Comparison of Preprocessing Methods in Estimating Total Nitrogen in Soils using PLSR based on Diffuse Reflectance Spectroscopy

This study aimed to compare and evaluate the impact of data preprocessing techniques on the performance of a Partial Least Squares Regression (PLSR) model for predicting total nitrogen (TN) in soils using Diffuse Reflectance Spectroscopy (DRS). A total of 120 soil samples were collected from four saline paddy fields located in Jangan-myeon, Hwaseong-si, Gyeonggi-do, South Korea. The total nitrogen content was determined through laboratory chemical analysis, and spectral data in the VIS-NIR range (400-2,500 nm) were acquired using an ASD FieldSpec PRO4 spectroradiometer after drying the samples. After applying Savitzky-Golay (SG) smoothing to the spectral data, two preprocessing techniques, Standard Normal Variate (SNV) and mean normalization, were separately applied, and the prediction performance of the PLSR models was evaluated for each of the three wavelength ranges: VIS (400-700 nm), NIR (700-2,500 nm), and VIS-NIR (400-2,500 nm). The results showed that the SNV-preprocessed models exhibited 2.42% and 2.34% higher RPD values than the mean-normalized models for the VIS and NIR ranges, respectively. Conversely, in the VIS-NIR range, the mean-normalized model outperformed the SNV model by 1.63% in RPD value. However, the differences were generally minor (RPD < 2.0), and no significant performance differences were observed between the two preprocessing methods. Therefore, for more accurate soil total nitrogen prediction, the acquisition of additional soil data and the combined application of various preprocessing techniques are recommended.