Prediction accuracy of local and regional soil total carbon models, calibrated based on visible-near infrared spectra, in the Djerid arid region

Abstract

Visible-near infrared diffuse reflectance spectroscopy (vis-NIR DRS) is recognized as a promising tool for predicting various soil physico-chemical and biological properties. However, models’ applicability, transferability, and scaling are still questionable. Our objective was to study, for total carbon, these aspects in arid context. QuickBird satellite images enabled us to establish parsimonious soil sampling strategies over three different sites selected in Djerid arid region. For each site, a spectral local database was built and merging them allowed us to obtain a regional database. The principal component analysis enabled us to select independent calibration and validation sets. Local spectral models were performing well for two sites and poorly for the site with high salinity. In cross-transfer, local models showed limited geographic robustness. The regional model was less efficient than one of the local models, yet quite satisfactory ( r2 = 0.67, bias = 0.18%, RMSEP = 0.93% and REP = 1.72). The choice of local or regional model should depend not only on performance of the model but also on the purpose of the intended application and the required precision.