A combined data mining approach for on-line prediction of key soil quality indicators by Vis-NIR spectroscopy

Successful modelling of visible and near-infrared (vis-NIR) spectra for on-line prediction of key soil quality indicators is crucial for accurate variable rate applications of farm input resources. The aim of this paper is to optimize modelling of on-line collected spectra for the prediction of soil pH, organic carbon (OC), extractable phosphorous (P) and potassium (K) by means of spiking, combined with clustering and/or extra-weighting. A mobile fiber-type vis-NIR spectrophotometer (CompactSpec from Tec5 Technology, Germany), with spectral range of 305 1700 nm was calibrated using 100 samples collected from five different fields, which were merged with 28 samples collected from a target field. The resulting dataset was subjected to spectral pretreatments followed by k-means clustering and 95 % confidence ellipsoid, resulting in three optimal datasets. Partial least squares regression (PLSR) analyses were carried out on the calibration set (75 % of samples) for four calibration strategies: (i) non-clustered and non-weighted (NCNW), (ii) clustered and non-weighted (CNW), (iii) nonclustered but extra-weighted (NCW), and (iv) clustered and extra-weighted (CW). Results showed that the quality of on-line predictions was the best after clustering combined with extra-weighting. Modelling based with CW significantly improved model prediction accuracy to be very good for pH (ratio of prediction deviation (RPD) = 2.32) and P (RPD = 2.05), and good for OC (RPD = 1.90) and K (RPD = 1.80), whereas results of NCNW (standard calibration approach) were the poorest to be fair for P (RPD = 1.74) and OC (RPD = 1.50), and poor for K (RPD = 1.1) and pH (RPD = 1.39). It can be concluded that optimal sample selection with k-mean clustering when combined with extra-weighting will result in accurate PLSR calibration models for on-line prediction of soil pH, P, OC and K using a multi-field diverse dataset.

Automated summary

Successful modelling of vis-NIR spectra is crucial for accurate variable rate applications of farm input resources. This study optimized on-line collected spectra for the prediction of soil quality indicators using spiking, clustering, and extra-weighting methods. Results showed that combining clustering with extra-weighting improved prediction accuracy for soil pH, organic carbon, extractable phosphorous, and potassium.