Evaluating validation strategies on the performance of soil property prediction from regional to continental spectral data

Visible-near infrared (vis–NIR) spectroscopy has been widely used to characterize soil information from field global scales. Before applying a calibrated spectral predictive model acquire information, either independent validation or k-fold cross is evaluate performance. However, there no consensus on which strategy more suitable and robust when evaluating performance for the studies in different The objective of this study compare two strategies coupling calibration sizes (a ratio 2:1, 4:1 9:1) sampling (random (RS), rank, Kennard-Stone (KS), rank-Kennard-Stone (RKS) conditioned Latin hypercube (cLHS)) across A total 17,272 vis–NIR spectra mineral soils LUCAS data (continental scale) their organic carbon (SOC) clay contents were study, dataset was further split into national (2761 samples France) five regional datasets (110 248 French administrative regions). To eliminate effect changing set performance, consistent test (20% at each all combinations involved strategies. Lin’s concordance correlation coefficient (CCC) cubist stable both SOC sizes, large size (>1400) continental larger can potentially improve small (<300) scale, wider range would result better No silver bullet found among scale. For regions (small set), we high variation (95th percentile minus 5th percentile) CCC models built 50 repeated RS (0.10–0.44 SOC, 0.16–0.52 clay) cLHS (0.08–0.40 0.12–0.36 clay). This finding indicates that one-time selecting uncertainty evaluation therefore should be with caution. Therefore, suggest following: (1) (thousands), random appropriate; (2) (dozens hundreds), and/or evaluation.