Soil organic carbon monitoring at the local and provincial scales using visible near-infrared spectroscopy

Soil organic carbon (SOC) has a significant impact on soil health and is vital for achieving net zero emissions in crop production systems. Therefore, developing cost-effective measuring and monitoring methods for SOC is an urgent priority. Visible near-infrared (Vis-NIR) spectroscopy has the potential to aid in SOC monitoring as an efficient and low-cost method. Using the Canadian Prairie Soil Spectral Library (CPSSL, n = 8392), which encompasses the provinces of Manitoba, Saskatchewan and Alberta, we aimed to examine the potential to predict SOC content for subsequent samplings at the local and provincial scales in the Canadian prairies. The local scale involved a long-term field experiment where SOC was measured in 1987 and 2003, with the last sampling in 2015 (n = 74). At the provincial scale, SOC was measured on producer fields in Saskatchewan in 1996, 2005 and 2011, with the last sampling in 2018 (n = 360). Partial least squares regression (PLSR), Cubist and Global-Local were compared by using site-specific (local scale, n = 150), study-specific (provincial scale, n = 1985) and the remaining CPSSL samples (global, n = 2628 and 2633 for the local and provincial scale, respectively) as the calibration sets. The site-specific with PLSR gave the best prediction of SOC for the 2015 sampling at the local scale (root mean square error (RMSE) = 0.14% SOC, ratio of performance to interquartile range (RPIQ) = 7.5) followed by the site-specific with Cubist (RMSE = 0.15%, RPIQ = 6.8), Global-Local (RMSE = 0.20%, RPIQ = 5.1) and the global calibration with Cubist (RMSE = 0.29%, RPIQ = 3.6). The study-specific with Cubist best predicted the 2018 sampling at the provincial scale (RMSE = 0.36%, RPIQ = 2.2) followed by the study-specific with PLSR (RMSE = 0.47%, RPIQ = 1.7), Global-Local (RMSE = 0.46%, RPIQ = 1.8) and the global calibration with Cubist (RMSE = 0.52%, RPIQ = 1.6). While the site- or study-specific calibrations, as expected, provided the most accurate predictions, data mining and machine learning models with the CPSSL showed great promise, especially at the local scale. Our results showed that SOC monitoring at the local scale with Vis-NIR using site-specific samples can be reliable, thereby reducing costs. At the larger provincial scale, models based on soil texture and/or soil classification, and inclusion of covariates may be necessary to improve prediction precision depending on the level of accuracy required. Overall, this study suggests that using Vis-NIR in combination with soil spectral libraries with temporal data and machine learning models can improve efficiency in SOC monitoring.