Building a near-infrared (NIR) soil spectral dataset and predictive machine learning models using a handheld NIR...

Area of Science:

• Soil Science
• Spectroscopy
• Machine Learning

Background:

• A large dataset of near-infrared (NIR) spectra from diverse mineral soils was compiled.
• Soil samples were sourced from the US and Africa, representing varied soil types.
• Spectra were collected using handheld spectrophotometers on dried and sieved soil samples (<2 mm).

Purpose of the Study:

• To develop machine learning models for predicting key soil properties from NIR spectra.
• To provide a valuable resource for soil monitoring and management applications.
• To enable broader use of spectral data in soil science.

Main Methods:

• Collected NIR spectral data from 2,106 mineral soil samples.
• Developed predictive models using Cubist and Partial Least Squares Regression (PLSR) algorithms in R.
• Utilized two modeling strategies: averaging spectral scans and using replicate scans across devices.

Main Results:

• Machine learning models were created to predict soil organic carbon (SOC), pH, bulk density (BD), carbonate (CaCO3), exchangeable potassium (Ex. K), sand, silt, and clay content.
• Internal performance of the developed Cubist and PLSR models was evaluated.
• The dry spectra and Cubist models are publicly available for download.

Conclusions:

• The developed spectral dataset and predictive models offer a powerful tool for soil property assessment.
• This resource facilitates advancements in soil monitoring and precision agriculture.
• Open access to data and code promotes collaborative research in soil spectroscopy.