On-line vis-NIR spectroscopy prediction of soil organic carbon using machine learning

Abstract

Accurate on-line visible and near infrared (vis-NIR) spectroscopy prediction of soil organic carbon (OC) is essential for food security and environmental management. This paper aims at using on-line vis-NIR spectra coupled with random forest (RF) modelling approach for the prediction of soil organic carbon (OC), comparing between single field (SF), non-spiked UK multiple-field (NSUK) and spiked UK multiple-field (SUK) calibration models. Fresh soil samples collected from 6 fields in the UK (including two target fields) were scanned with a fibre-type vis-NIR spectrophotometer (tec5 Technology for Spectroscopy, Germany), with a spectral range of 305–2200 nm. After dividing spectra into calibration and independent validation sets, RF was run on the calibration set to develop calibration models for OC for the three studied datasets. Results showed that the model prediction performance depends on the dataset used and varies between fields. Less accurate prediction performance was obtained for the on-line prediction compared to laboratory (samples scanned in the laboratory under non-mobile measurement) prediction, and for non-spiked models compared to spiked models. The best model performance in both laboratory and on-line predictions was obtained when samples from the SF were spiked into the UK samples, with coefficients of determination (R2) values of 0.80 to 0.84 and 0.74 to 0.75, root mean square error of prediction (RMSEP) values of 0.14% and 0.17 to 0.18%, and ratio of prediction deviation (RPD) values of 2.30 to 2.5 and 1.98 to 2.04, respectively. Therefore, these results suggest that RF modelling approach when coupled with spiking provides high prediction performance of OC under both non-mobile laboratory and on-line field scanning conditions.