Spiking improves SOC and TN Model Performance by Shifting the Heterogeneity of Training-Set Toward that of Testing Set

Fourier Transform Near Infrared (FT-NIR) spectroscopy is an important tool for soil organic carbon (SOC) and total nitrogen (TN) measurement and monitoring, but calibration models often fail when applied to new locations. Here, we evaluated FT-NIR spectral spiking as a strategy to improve model performance and transferability across diverse Saskatchewan agricultural soils. Spectra (SOC: n= 1876; TN: n= 1383) from six Saskatchewan agricultural regions were preprocessed using continuous wavelet transform and modelled using site-specific (70:30 train-test split) and leave one site out ( LOSOV ) approaches. Spiking samples were selected using conditioned Latin Hypercube sampling (cLHS), k-means clustering and Kennard-Stone methods, and incorporated into LOSOV calibration sets. Site-specific models achieved higher accuracy for all regions (SOC: coefficient of determination (R 2 ) = 0.71 – 0.90, Lin’s concordance (CCC) = 0.81 – 0.93, ratio of performance to deviation (RPD) = 1.50 – 2.53; TN: R 2 = 0.88 – 0.91, CCC = 0.92 – 0.94, RPD = 2.45 – 3.45) than LOSOV , except Rosetown. Spiking improved LOSOV performance till 20 – 30% inclusion of target-site samples, although most improvement occurred within the first 10%, followed by diminishing returns. The cLHS was most effective for selecting representative samples across all regions. Model improvements arose via shifts in the coefficient of variation of the training dataset; neither selected features nor optimized hyperparameters were significant. This study demonstrated that aligning the variance structure between training and testing sets is important for FT-NIR model transferability. Therefore, incorporating a small, strategically selected set of local samples provides a cost-effective approach to improve prediction accuracy, while spiking beyond 20 – 30% offers minimal additional benefits at higher cost.