Machine Learning-Based Organic Soil Classification Using Cone Penetrometer Tests

Hanssel Omar Ulloa

doi:10.1061/JGGEFK.GTENG-12322

Machine Learning-Based Organic Soil Classification Using Cone Penetrometer Tests

2024-09-01

0

PRE

AI

Sabarethinam Kameshwar

I

Ignacio Harrouch

DOI：10.1061/JGGEFK.GTENG-12322

Original

Original request for help

Save

Abstract

En 中文

Soil classification methods currently rely on soil borings or cone/piezocone penetrometer tests (CPT/CPTu). Literature provides several methods that classify soils based on two parameters (typically tip resistance and friction sleeve/porewater pressure) obtained from CPTu data, defining a soil behavior type. However, these methods face challenges in reliably classifying certain soils, such as organic soils. Robust and complex analyses are required to classify organic soils accurately. In this study, a random forest (RF) based method is developed to predict the presence and depth of organic soils. The RF model utilized features from CPTu soundings including tip resistance, sleeve friction, and pore pressure. The true location of organic soils, derived from index properties obtained from soil borings, served as the model's outputs. Unseen CPTu data was used to predict organic soils throughout the entire project alignment, serving as validation for the model. The model achieved an F1 score of 0.89. Settlement analyses were conducted to evaluate the practical implications of the model's predictions on levee fill costs. The RF-based model yielded settlement predictions that closely matched those obtained through boring predictions. In contrast, traditional classification methods underestimated settlements, resulting in lower estimates for levee fill costs.

Keywords:

Organic soils

Cone penetrometer test (CPT)

Machine learning

Soil classification

Random forest (RF)

Machine Learning-Based Organic Soil Classification Using Cone Penetrometer Tests

Abstract

Journal

Organization