Dual-spectral deep learning with NGO-based hyperparameter optimization for rapid chili paste quality evaluation

The CNN-LSTM models exhibited suboptimal performance in predicting key quality indicators of fermented chili paste, including reducing sugars, aflatoxin B₁ (AFB₁), and color value. To address this limitation, an attention-enhanced dual-spectral deep learning framework was developed by integrating hyperspectral imaging (HSI), near-infrared spectroscopy (NIRS), a Mixup-Gaussian data augmentation strategy, and Northern Goshawk Optimization (NGO). MSC preprocessing outperformed SG, while UVE provided more robust feature selection than CARS for PLSR. Under HSI-NIR fusion, the CNN-BiLSTM-Attention model with NGO-optimized hyperparameters achieved the best accuracy, improving R² from 0.850 to 0.963 for reducing sugars, 0.815 to 0.956 for AFB₁, and 0.899 to 0.984 for color value compared with MSC-UVE-PLSR. Ablation studies confirmed a consistent trend: HSI-NIR > HSI > NIR. This study developed a novel framework for the rapid and non-destructive prediction of key quality and safety indicators of fermented chili paste in food processing automation.