Enhancing the generalization ability of blueberry impact damage classification model based on SEM-STMMN-SVM

Abstract

To improve the generalization ability of the model, this study proposed the Structural Equation Modeling-Source to Target Min-Max Normalization-Support Vector Machine (SEM-STMMN-SVM) model to classify impact damage in blueberries from different regions and varieties. First, the spectral relative reflectance of samples in the range of 935 nm-1720 nm was measured, and the changes in the internal substances of the fruit before and after bruising were analyzed. Subsequently, the spectral relative reflectance corresponding to 90 feature bands related to blueberry damage, including water content, pigments (anthocyanins), cell structure and tissue state, and biochemical composition (sugars and acids) changes, was selected as input factors. SEM was used to establish correlations between the source domain and the target domain. Then, the STMMN algorithm was used to map the correlated source and target domains to the same data range. Finally, an SVM classification model was constructed. The experimental results indicated that the SEM-STMMN-SVM model outperformed the SVM model in classification accuracy for 12 transfer tasks. The average accuracy in the target domains D1, D2, D3, and D4 increased from 60.26 %, 65.53 %, 61.40 %, and 64.24 % to 87.04 %, 87.04 %, 88.89 %, and 87.74 %, respectively. Overall, the proposed method showed potential in predicting impact damage in blueberries from different sources and varieties. The following research will explore the model transfer effects between different fruits.