Comparative study of different wavelength selection methods in the transfer of crop kernel qualitive near-infrared models

Abstract

The calibration transfer of qualitative models is of great significance for the further application and popularization of near-infrared spectroscopy (NIRS) in agriculture. Appropriate wavelength selection can effectively improve the transfer efficiency. In this study, the performances of some wavelength selection methods in the transfer of qualitative models were evaluated and compared under the combinations of different pretreatment conditions and transfer algorithms. The compared methods included full-spectrum, synergy interval partial least squares (SIPLS), variable importance in projection (VIP), correlation analysis–based wavelength selection (CAWS) method, and screening wavelengths with consistent and stable signals method (SWCSS). The transfers of two batches of crop kernel datasets (wheat kernel and corn kernel datasets) between instruments were used for testing. The differences in spectral ranges, pretreatments, and transfer algorithms affected the accuracy of the qualitative models during the transfer, and the results showed that CAWS showed a better transfer efficiency than other wavelength selection methods under the combinations of various pretreatment conditions and transfer algorithms. The validation Mathews correlation coefficients (tMCCp) of the transferred CAWS-optimized wheat and corn models (0.718 and 1, respectively) were respectively the second-highest and highest among the tMCCp of the models under all treatment combinations. Additionally, the comparison of the spectral ranges selected by these methods suggested that a better transfer efficiency can be achieved under the co-occurrence of the following conditions: the spectral ranges used for modeling contain characteristic absorption peaks are related to the target to be identified, and the response between instruments is relatively consistent.