Effect of spiking strategy and ratio on calibration of on-line visible and near infrared soil sensor for measurement in European farms

A previously developed on-line visible and near infrared (vis–NIR) spectroscopy-based soil measurement system was implemented for the measurement of soil organic carbon (OC), total nitrogen (TN) and moisture content (MC) in three fields at three European farms. The on-line sensor platform was coupled with a mobile, fibre type, vis–NIR spectrophotometer (AgroSpec from tec5 Technology for Spectroscopy, Germany), with a measurement range of 305–2200 nm, to acquire soil spectra in diffuse reflectance mode. A general calibration set of 425 soil samples, spiked with different number of spectra from the three validation fields were used to establish calibration models for the studied soil properties using partial least squares (PLS) regression analysis. Different spiking strategies and spiking ratios were investigated and results revealed that the best prediction accuracy was obtained after 20% spiking ratio with samples whose spectra were measured in the laboratory. Evaluated by the values of residual prediction deviation (RPD), which is the ratio of standard deviation to root mean square error of prediction (RMSEP), the accuracy of the on-line measurement was classified as excellent for MC (RPD = 2.76–3.96), good to very good for OC (RPD = 1.88–2.38) and good to excellent for TN (RPD = 1.96–2.52). Reducing the number of samples used for spiking resulted in deteriorating the prediction accuracy, although 1–2 samples per ha were found to provide good predictions. There was a distinguishable spatial similarity between the on-line and laboratory measured maps for all studied properties, although the full-data point maps provided more detailed information about the spatial variation. This confirms that the on-line vis–NIR soil sensor provides correct and detailed information about soil OC, TN and MC at high sampling resolutions.

► The accuracy of on-line measurement of OC, TN and MC is affected by spiking strategy.
► Reducing number of samples used for spiking results in deteriorating prediction accuracy.
► About 1–2 samples per ha are recommended for spiking to establish models at farm scale.