PCA- and PLSR-Based Machine Learning Model for Prediction of Urea-N Content in Heterogeneous Soils Using Near-Infrared Spectroscopy

Determining the soil's nitrogen supply accurately and quickly is essential for effective agricultural management. This study explores the use of near-infrared (NIR) spectroscopy combined with spectral pre-processing techniques (such as Savitzky-Golay filtering) and partial least squares regression (PLSR) to assess soil nitrogen levels. Six soil types of varying compositions, treated with different levels of Urea-N fertilizer, were examined. Nitrogen-specific NIR peaks were identified, and regression models were consequently developed. Through a comparison of the performance of the models, the most effective model for nitrogen detection was selected. In calibration, the models performed well, with high R² (over 0.9) and low root mean square error (RMSE) values. The second derivative-based (SD) model slightly outperformed the first derivative-based (FD) model in terms of accuracy. Both models showed minimal bias, indicating reliable performance. During validation, the FD model outperformed the SD model in terms of R², root mean square error of prediction (RMSEP), and residual prediction deviation (RPD). Thus, the FD model demonstrated good predictive ability (R² = 0.77, RPD = 2.06), while the SD model was less effective (R² = 0.65, RPD = 1.77). Compared to previous studies, this study uniquely combines real-time online detection capability with low computational cost, unlike most prior offline approaches, and includes model validation across various soil types. Overall, NIR spectroscopy coupled with multivariate models proves to be a promising tool for the detection of nitrogen levels in various soils.