The Utility of Laboratory Measurement Uncertainty: A Case-Study Using the Open Soil Spectral Library Service

Abstract

Soil spectroscopy offers the promise of rapid and cost-effective alternatives to traditional wet chemistry methods for analyzing soil properties, with broad applicability across soil types and geographic regions. This promise has led to the creation of large spectral libraries and online modeling tools, including the Open Soil Spectral Library (OSSL). While these tools can advance soil spectroscopy, they often rely on the assumption that laboratory reference measurements are error-free. In this study, we evaluated the accuracy of the OSSL engine in predicting 27 soil properties from mid-infrared spectra, explicitly accounting for uncertainty in laboratory measurements. Predicted properties included carbon fractions (total, organic, inorganic), total nitrogen, cation exchange capacity, electrical conductivity, 1:1 pH, 1:2 CaCl₂ pH, particle-size fractions (clay, silt, sand), NH₄OAc-extractable K, Ca, Mg, and Na; KCl-extractable Al; Mehlich III extractable K, Ca, Mg, Na, Al, Mn, Fe, Cu, and B; and phosphorus extracted by both Bray and Olsen methods. The quality of spectral predictions was evaluated using analytical results from the North American Proficiency Testing (NAPT) program—a geographically diverse dataset comprising standardized soil measurements from multiple laboratories. To quantify predictive performance, OSSL predictions were compared to the median NAPT values using several common statistical metrics, including the Nash–Sutcliffe efficiency coefficient (NSE), concordance correlation coefficient (CCC), ratio of performance to interquartile range (RPIQ), and standardized bias. When compared against median measurements, the carbon and particle size fractions and CEC were predicted well. Total N, 1:1 and 1:2 pH, Mehlich III extractable Ca, Al, and Mg, and NH₄OAc exchangeable Mg were moderately well predicted. All other soil properties were not well predicted. To further evaluate the reliability of OSSL predictions in the context of laboratory measurement uncertainty, we assessed the proportion of predicted values falling within predefined tolerance ranges. Specifically, we calculated the percentage of predictions that fell within 2.5 and 4 times the median absolute deviation (MAD) from the median NAPT value. These thresholds are consistent with those used by the NAPT program to identify potentially erroneous laboratory measurements and serve as a proxy for acceptable uncertainty bounds. More than 80% of OSSL predictions were within NAPT-acceptable measurement ranges for CEC, inorganic C, and clay. Silt, total and organic carbon, sand, and total N predictions were within the acceptable ranges more than 50% of the time. Less than 50% of all other soil properties were predicted within acceptable measurement ranges. These results suggest that OSSL predictions could replace CEC measurements for agricultural surface soils. These results highlight the importance of including the uncertainty of laboratory measurements when evaluating OSSL predictions. We encourage the inclusion of uncertainty estimates when evaluating the quality of all soil spectral models.