Machine Learning-Based Classification of Soil Parent Materials Using Elemental Concentration and Vis-NIR Data.

Area of Science:

• Soil Science
• Geochemistry
• Remote Sensing

Background:

• Accurate soil sample origin allocation is vital in soil science for effective utilization.
• Advancements in proximal sensing and machine learning offer new approaches for soil analysis.

Purpose of the Study:

• To investigate soil classification based on four parent materials using machine learning algorithms.
• To evaluate the effectiveness of different analytical techniques (XRF, ICP-OES, Vis-NIR) and machine learning models for soil classification.

Main Methods:

• Collected 59 soil samples from 12 profiles and surrounding areas (0-30 cm depth).
• Analyzed samples for elemental concentrations (XRF, ICP-OES) and spectral data (Vis-NIR).
• Applied machine learning algorithms: Support Vector Machine (SVM), Ensemble Subspace k-Near Neighbor (ESKNN), and Ensemble Bagged Trees (EBTs) for classification.
• Validated models using five-fold cross-validation and an 80% calibration/20% validation split.

Main Results:

• Classification success rates ranged from 70% to 100%, depending on the dataset and algorithm.
• Ensemble Subspace k-Nearest Neighbor (ESKNN) achieved the highest accuracy at 99%.
• Relief algorithms identified key variables: CaO, Fe2O3, Al2O3, MgO, MnO for ICP-OES; SiO2, CaO, Fe2O3, Al2O, MnO for XRF; and specific wavelengths (567-574 nm) for Vis-NIR.

Conclusions:

• Machine learning, particularly ESKNN, is highly effective for classifying soil parent materials.
• Elemental and spectral data combined with machine learning provide robust soil classification capabilities.
• Identification of key variables can potentially streamline soil analysis while maintaining high classification accuracy.