory discriminative results. This paper, based on three in situ experimental data (standard penetration test (SPT), cone penetration test (CPT), and shear wave velocity (V_s)), adopts the Random Forest (RF) method to analyze, the effects of five probabilistic sampling methods (Simple Random Sampling (SRS), Unordered Systematic Sampling (USS), Ordered Systematic Sampling (OSS), Stratified Random Sampling (StrRS), and Cluster Sampling (CS)) and five integration methods (sequential integration, voting, simple averaging, weighted averaging, and Bayesian model averaging) on the RF models of seismic liquefaction, and constructs three RF model based different in-situ tests data and a Combined RF Model (CRF). The results show that the sampling methods have a large impact on the performance of the RF model. Among them, the OSS method performed the best in different in-situ test databases with Acc = 0.9 and F₁ = 0.930 for the RF-SPT model (the RF model based on the SPT data), Acc = 0.88 and F₁ = 0.918 for the RF-CPT model (the RF model based on the CPT data), Acc = 0.872 and F₁ = 0.913 for the RF-Vs model (the RF model based on the V_s data), whereas, the CS method performed the worst in the datasets. In addition, sensitivity analysis of the RF models under the optimal sampling method was performed. In combined models, integration modes do not always improve model performance, and sequential integration fails to improve model performance in this study. However, the CRF based on the Bayesian model averaging method performed the best with Acc = 0.924 and F₁ = 0.947, which is better than the RF-SPT model.