Optimal sample selection for measurement of soil organic carbon using on-line vis-NIR spectroscopy

Abstract

The selection of samples for modelling of visible and near infrared (vis-NIR) spectra for prediction of soil organic carbon (OC) is a crucial step for improving model prediction performance. This paper aims at comparing three soil sample selection methods coupled with spiking technique for improving on-line prediction performance of OC. Sample selection methods included random selection (RS), Kennard-Stone (KS) algorithm and similarity analysis (SA). Soil vis-NIR spectra was measured with an on-line fibre-type vis-NIR spectrophotometer (tec5 Technology for Spectroscopy, Germany), with a spectral range of 305–2200 nm. A multiple field sample set (268 samples) was merged with samples (148 samples) collected from one target field, and the resulted sample set was subjected to the three sample selection methods. After dividing spectra into calibration and prediction sets, partial least squares regression (PLSR) was run on the calibration set to develop calibration models for OC, and resulted models were validated using samples of the prediction set. Results show that SA performed generally better than its competitors, especially when there were 58 spiked samples used in the calibration set (54% of total spiked samples of 106), with the best residual prediction deviation (RPD) and root mean squares error of prediction (RMSEP) of 2.14–2.54 and 0.16–0.15% for laboratory and on-line prediction. KS and RS performed similarly, but depending on the size of the calibration set, KS produced slightly better models. This indicates that the proposed SA coupled with spiking holds great potential in the optimization of a calibration set size and may serve as a novel and efficient tool for balancing the cost and quality of vis–NIR calibrations for estimating OC.