Open Soil Spectral Library (OSSL): Building reproducible soil calibration models through open development and...

Area of Science:

• Soil Science
• Spectroscopy
• Machine Learning

Background:

• Soil spectroscopy is vital for environmental and agricultural monitoring.
• Large reference datasets (soil spectral libraries, SSLs) are needed for machine learning (ML) models.
• Limited data diversity hinders prediction accuracy for new soil samples.

Purpose of the Study:

• To establish the Open Soil Spectral Library (OSSL) as a centralized, open-access resource.
• To describe data collection, harmonization, and analysis procedures for the OSSL.
• To evaluate ML model performance for predicting soil properties using spectral data.

Main Methods:

• Collected and harmonized multiple SSLs into the OSSL.
• Performed exploratory data analysis and predictive modeling.
• Utilized 10-fold cross-validation and independent model evaluation.

Main Results:

• Mid-infrared (MIR) models showed significantly higher accuracy than visible and near-infrared (VisNIR) or near-infrared (NIR) models.
• The Cubist ML algorithm demonstrated the best performance for calibration and reliable output delivery.
• Prediction accuracy varied, with total sulfur, extractable sodium, and electrical conductivity performing poorly across all spectral regions.

Conclusions:

• The OSSL facilitates affordable soil data collection and enhances ML model development.
• MIR spectroscopy offers superior accuracy for predicting many soil properties compared to VisNIR or NIR.
• Limitations exist for spectral prediction of certain soil properties, highlighting the need for integrated approaches.