DOI: https://doi.org/10.36347/sjavs.2026.v13i02.002

Pages: 40-45

Developing a stable and reliable universal model for different sample characteristics is a complex undertaking. During spectral calibration, variable shifts arising from sample characteristics and environmental conditions can compromise algorithmic accuracy. To address substantial errors in sample detection for universal models, a wavelength screening algorithm is incorporated into the pre-processing stage of model transfer. This removes wavelengths associated with irrelevant variables and information correction, retaining only effective variables as the final feature subset. This facilitates spectral data correction by model transfer algorithms, thereby realising universal modelling. Selecting CARS and RF, along with their combined wavelength screening methods, we constructed universal models for different maize traits in conjunction with the DS method. Experimental validation demonstrated that the combined wavelength screening methods outperformed single-wavelength screening models across all evaluation metrics in the universal models. Among 12 spectral datasets for distinct traits, the CARS-RF-DS model constructed by integrating features extracted from outlier sample sets—demonstrated optimal performance. Its correlation coefficient R² reached 0.9603, RMSEP was 0.0102, with an MAE of 0.0080. For the ear grain sample set, R² exceeded 0.93 and RPD surpassed 3.8, indicating that the CARS-RF-DS model possesses a degree of universality and generalisability.