Predicting the oil content of individual corn kernels combining NIR-HSI and multi-stage parameter optimization techniques.

ABSTRACT

Predicting the oil content of individual corn kernels using hyperspectral imaging and ML offers the advantages of being rapid and non-destructive. However, traditional methods rely on expert experience for setting parameters. In response to these limitations, this study has designed an innovative multi-stage grid search technique, tailored to the characteristics of spectral data. Initially, the study automatically screening the best model from up to 504 algorithm combinations. Subsequently, multi-stage grid search is utilized for improving precision. We collected 270 kernel samples from different parts of the ear from 15 high oil and regular corn materials, with oil contents ranging from 1.4% to 13.1%. Experimental results show that the combinations SG + NONE+KS + PLSR(R2 0.8570) and MA + LAR+Random+MLR(R2 0.8523) performed optimally. After parameter optimization, their R2 values increased to 0.9045 and 0.8730, respectively. Additionally, the ACNNR model achieved an R2 of 0.8878 and an RMSE of 0.2243. The improved algorithm significantly outperforms traditional methods and ACNNR model in prediction accuracy and adaptability, offering an effective method for field applications.