Impact of CPT Feature Selection on Random Forest Performance for Soil Behavior Type Classification

Accurate soil classification is essential for safe and economical geotechnical design such as in foundation selection, slope stabilization, and earthwork construction. Cone penetration testing (CPT) provides continuous resistance profiles, but conventional Soil Behavior Type (SBT) charts can be ambiguous near class boundaries. This study evaluates how progressive input feature sets affect a hyperparameter- optimized Random Forest classifier for normalized SBT (SBTn) prediction. A public dataset of 1,010,153 cleaned CPT datapoints from Austria and southern Germany was stratified, sampled, and divided into training and test subsets. Four feature sets were compared using accuracy, macro-averaged F1-score, and confusion matrices. The baseline set of depth, cone resistance qc, and sleeve friction fs achieved an accuracy of 0.9554 and a macro-averaged F1-score of 0.9397. Adding friction ratio Rf improved accuracy and macro-averaged F1-score metrics to 0.9648 and 0.9544 respectively. Adding the soil behavior type index Ic produced the best performing results with an accuracy of 0.9923 and a macro-averaged F1-score of 0.9845. However, Ic is calculated from the same normalized parameters used to assign SBTn labels. Feature sets that only use measured or simply derived quantities, such as depth, qc, fs, and Rf aremore informative for models trained on raw CPT measurements. Lastly, adding explicit overburden stresses provided no further improvement. The outcome of this study recommends a feature combination of depth, qc, fs, Rf, and Ic to assign Robertson SBTn classes automatically when Ic is already available. For the purposes of predicting soil classes from raw CPT measurements without Ic, this study recommends the feature combination of depth, qc, fs, and Rf.