Can soil organic carbon in long–term experiments be detected using Vis-NIR spectroscopy?

The determination of soil organic carbon (SOC) stock and its evolution over time is important for understanding the cycling of carbon in soils, and its relationship with agricultural practices and agro-environmental conditions to develop more sustainable management practices. Moreover, this issue is of great importance in the development of monitoring, reporting, and verification (MRV) systems as part of climate change mitigation measures affecting the agricultural sector. In this context, the present study aims to investigate the reliability of visible and near-infrared (Vis-NIR) spectroscopy as a fast method to determine and monitor the SOC content in an MRV system framework. To that aim, soils from a Long-term field experiment (LTE), located in northern Italy and comparing various maize-based forage systems with contrasting residues and fertilization management was used as a case study. Three sampling campaigns (i.e. 2003, 2012, and 2018) were utilized for a total of 162 soil samples collected in the LTE (54 each). Soil samples archived after each campaign were retrieved and scanned using a Vis-NIR spectrometer to create a site-specific soil spectral library (Site-SSL). Aiming to implement a local prediction model, the 54 soil samples and spectra collected in 2003 were used as a training data set to estimate the SOC of the soil samples collected in 2012 and 2018 (108 in total). In practice, 2003 soil survey represented the time zero reference system. Concurrently, a second prediction model was performed from a regional soil sample archive (Reg-SSL). This is a dataset containing 122 soil samples and spectra collected in alluvial soils from the Po Plain, in the same pedo-landscape of the LTE. The Reg-SSL was merged with the 2003 Site-SSL to implement a regional prediction model. The objective of this activity was to compare the SOC estimations for 2012 and 2018 obtained with both Site-SSL and Reg-SSL models. In addition, three strategies were compared, including Random Forest (RF), Cubist (CU), and Memory based learning (MBL) on Site-SSL and Reg-SSL models (totally, six models were created). We also carried out a sensitivity analysis about the influence of training sample size on the accuracy of different models. Results revealed high correlations between measured and predicted values and for this kind of application, the best results were identified for the MBL model and the Site-SSL dataset. This exercise pointed out that Vis-NIR spectral libraries could be used for detecting changes in SOC. In addition, comparing the use of regional and site-specific libraries for SOC estimation, we confirmed that the accuracy of the Reg-SSL model, although more stable, was lower than that of Site-SSL. A sample size of N=40 samples resulted large enough for building a site-specific spectral library.