Wavelength selection method for near-infrared spectroscopy based on standard-sample calibration transfer of mango and apple

Modern near-infrared spectroscopy (NIRS) is widely used in nondestructive testing technology, playing an essential role in the quality sorting of agricultural products, among other application. Nevertheless, the reproducibility of results between instruments limits the application of NIRS commercial processes. Therefore, calibration transfer is usually employed to correct for the differences between instruments. To achieve good calibration transfer results, appropriate consistent wavelength selection can reduce the redundancy of variables and optimize the model. Therefore, this research proposed a new relative-error-analysis (REA) wavelength selection method based on relative errors to eliminate the bands with large deviations. This method also improved the robustness of the model. Mango and apple datasets were used to verify the feasibility of REA. First, the research used a calibration transfer method to calibrate the spectra. Then the REA was applied to filter the bands; the optimized result was compared with the result of other wavelength selection methods. The root means squared error of the prediction (RMSEP) of mango and apple models improved by 48.80% and 78.82%, respectively. This result demonstrates that the REA wavelength selection method can not only improve the prediction effect but also simplify the model, at the same time, ensuring the accuracy of calibration transfer between spectrographs.

Automated summary

This study proposed a new relative-error-analysis (REA) wavelength selection method based on relative errors to improve the model's robustness by eliminating bands with large deviations, aiming to simplify the model and enhance prediction accuracy.