Soil organic carbon and total nitrogen multivariate modelling from diverse FT-NIR spectral dataset

This study linked soil FT-NIR spectroscopy with soil organic carbon (SOC) and total nitrogen (TN) content in Saskatchewan (SK) agricultural soils, using a multivariate approach. Soil spectra were acquired along with laboratory measurements of SOC and TN from 1965 Saskatchewan soil samples. Spectral data were transformed using a variety of common pre-treatment approaches: Savitszky-Golay, first and second derivative, standard normal variate, multiplicative scatter correction and continuous wavelet transform. Models were next built using cubist regression tree (Cubist), support vector machine (SVM), and partial least square regression (PLSR) to evaluate the performance of the different pre-treatment/modelling approaches. The continuous wavelets transform was the best performing spectral treatment method for SK agricultural SOC and TN. For predictive model using an extensive dataset, the cubist model performed best for SOC and TN (R² = 0.80 and 0.85) followed by SVM (R² = 0.77 and 0.85) and PLSR (R² = 0.63 and 0.73). However, all models demonstrated the same correlation between predicted and observed values for SOC and TN (CCC = 0.87 and 0.93). The consistent model accuracy with extensive soil dataset suggests model's ability to generalize well beyond the data it was trained on. However, model accuracy varies if trained using different soil zones and Sk agricultural sites, and this suggest the need for careful selection of specific site or soil-zone on which model should be trained. Additionally, this study also underscores the influence of factors beyond sample size and spectra variability, such as coefficient of variation, on the accuracy of SOC and TN predictions.