A novel spectral correction method for eliminating shell interference and improving NIR detection accuracy of...

Spectral interference from the shell significantly affects the detection of internal defect of in-shell Torreya grandis nuts, reducing detection accuracy. To address this issue, a near-infrared spectral correction method combining a fully connected neural network (FNN) with Gaussian filtering (GS) was developed to eliminate interference from the shell spectra and improve the accuracy of black kernel defect detection. The correction performance of three methods, including direct standardization (DS), piecewise direct standardization (PDS), and the proposed FNN combined with GS, was compared. And detection models were constructed using RF, SVM, XGBoost, and LightGBM. Results shows that DS and PDS provide limited correction, leading to no significant improvement and even a decline in model performance. In contrast, the FNN combined with GS achieves more effective correction and significantly enhances classification accuracy. On the prediction set, RF and SVM models both reach 100% in accuracy, sensitivity, and specificity. XGBoost and LightGBM achieve 95.83% accuracy, with 100% specificity and 91.67% sensitivity. Compared to models developed using uncorrected raw spectral, RF and SVM models show improvements of 16.67% and 19.44% in sensitivity, and 8.33% and 9.72% in accuracy, respectively. Compared to models developed using kernel spectra, RF and SVM models surpass them in both accuracy and sensitivity. The proposed correction method effectively removes shell-related spectral interference, enhances defect-related spectral features, and enables high-precision, non-destructive detection of black kernel defect. This approach offers valuable reference for internal quality assessment in other nut varieties using NIRS.