Performance of the Global-Local modelling approach for FT-NIR predictions of SOC and TN in diverse Saskatchewan agricultural soils

Precision agriculture requires a reliable, cost-effective method to measure soil organic carbon (SOC) and total nitrogen (TN), and Fourier Transform Near Infrared (FT-NIR) spectroscopy offers a promising solution. Here, we applied the Global-Local model to improve FT-NIR SOC and TN predictions in Saskatchewan agricultural soils. Soil samples (SOC: n = 1876; TN: n = 1442) were collected in 2020 and 2021 from six Saskatchewan agricultural regions. Spectral data were acquired, preprocessed using continuous wavelet transform (CWT), and modelled using Cubist regression. The Global-Local model was applied by combining a small subset of site-specific samples (Lab) with their k-nearest neighbours (Neighbour) from Saskatchewan spectral datasets. Its performance was compared with Leave-One-Site-Out (LOSOV), site-specific, Lab, Neighbour, and traditional spiking. Compared to LOSOV (SOC: R² = 0.55 – 0.76, CCC = 0.67 – 0.79, RPD = 1.20 – 1.44), site-specific models gave higher performance (SOC: R² = 0.71 – 0.88, CCC = 0.82 – 0.92; RPD = 1.59 – 2.70). The Global-Local model performed better than LOSOV and performed similarly to the best Lab or Neighbour models. Compared to the Global-Local, traditional spiking either improved or gave similar results due to higher variability in target variable and spectra datasets. The more accurate models using either spiking or Global-Local than LOSOV confirms the importance of incorporating site-specific samples into training datasets. Our results indicate that the application of the Global-Local model should be restricted to an individual field level, which was its original purpose. Future studies on optimization of the Global-Local model is needed to scale-up its application.