Search Results for ensemble-methods

Abstract

Soil classification is a significant tool for the interpretation of soil properties and the selection of the appropriate parameters to express soil types in geotechnical engineering applications. In this paper, a comparative analysis on the prediction of soil classification using different soil parameters via Neural Network, Decision Tree, and Ensemble classifier algorithms in the machine learning tool box of MATLAB is conducted. The data is classified into 14 sets of input parameters according to the parameters of the Unified Soil Classification System (USCS), which is a standard system in practice for classifying soils based on their physical properties and their engineering behavior. The input parameters of the USCS, Passing No. 4 Sieve, Passing No. 200 Sieve, Gravel, Sand, Fines, Liquid Limit, Plastic Limit and Plasticity Index, are specified within the certain ranges to generate a synthetic data base including 478 data for the purpose of training and testing. Five Neural Network models (Medium, Trilayered, Bilayered, Narrow and Wide), three Decision Tree models (Fine, Medium, Coarse) and three Ensemble models (Boosted Trees, Bagged Trees, RUSBoosted Trees) are tested for the different test split ratios (10, 20, and 30). The results show that the Neural Network models proposed in this study have the highest prediction capability, with a classification accuracy of up to 100 % for some variable sets, and can classify soils automatically and precisely based on the soil index properties. Among the individual models, WNN and NNN generally showed the best performance for all variable sets. The Ensemble models presented high prediction capability with respect to generalization. Among the DT models, although they are fast, they presented relatively lower accuracy than the other models, especially the Coarse Tree model. This study can be used as a reference for the future development of automated soil classification systems and can offer practical suggestions for the selection of machine learning models for geotechnical engineering problems.