Skip to main content
ARS Home » Midwest Area » Columbia, Missouri » Cropping Systems and Water Quality Research » Research » Publications at this Location » Publication #358096
Title: Predicting profile soil properties with reflectance spectra via Bayesian covariate-assisted external parameter orthogonalization

Author
item Veum, Kristen
item PARKER, PAUL - University Of Missouri
item Sudduth, Kenneth - Ken
item HOLAN, SCOTT - University Of Missouri

Submitted to: Sensors
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/9/2018
Publication Date: 11/10/2018
Citation: Veum, K.S., Parker, P., Sudduth, K.A., Holan, S.H. 2018. Predicting profile soil properties with reflectance spectra via Bayesian covariate-assisted external parameter orthogonalization. Sensors. 18(11):3869. https://doi.org/10.3390/s18113869.
DOI: https://doi.org/10.3390/s18113869

Interpretive Summary: Profile diffuse reflectance spectroscopy (DRS) soil sensors have the potential to provide rapid, high-resolution measurement of soil profiles for precision agriculture, soil health assessment, and other applications related to environmental protection and agronomic sustainability. However, the effects of soil moisture, other environmental factors, and the in-field DRS data collection process often hamper the utility of in-situ spectral data. Various processing and modelling techniques have been developed to overcome these challenges. The objectives of this study were to estimate soil organic carbon (SOC), total nitrogen (TN), and texture fractions using a large, regional dataset of profile DRS spectra and compare the performance of traditional modelling approaches with newer techniques. In this study, soil cores and profile DRS spectrometer scans were obtained to ~1 m depth from 22 fields across Missouri and Indiana, USA, totaling 708 samples. Estimates of all soil properties were dramatically improved by the newer spectral modelling techniques. Overall, the study illustrates the potential to overcome the effects of in-field data collection to advance in-field DRS spectroscopy as a tool for rapid, high-resolution estimation of soil properties. The results of this study benefit scientists and producers by improving precision agriculture tools for more informed agroecosystem management.

Technical Abstract: In-situ, diffuse reflectance spectroscopy (DRS) profile soil sensors have the potential to provide rapid, high-resolution prediction of multiple soil properties for precision agriculture, soil health assessment, and other applications related to environmental protection and agronomic sustainability. However, the effects of soil moisture, other environmental factors, and artefacts of the in-field spectral data collection process often hamper the utility of in-situ DRS data. Various processing and modelling techniques have been developed to overcome these challenges, including external parameter orthogonalization (EPO) transformation of the spectra. In addition, Bayesian modelling approaches may improve prediction over traditional partial least squares (PLS) regression. The objectives of this study were to predict soil organic carbon (SOC), total nitrogen (TN), and texture fractions using a large, regional dataset of in-situ, profile DRS spectra and compare the performance of 1) traditional PLS analysis, 2) PLS on EPO transformed spectra (PLS-EPO), 3) PLS-EPO with the Bayesian Lasso (PLS-EPO-BL), and 4) covariate assisted PLS-EPO-BL models. In this study, soil cores and in-situ profile DRS spectrometer scans were obtained to ~1 m depth from 22 fields across Missouri and Indiana, USA. In the laboratory, soil cores were split by horizon, air-dried, and sieved (< 2mm) for a total of 708 samples. Soil properties were measured and air-dry DRS spectra were collected in the laboratory. The data were split into training (n=308), testing (n=200), and EPO calibration (n=200) sets, and soil textural class was used as the categorical covariate in the Bayesian models. Model performance was evaluated using root mean square error of prediction (RMSEP). For the prediction of soil properties using a model trained on dry spectra and tested on field-moist spectra, the PLS-EPO transformation dramatically improved model performance relative to PLS alone, reducing RMSEP by 66% and 53%, for SOC and TN, respectively, and by 76%, 91%, and 87% for clay, silt, and sand, respectively. The addition of the Bayesian Lasso further reduced RMSEP by 4% – 11% across soil properties, and the categorical covariate reduced RMSEP by another 2% - 9%. Overall, the study illustrates the strength of the combination of EPO spectral transformation paired with Bayesian modelling techniques to overcome environmental factors and in-field data collection artefacts when using in-situ DRS data, and highlights the potential for in-field DRS spectroscopy as a tool for rapid, high-resolution prediction of soil properties.