Comparison between artificial neural network and partial least squares for on-line visible and near infrared spectroscopy measurement of soil organic carbon, pH and clay content

Abstract Soil organic carbon (OC), pH and clay content (CC) can be measured with on-line visible and near infrared spectroscopy (vis-NIRS), whose calibration method may considerably affect the measurement accuracy. The aim of this study was to compare artificial neural network (ANN) and partial least squares regression (PLSR) for the calibration of a visible and near infrared (vis-NIR) spectrophotometer for the on-line measurement of OC, pH and CC in two fields in a Danish farm. An on-line sensor platform equipped with a mobile, fiber type, vis-NIR spectrophotometer (AgroSpec from tec5 Technology for Spectroscopy, Germany), with a measurement range of 305–2200 nm was used to acquire soil spectra in diffuse reflectance mode. Both ANN and PLSR calibration models of OC, pH and CC were validated with independent validation sets. Comparison and full-point maps were developed using ArcGIS software (ESRI, USA). Results of the on-line independent validation showed that ANN outperformed PLSR in both fields. For example, residual prediction deviation (RPD) values for on-line independent validation in Field 1 were improved from 1.93 to 2.28, for OC, from 2.08 to 2.31 for pH and from 1.98 to 2.15 for CC, after ANN analyses as compared to PLSR, whereas root mean square error (RMSEP) values decreased from 1.48 to 1.25%, for OC, from 0.13 to 0.12 for pH and from 1.05 to 0.96% for CC. The comparison maps showed better spatial similarities between laboratory and ANN predicted maps (higher kappa values), as compared to PLSR predicted maps. In most cases, more detailed full-point maps were developed with ANN, although the size of spots with high concentration of PLSR maps matches the measured maps better. Therefore, it was recommended to adopt the ANN for on-line prediction of OC, pH and CC.