Date Available

11-19-2025

Year of Publication

2025

Document Type

Master's Thesis

Degree Name

Master of Science (MS)

College

Agriculture

Department/School/Program

Plant and Soil Sciences

Faculty

Dr. Christopher Shepard

Abstract

Soil health is critical to sustaining life on Earth, influencing plant growth, agricultural productivity, water quality, and overall ecosystem sustainability. In response to climate change and intensifying land use, the need to improve our understanding of soil health has become increasingly urgent. The application of mid-infrared (MIR) spectroscopy coupled with statistical modeling has emerged as a promising, cost-effective approach to predict soil health parameters, offering a quicker alternative to traditional wet chemistry analyses. MIR methods have been successfully applied to predict various soil health properties, such as bulk density (BD), cation exchange capacity (CEC), base saturation (BS), electrical conductivity (EC), soil organic carbon, and total nitrogen, although with varying degrees of success.

However, there remains a lack of standardization in the application of MIR for soil health assessment, particularly regarding sample size, statistical modeling, and the handling of spectral data. In this study, we conducted a meta-analysis comparing different statistical approaches, including machine learning and partial least squares regression (PLSR), to assess their effectiveness in predicting soil health metrics using R² and root mean square error (RMSE) values. We found that machine learning methods generally outperformed PLSR in both R² and RMSE, but sample sizes greater than 500 observations did not lead to further improvements in predictions. Furthermore, there was considerable variability in training data properties, number of components used in models, and spectral treatments across studies.

To advance MIR-based soil health assessments, particularly for under-studied soil types such as those with fragipans, we applied MIR spectroscopy to 90 soil core samples from four different soil series in Western Kentucky. These samples were taken from long-term no-till and pasture agroecosystems, with the goal of refining standardization and model comparison in soil health research. Our findings highlight the importance of reporting consistent model parameters and evaluation metrics in future studies. Improved standardization could enable better inter-model and site comparisons, advancing our ability to predict soil health and understand soil-forming processes in diverse ecosystems.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2025.511

Funding Information

Support was provided by the NRCS Natural Resources Conservation Service grant #NR223A750023C009 from 2023-2025

Download

Included in

Soil Science Commons

Share

COinS