Comparing predictive performance of near infrared spectroscopy at a field, regional, national and continental scales by using spiking and data mining techniques

The development of accurate visible and near infrared (vis-NIR) spectroscopy calibration models for selected soil properties is a crucial step for variable rate application in precision agriculture. The objective of the present study was to compare the prediction performance of vis-NIR spectroscopy at local, regional, national and continental scales using data mining techniques including spiking. Fresh soil samples collected from farms in the UK, Czech Republic, Germany, Denmark and the Netherlands 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 (75%) and validation (25%) sets, spectra in the calibration set were subjected to three multivariate calibration models. The partial least squares regression (PLSR), multivariate adaptive regression splines (MARS) and support vector machines (SVM), with leave-one-out cross-validation were used to establish calibration models of total nitrogen (TN), total carbon (TC) and soil moisture content (MC). The results showed the lowest model performance to be obtained when the single field (local scale) data were used in the calibration models. The effect of spiking was significant and the best model performance was obtained when local samples collected from two fields in the UK were spiked with European soil samples (continental), followed by when the same samples were spiked with UK samples (national). Therefore, these results suggest that continental and national vis-NIR calibration models can be successfully used to predict TN, TC and MC. Therefore, selection of the optimal soil samples with the appropriate data mining technique should be considered when developing vis-NIR calibration models for a non-standard soil to cover a wide variation range