Pathways to improving MIR prediction accuracy for soil carbon, nitrogen and pH using external libraries at national scale

• Determining inter-laboratory bias in reference data can improve MIR model accuracy. • Two pathways presented enhance model accuracy: range alignment or spectral control charts. • Range alignment optimises prediction accuracy for soil properties linked to MIR peaks. • Spectral control charts optimise accuracy for properties predicted from other attributes. • This study gives spectroscopy users a variable-specific guide to optimize prediction accuracy. The European Union’s proposed Soil Monitoring and Resilience Law will require reliable methodologies for monitoring soil properties such as total carbon (TC), total organic carbon (TOC), total nitrogen (TN), and soil pH (Water). This study investigates the potential of MIR spectroscopy in predict these soil attributes at a national level using an external archive of ∼3700 samples collected in 2023. Predictions were based on support vector regression models calibrated with regional archives of ∼850 samples for TC, TOC, and TN, and 1,264 samples for pH scanned in 2021. The findings shows the importance of spectral control charts for properties like pH, which does not present specific peaks in the MIR spectra and is derived from combinations of other chemical groups, like carboxylic acids and carbonates. In this case, spectral control charts improved pH prediction, reducing RMSEP by ∼50 % (0.31 to 0.22), but classified 876 of 3,545 samples as unrepresentative, requiring classical analysis. In contrast, for parameters with distinct MIR peaks, such as TC, TOC, and TN, comparable accuracy was obtained without control charts by simply aligning the range of predicted values with those in the model calibration, i.e. excluding predicted values that fell outside the calibration range. This is a simple strategy and only excluded, on average, 55 samples of ∼3700. In conclusion, spectral control charts are essential for indirect predictions but unnecessary for direct ones. For direct predictions, excluding out-of-range values is an effective strategy to ensure accuracy, with unrepresentative samples requiring reference analysis.