Evaluation of Optimized Preprocessing and Modeling Algorithms for Prediction of Soil Properties Using VIS-NIR...

Area of Science:

• Soil Science
• Spectroscopy
• Data Analysis

Background:

• Visible and near-infrared (VIS-NIR) spectroscopy offers a rapid, non-destructive method for soil analysis.
• Optimizing preprocessing and modeling techniques is crucial for accurate soil property prediction.

Purpose of the Study:

• To evaluate various preprocessing and modeling techniques for predicting soil properties using VIS-NIR spectroscopy.
• To identify the optimal combination of algorithms for predicting soil organic matter (SOM), pH, electrical conductivity (EC), and particle size fractions.

Main Methods:

• Collected VIS-NIR absorbance spectra (343-2200 nm) from Ontario soil samples.
• Tested thirteen preprocessing algorithms (e.g., 1st derivative, SNV) and four modeling approaches (PLSR, Cubist, RF, ELM).
• Focused on the best-performing preprocessing methods (1st derivative + gap, 2nd derivative + gap, SNV) for predicting multiple soil properties.

Main Results:

• Soil organic matter (OM), pH, %sand, %silt, and %coarse sand (CS) were predicted with high confidence (R² > 0.60).
• The combination of 1st derivative + gap preprocessing and the random forest (RF) model achieved the best prediction performance.
• Different soil properties required distinct preprocessing and modeling algorithms for optimal prediction accuracy.

Conclusions:

• VIS-NIR spectroscopy, coupled with appropriate data processing, is a reliable tool for soil property assessment.
• The random forest (RF) model and 1st derivative + gap preprocessing represent a highly effective combination for predicting various soil characteristics.
• Tailoring algorithm selection to specific soil properties enhances predictive power and analytical efficiency.