Evaluating visible near-infrared spectroscopy in context of a repeated sampling survey across the European Union

• We obtained soil organic carbon (SOC) predictions by visible near-infrared spectroscopy (VNIRS) for a repeated soil survey at pan-European scale. • Prior survey information led to improvements in SOC predictions, in particular for underrepresented soil types. • We assessed SOC predictions relative to the laboratory measurement error associated with dry-combustion measurements. • Proportion of SOC predictions within the laboratory confidence interval is proposed as a new metric. • A loss function quantifies the cost-effectiveness of VNIRS based on the relative tolerated error, prediction performance and laboratory costs. Visible near-infrared spectroscopy (VNIRS) has potential to fulfill the increasing need for soil organic carbon (SOC) data to support more cost-effective monitoring. However, VNIRS predictions for large-scale repeated surveys have not yet been systematically compared to the laboratory measurement error from dry-combustion. This study assessed 15,134 SOC pan-European predictions by VNIRS for a second campaign, LUCAS 2015, based on the LUCAS 2009 survey. Models performed well considering the mean prediction metrics (e.g. RMSE: 26–27 g C kg −1 , CCC: 0.94) with marginal differences between approaches. However, relative differences between model approaches performance changed when assessed based on the confidence interval coverage probability (CICP). The CICP assesses whether VNIRS predictions lie within the confidence interval of measured SOC given the laboratory error. Furthermore, we quantified with a loss function how the cost-effectiveness of VNIRS depends on both the laboratory measurement error and the tolerated error in SOC predictions. Depending on the confidence interval considered, the best-case scenario for VNIRS would equate to estimated cost savings between 14–31 k (€) by replacing dry-combustion in the second campaign of a repeated survey on SOC. Overall, prior information from the first survey led to modest to large improvements of VNIRS SOC predictions, depending on the metric considered. Our results showed how mean prediction metrics, the CICP and the loss function can lead to variable interpretations of model performance and ideally should not be evaluated in isolation. Our results further indicated that future research is warranted on calibration models that are interpretable and allow for adequate uncertainty quantification.