Determining soil particle-size distribution from infrared spectra using machine learning predictions: Methodology and...

Area of Science:

• Soil Science
• Analytical Chemistry
• Machine Learning

Background:

• Accurate soil particle-size distribution (PSD) measurement is crucial for soil management.
• Traditional methods like sedimentation and laser diffraction can be time-consuming and resource-intensive.
• Infrared (IR) spectroscopy offers a rapid, non-destructive approach for soil analysis.

Purpose of the Study:

• To evaluate the accuracy of infrared (IR) models for predicting soil particle-size distribution (PSD) across diverse methods and soil properties.
• To explore the utility of machine learning, specifically Stochastic Gradient Boosting (SGB), in enhancing IR-based PSD predictions.
• To identify key soil features and spectral parameters that improve the accuracy of near-infrared (NIR) and mid-infrared (MIR) spectroscopy for PSD analysis.

Main Methods:

• Near-infrared (NIR) and mid-infrared (MIR) spectra were collected from 1298 soil samples from eastern Canada.
• Stochastic Gradient Boosting (SGB) machine learning was employed to predict sand, silt, clay, and carbon content.
• Log-ratio (ilr) transformation was applied to compositional soil data to mitigate numerical biases.
• Performance was compared between IR spectra, log-log relationships of settling time and suspension density function (SDF), and various spectral processing techniques (e.g., gain, method).

Main Results:

• SGB models using NIR spectra achieved high prediction accuracies (R2 = 0.76-0.94) for PSD, comparable to traditional SDF methods (R2 = 0.84-0.92).
• The inclusion of soil properties like carbon, bulk density, pH, and color significantly improved texture prediction accuracy (R2 = 0.86-0.91).
• Mid-infrared (MIR) spectra generally outperformed near-infrared (NIR) spectra for PSD prediction.
• Specific parameters like the slope and intercept of SDF, 4X gain, and soil carbon were identified as crucial for accurate SGB-processed NIR analysis.

Conclusions:

• Machine learning significantly enhances the capability of IR spectroscopy for cost-effective and accurate soil particle-size distribution analysis.
• Incorporating additional soil properties alongside spectral data provides a more robust prediction of soil texture.
• The study demonstrates the potential of routine IR analysis, supported by machine learning, for soil characterization.