Bayesian-optimized stacking ensemble with multimodal feature selection for tobacco leaf maturity classification

The maturity of fresh tobacco leaves significantly impacts the quality of subsequent curing processes, making accurate recognition of tobacco leaves maturity crucial. The development of computer vision has promoted the recognition of tobacco maturity. However, the subtle changes in color, shape, and texture of tobacco leaf images at various maturity levels are one of the main challenges for effective recognition. In this study, we preprocess tobacco leaf images and extract features including color, shape, and texture, collectively forming a multimodal feature set. We then filter, rank, and select these features to enhance their discriminative power and improve classification accuracy.Furthermore, we employ stacking methods to construct composite models integrating multiple base models, namely Random Forest (RF), LightGBM (LGBM), and XGBoost (XGB). We train these base models individually and optimize their parameters using Bayesian optimization to obtain prediction results. Subsequently, we utilize the prediction results of these base models as new features inputted into a meta-model. Our approach achieves a classification accuracy of 95.56 % in tobacco leaf maturity classification, demonstrating superior performance compared to alternative methods and indicating significant practical value for real-world applications.