A multi-feature fusion based method for detecting the moisture content of withered black tea

Abstract

This paper proposes a multi-feature fusion-based method for detecting the moisture content of withered black tea. Hyperspectral images at different withering stages were collected, and their spectra, texture and shape features were extracted. Texture features were extracted using the Gray Level Co-occurrence Matrix (GLCM). This approach described the spatial relationships between neighboring or adjacent pixels for calculating the texture characteristics of tea leaves during the withering process. Shape features were extracted through dimensionality reduction applied to the features extracted by a pre-trained Visual Geometry Group 19 (VGG-19) deep convolutional neural network. VGG-19 was capable of extracting low-level image features, such as edges and contours, representing the shape changes in tea leaves caused by moisture loss. These features were then fused with spectral features to build detection models. The impact of spectra, texture, and shape features on prediction accuracy was analyzed using five types of regressors. Results showed substantial improvement in prediction accuracy with multi-feature fusion. For the best partial least squares regression (PLSR) model, fusing all three features achieved a coefficient of determination (R²) of 0.7968 on the test set, improving by 0.0506, 0.054, and 0.0596 compared to models using individual spectra, texture, and shape features, respectively. The proposed method was also validated with an external dataset, consisting of 72 samples, covering two different tea varieties in two seasons. On the external dataset, the PLSR model maintained good generalization, with an R² of 0.7384, improving by 0.1027, 0.1097, and 0.1326 over individual features. This demonstrates that the fusion of spectra, texture, and shape features significantly improves the model's accuracy and robustness across different tea varieties and seasonal variations. This study provides a fast, non-destructive detection method of moisture content in withered black tea, which can facilitate more precise monitoring of the withering process and ensure consistent quality in large-scale tea production.