Strategies to improve the prediction of bulk soil and fraction organic carbon in Brazilian samples by using an Australian national mid-infrared spectral library

• Published in: Geoderma Vol. 373; p. 114401
• Main Authors: Briedis, Clever, Baldock, Jeff, de Moraes Sá, João Carlos, dos Santos, Josiane Burkner, Milori, Débora Marcondes Bastos Pereira
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 15.08.2020
• Subjects: Chemometrics; Cubist; MIR spectroscopy; Partial least square regression; Spectrum-based learner; Spiking; MIR spectroscopy; Cubist; Spectrum-based learner; Chemometrics; Spiking; Partial least square regression
• ISSN: 0016-7061; 1872-6259
• DOI: 10.1016/j.geoderma.2020.114401

•MIR spectroscopy with different chemometric approaches was applied to predict OC in bulk soil and fractions.•An Australian national library and a Brazilian regional library were used to predict Brazilian target samples.•Calibrating a PLSR from Brazilian samples only returned the greatest OC prediction accuracy, but it was costly.•Calibrating models using exclusively the Australian library returned unsure OC predictions.•Spiking few Brazilian regional samples to the Australian MIR library yield suitable OC prediction with low laboratory costs. At present, organic carbon (OC) levels are measured using techniques that can be very expensive and time-consuming or can generate toxic residue. Mid-infrared (MIR) spectroscopy, which is an inexpensive and safe technique, can be used for OC prediction; however, appropriate modeling is required. This study aimed to evaluate different strategies for MIR modeling calibration in order to improve the accuracy and cost-effectiveness of OC prediction in bulk soils and fractions (e.g., particulate OC, POC; humic OC, HOC; resistant OC, ROC). We used regional topsoil samples (i.e. 0–20 cm) from Brazil (BRreL) and an existing Australian national library (AUnaL) to calibrate different models, which were used to predict OC content in target samples (i.e. subset of the Brazilian sample set). In total, eight strategies were tested, including the use of different soil libraries, different chemometric approaches (i.e. common multivariate regression and local-type modeling), and spiking. The highest accuracy for OC prediction in Brazilian regional target samples for both bulk soils (RPIQ = 5.86; RMSE = 3.17 g kg−1) and fractions (RPIQ = 1.58, 5.72, and 5.21 for POC, HOC, and ROC, respectively) was obtained when the BRreL was used to calibrate a partial least square regression. When AUnaL was used alone as the calibration set, the OC prediction accuracy sharply decreased and even the local-type models Cubist and spectrum-based learning, in general, were not able to achieve high OC predictions accuracy. With this strategy, the averaged RPIQ values (among the three models) were 2.92, 0.55, 2.17, and 0.65 for bulk OC, POC, HOC, and ROC, respectively. However, when 20 Brazilian regional samples (representing 8 and 42% of the original calibration samples used for total and fraction OC, respectively) were spiked to the AUnaL, the OC predictions accuracy markedly improved for all OC pools, regardless of the modeling approach used. With this strategy, the averaged RPIQ values were 4.74, 1.35, 4.49, and 2.29 for bulk OC, POC, HOC, and ROC, respectively. In addition, high OC prediction accuracy (RPIQ = 4.49 for bulk OC; RPIQ = 1.58, 6.87, and 4.26 for POC, HOC, and ROC, respectively) was obtained when models were calibrated with the 20-spiking subset only. Our results suggest that a proper selection of a small calibration set, that could be used alone or to augment an existing library, is critical and could make feasible the OC prediction in bulk soils and fractions by MIR spectroscopy.