Chemical composition imaging and class imbalance handling using a 2D-CNN with SMOTE and threshold moving for tobacco...

To address complex chemical relationships and geographical sample imbalance in tobacco origin classification, a method combining chemical composition imaging with a two-dimensional convolutional neural network (2D-CNN) and the synthetic minority over-sampling technique (SMOTE) with threshold moving is proposed. In this approach, multidimensional chemical components are transformed into structured 2D images. Then, a 2D-CNN deep learning model is constructed to capture intricate correlations among chemical indicators through two-dimensional convolutions. Classifier bias arising from class imbalance is mitigated by combining SMOTE with threshold-moving techniques. The results show that the 2D-CNN classification model achieved an overall accuracy of 0.9764 on the test set, with an average precision of 0.9477, a recall of 0.9511, and an F1-score of 0.9492 across eight ecological areas, indicating high model performance. Under the same imbalance handling, the 2D-CNN outperformed a one-dimensional CNN (1D-CNN) by 2.51% in average F1-score, confirming that chemical composition imaging effectively extracts complex inter-indicator relationships. The integration of SMOTE and threshold moving effectively alleviates the impact of class imbalance, significantly enhancing the recognition rates for minority areas. Furthermore, to independently validate the effectiveness of the imbalance handling strategy, it was applied to a separate public image dataset (the Niphad Grape Leaf Disease Dataset). Compared to the baseline model without any imbalance mitigation, the absolute recall of the minority class increased by 40 percentage points.