Comparison of GRNN-MC and RF models for predicting soil hydrogeological and geotechnical profile using borehole data

Soil stratification plays a critical role in geological and geotechnical engineering, motivating the development of algorithms capable of reconstructing reliable subsurface profiles. Among these, the General Regression Neural Network–Markov Chain (GRNN-MC) model offers a promising data-driven approach, although it is susceptible to noise and outliers. In contrast, the Random Forest (RF) algorithm is widely recognized for its resistance to overfitting and data variability. This study compares the performance of GRNN-MC and RF in predicting soil profiles using borehole data. Results show that the GRNN-MC model requires approximately 79% of the RF runtime, whereas the RF model attains an average error of about 60% of that of GRNN-MC. When information entropy is used as an indicator of prediction uncertainty, the comparative behavior of the two models is strongly influenced by spatial mesh resolution. Under the coarser mesh used in the baseline RF simulations, the average RF entropy was about 67% of the GRNN-MC value. However, when both models were evaluated on the same refined mesh, RF exhibited substantially higher entropy—approximately three times that of GRNN-MC. Overall, although RF consistently shows lower prediction error, the two models demonstrate similar sensitivity to thin-layer detection and to mesh-resolution effects.