MA Dian-xu, CAI Yan, YANG Hai-tao, SHAN Chang-ji, DU Guo-fang, LI Zhangyan, SHAN Yuqiong, CHEN jiu-fu
Abstract (
)
Download PDF (
)
Knowledge map
Save
This study employed Attenuated Total Reflection Fourier Transform Infrared Spectroscopy (ATR-FTIR) and second derivative spectroscopy for the analysis of 10 apple varieties, complemented by Convolutional Neural Network (CNN) and its hybrid with Long Short-Term Memory Network (CNN-LSTM) for discrimination. In the FTIR spectra of the 10 apple varieties, strong absorption was observed in the ranges of 3500~2850 cm-1, 1650~1400 cm-1, and 1200~900 cm-1, indicating that apples are rich in carbohydrates, vitamins, amino acids, lipids, organic acids, phenols, and flavonoids. Notably, in the fingerprint region of 1200~900 cm-1, variations in the contents of carbohydrates, pectin, and other components were observed among different apple varieties. Second derivative spectral analysis was performed on the 1200~900 cm-1 fingerprint region. In the second derivative spectra, differences in chemical composition among different apple varieties could be identified. Based on the changes in peak positions and shapes, the Yellow Delicious varieties from Gansu and Xinjiang could be effectively distinguished, and the structural characteristics of different lines of Red Fuji and Qinguan could also be differentiated. This study provides a reliable fingerprint basis for quality control and traceability identification of apples. Further CNN analysis was performed on 329 spectra of 10 apple varieties. A stratified sampling method and K-fold cross-validation were used to divide the spectral data of each variety into a training set and a prediction set at a ratio of 7:3. After a certain number of iterations and training, the CNN model achieved an optimal state with 100% classification accuracy on the training set. Then, the spectra of 66 samples were predicted, and the accuracy in the CNN analysis was 86.4%. To improve the model accuracy, a hybrid CNN-LSTM model combining CNN and LSTM was further applied for discrimination, which increased the accuracy to 90.9%. Both models exhibited excellent classification accuracy. Therefore, the analytical methods of ATR-FTIR, second derivative spectroscopy, CNN, and CNN-LSTM neural networks complement each other in apple analysis and discrimination research, enabling accurate classification of apples. Moreover, this methodological framework can be extended to the classification and discrimination analysis of other substances.