Abstract:
To address the susceptibility of tobacco strips to mildew during storage and the interference of sample scattering and measurement noise with spectral responses, a method based on hyperspectral imaging and an optimized residual temporal convolutional network (ResTCN) was proposed for identifying the mildew status of tobacco strips. A total of 960 tobacco strips from eight batches defined by production year and geographical origin were collected, and hyperspectral images were acquired before and after mildew incubation. Multiplicative scatter correction (MSC) and band-wise standardization (StandardScaler) were applied to preprocess the spectral data, and combined data augmentation using amplitude scaling, baseline shifting, and Gaussian noise injection was performed exclusively on the training spectra. An optimized ResTCN model comprising a stem convolutional layer and four residual dilated convolutional blocks was then constructed. The results showed that: 1) the optimized ResTCN model achieved an accuracy, precision, recall, F1-score, and AUC of 97.06%, 96.59%, 97.79%, 97.09%, and 99.47%, respectively, indicating good performance in distinguishing healthy from mildewed tobacco strips; 2) the ROC curve of the optimized ResTCN model was generally closer to the upper-left corner, and the predicted probabilities of mildew were concentrated near 1.0 for mildewed samples and near 0 for healthy samples, indicating good discrimination between the two classes; 3) wavelength-window occlusion analysis identified 596-606 nm and 626-651 nm as important discriminative regions for identifying the mildew status of tobacco strips, with occlusion of the 596-606 nm region resulting in a relatively large decrease in AUC (ΔAUC) for the optimized ResTCN model. These findings demonstrate that the optimized ResTCN model can effectively distinguish healthy tobacco strips before incubation from mildewed tobacco strips after incubation, providing a methodological reference for spectral identification of mildew in tobacco strips and key wavelength selection.