Comprehensive comparison on different wavelength selection methods using several near-infrared spectral datasets with different dimensionalities

NIR spectroscopy is widely used in chemical analysis, agricultural science, food safety, and other fields, but its high dimensionality and data redundancy bring analytical challenges. This study aims to compare the performance of different wavelength selection methods in NIR spectral datasets with different dimensionalities to provide a reference for researchers. The wavelength selection methods in this study were classified into four categories according to their principles, which are partial least squares (PLS) parameter-based methods, intelligent optimization algorithms (IOA)-based methods, model population analysis (MPA)-based methods and wavelength interval selection (WIS) methods. The performance of the models was compared in terms of R²_C, R²_P, root mean square error of calibration (RMSEC), root mean square error of prediction (RMSEP), the number of selected variables, computational time, and the improvement ratio of RMSEP (iRMSEP). The results showed that the models established by MPA-based and WIS methods were more stable and superior to the other categories of wavelength selection methods in most datasets. During the twenty characteristic wavelength selection methods in this study, bootstrapping soft shrinkage (BOSS) and genetic algorithm interval partial least squares (GA-iPLS) show the best performance at the overall level.