Predicting Soil Properties and Interpreting Vis-NIR Models from across Continental United States

Area of Science:

• Soil Science
• Geospatial Analysis
• Chemometrics

Background:

• The U.S. NRCS soil database contains extensive historical data, including soil spectral scans.
• This valuable data is underutilized for comprehensive soil property analysis.
• Soil spectral data offers a powerful, non-invasive method for characterizing soil properties.

Purpose of the Study:

• To extract and analyze soil pedon, horizon, and spectral data from the NRCS database (2011-2015).
• To model and predict eight key soil properties using chemometric methods.
• To evaluate the performance of four different prediction models: PLSR, RF, Cubist, and MARS.

Main Methods:

• Extracted over 14,000 soil samples with complete data from the NRCS database.
• Utilized random subsetting for calibration (70%) and validation (30%) datasets.
• Employed four chemometric methods (PLSR, RF, Cubist, MARS) for predicting soil organic carbon, total nitrogen, total sulfur, clay, sand, exchangeable calcium, CEC, and pH.

Main Results:

• High prediction accuracy (R² > 0.9) was achieved for soil organic carbon (SOC), total nitrogen (TN), and total sulfur (TS).
• Moderate prediction performance was observed for exchangeable calcium (Ca_ex), cation exchange capacity (CEC), and pH.
• The Cubist model demonstrated the strongest predictive performance, while PLSR showed the weakest results.

Conclusions:

• Soil spectral data, when analyzed with appropriate chemometric models like Cubist, can accurately predict key soil properties.
• The study highlights the potential for leveraging large soil databases for enhanced soil characterization.
• Future research should explore incorporating environmental variables and consider alternative models for complex datasets.