Integrating digital image analysis, flash GC E-nose, and SHAP-driven interpretable deep learning for non-destructive aging assessment of citri reticulatae pericarpium

This study proposes a novel non-destructive method for evaluating the aging time and quality of citri reticulatae pericarpium (CRP) by integrating digital image and flash gas chromatography electronic nose (GC E-nose). Digital imaging was employed to extract color features and texture characteristics, while a self-developed one-dimensional convolutional neural network-gated recurrent unit (1D-CNN-GRU-Attention) deep learning model achieved an exceptional classification accuracy of 98.19 % for CRP samples aged between 0 and 12 years. SHapley Additive explanation (SHAP) analysis enhanced model transparency by identifying critical color and texture attributes influencing classification. Flash GC E-nose identified 33 volatile compounds, with Lasso and Random Forest (RF) models pinpointing seven key aroma markers linked to aging. Furthermore, partial least squares regression (PLSR) models demonstrated strong correlations between image features and major chemical components (total flavonoids, phenolic acids, and (+)-limonene). This approach overcomes limitations of traditional methods and provides an interpretable framework for non-destructive quality assessment.