[Using different data mining algorithms to predict soil organic matter based on visible-near infrared spectroscopy].

Area of Science:

• Soil Science
• Spectroscopy
• Machine Learning

Context:

• Visible/near infrared (Vis-NIR) spectroscopy is crucial for rapid soil property assessment and precision agriculture.
• Accurate soil organic matter (SOM) prediction is vital for sustainable land management.

Purpose:

• To evaluate and compare the predictive performance of nonlinear models (Random Forests, Support Vector Machines, Artificial Neural Networks) and a linear model (Partial Least Squares Regression) for soil organic matter estimation using Vis-NIR spectroscopy.
• To investigate the impact of different calibration and validation dataset divisions on model prediction accuracy.

Summary:

• Nonlinear models, particularly Support Vector Machines (SVM) utilizing all Vis-NIR wavelengths, showed high predictive ability for soil organic matter, outperforming the linear Partial Least Squares Regression (PLSR) model.
• The combined PLSR-Artificial Neural Network (ANN) approach significantly enhanced PLSR's predictive capability, offering good predictions with improved interpretability despite ANN's 'black box' nature.

Impact:

• Demonstrates the effectiveness of nonlinear machine learning models, especially SVM, for accurate soil organic matter prediction using Vis-NIR spectroscopy.
• Highlights the potential of hybrid modeling approaches (e.g., PLSR-ANN) to improve predictive performance and interpretability in soil sensing research.
• Provides valuable insights for developing advanced soil sensing technologies for precision agriculture and environmental monitoring.