Abstract
Biomass is greatly influenced by geographic location, soil composition, environment, and climate, making the efficient and accurate identification of growing areas highly significant. This study proposes a classification model for tobacco growing areas based on time series features from thermogravimetric analysis (TGA). This study combines Convolutional Neural Networks (CNN) with Long Short-Term Memory (LSTM) model to process the derivative thermogravimetric (DTG) data, aiming to uncover the inherent time series properties and the continuous and dynamic relationship between temperatures for classifying tobacco growing areas. By analyzing 375 tobacco samples from ten different provinces, CNN is employed to extract local features, while LSTM captures long-term dependencies in the DTG data. The dataset used in this study has a limited sample size, a wide variety of classes, and an imbalance in the number of samples across these classes. Despite these challenges, the model achieves 86.4% accuracy on the test set, significantly surpassing the performance of the traditional Support Vector Machine model, which only achieves 68.2% accuracy. Additionally, the model reveals key temperature ranges crucial for growing area classification associated with the pyrolysis temperature ranges of volatile components, hemicellulose, cellulose, lignin, and CaCO(3) in the tobacco. This model lays the groundwork for the future use of geographical labels to accurately represent tobacco's style and quality, enabling more precise differentiation and improved quality control.