Soil Type Factor Estimation Based on ERT Apparent Resistivity Using Supervised Machine Learning Models

Geologists and professionals traditionally conducted manual analysis when interpreting resistivity models acquired through geophysical techniques such as electrical resistivity tomography. It remains a challenge since it needs to be integrated with borehole drilling, which is known to be laborious, costly, and totally destructive. With that, this research aims to establish a soil type factor estimation based on resistivity values to establish an AI model that will help in the interpretation of soil layers for initial rapid decision-making in areas such as construction and agriculture. Thus, supervised machine learning regression models such as linear, lasso, ridge, k-NN (k-nearest neighbor), SVR (support vector regressor), and random forest, as well as some models added with interaction and polynomial features, were compared based on evaluation metrics to select the best performing model. The dataset employed was from an existing study that performed actual ERT surveying using a dipole-dipole electrode configuration. Fortunately, standard resistivity value ranges for various soil/rock types were available, which were incorporated to the extracted apparent resistivity dataset. The result shows that the top three performing models were linear regression, random forest, and linear regression with polynomial features with 24.0826, 24.4457, and 20.7549 values of MSE, respectively. With the highest R-squared value and lowest RMSE value, linear regression with polynomial features is the best performing model that could be ideally used for the intended application. Hence, the primary benefit of this work will be the establishment of subsurface decision making for geological interpretation in a rapid and continuous manner without major cost failures.