National calibration of soil organic carbon concentration using diffuse infrared reflectance spectroscopy

Abstract

This study presents the potential of infrared diffuse reflectance spectroscopy (DRS) to predict soil organic carbon (SOC) content. A large national soil library was used, including about 3800 samples collected at two soil depths (0–30 and 30–50cm) using a 16×16km plot grid over the French metropolitan territory (552,000km2). Reflectance spectra were collected in the laboratory using visible and near infrared (VNIR), near infrared (NIR) and mid infrared (MIR) spectrophotometers. The soil data library was broken down into calibration and validation sets through sample selection at random or based on spectral representativeness. The calibration intensity was investigated in order to assess the optimum number of calibration samples required to obtain accurate models. Predictions were achieved using global or local partial least square regression (PLSR) built using VNIR, NIR and MIR spectra separately or in combination. Local PLSR uses only calibration samples that are spectral neighbors of each validation sample, thus builds one model per validation sample. Model performance was evaluated on the validation set based on the standard error of prediction (SEP), the ratio of performance to deviation (RPDv), and the ratio of performance to interquartile range (RPIQv).Using all calibration samples, the global PLSR model provided the most precise predictions of SOC content with the MIR spectra, then with the NIR spectra, and less accurate predictions with the VNIR spectra (SEP=2.6, 4.4 and 4.8gkg−1, RPDv=2.7, 2.3 and 1.5, and RPIQv=3.3, 2.2 and 1.9, respectively). The combination of spectral libraries did not improve model performance noticeably. Local PLSR provided better models than global PLSR, allowing accurate predictions with only 30% of the calibration set, whatever the spectral library (RPDv and RPIQv>2.0). Optimum calibration intensity was estimated at about 60% for MIR spectra with both global and local PLSR, 30–40% for VNIR and NIR spectra with global PLSR, but 50% for VNIR spectra and 70% for NIR spectra with local PLSR. The most accurate models, which were obtained using the MIR spectra and local PLSR with calibration intensity higher than 50%, allowed very good SOC determination for the most frequent French soils (SEP<2gkg−1). This highlights the potential of infrared DRS for national SOC monitoring, provided that calibration database is strengthened with samples from less frequent soil types.