A Hoeffding Decision Tree Based Approach for Soil Classification

Abstract

Soil classification is an important but challenging problem for the research community. As such, current solutions for classifying soil for a wide variety of reasons are out of the reach of hobbyists and the small research firms. This research study focuses on comparing various machine learning algorithms on a custom database generated from Canadian System for Soil Classification (CSSC) attributes to reveal a solution for identifying a soil Pedon. Discussion centres around acquainting the user with soil terminology, current solutions to the problem of soil classification, and the proposed solution. A database using these attributes was constructed, and six algorithms were analyzed using validation, test case, and 70-30 split testing via WEKA. Among the comparing algorithms, the Hoeffding decision tree was found to perform best, and it was subsequently used in developing a simple prototype using Java Graphical User Interface (GUI). Finally, the Hoeffding decision tree was compared to the other algorithms that were used to see why it was more accurate than its competitors.