Dr Clever Briedis1, Dr Jeff Baldock2, Dr João Carlos de Moraes Sá3, Dr Débora Marcondes Bastos Pereira Milori1
1Brazilian Agricultural Research Corporation, São Carlos, Brazil, 2CSIRO Agriculture and Food, Glen Osmond, Australia, 3State University of Ponta Grossa, Ponta Grossa, Brazil
Currently, the measurement of organic carbon concentration in soil is completed using techniques that can be expensive, slow or generate toxic residue. Mid-infrared spectroscopy (MIR), when combined with chemometric analyses, can provide a fast and clean approach but proper model calibration is required for achieving reliable predictions of OC concentration. This study aimed to evaluate different strategies to calibrate the application of MIR to predicting OC concentration in the soil in order to improve the accuracy and cost-effectiveness values predicted for bulk soils and fractions of target samples. For this purpose, we used regional soils from Brazil (target samples – BRreL) and an existing Australian national library (AUnaL). In total, nine different model calibration strategies were tested for OC prediction in the target samples. Partial least square regression (PLSR) using only BRreL for calibration provided the highest accuracy for OC prediction in bulk soils and fractions. When only the AUnaL was used for PLSR calibration, the accuracy decreased, and OC predictions were acceptable for bulk soils but not for soil fractions. Alternative algorithms (e.g., cubist and spectrum-based learning – SBL) applied on the AUnaL, in general, did not improve OC predictions. The most promising results were found when the calibration of the model was performed by adding 20 BRreL samples in with the AUnaL samples. Through this spiking technique, regardless of the algorithm used (e.g., PLSR, cubist or SBL), the OC predictions were improved, making it very accurate for both bulk soils and fractions. In addition, this technique was more cost-effective since only a small number of samples from the target location had to be analysed in the laboratory to derive the analytical data required for model calibration. This makes the MIR technique a valuable resource for accurate, fast and cheap predictions of OC in bulk soils and fractions.
Jeff Baldock is a research scientist working with CSIRO studying the cycling of organic carbon in a range of natural environments.