The prediction of kiwi quality attributes based on multi-source data fusion comprehensive analysis model using HSI and FHSI

Abstract

The soluble solid content (SSC), dry matter content (DMC), and hardness (HD) are widely recognized as important indicators for evaluating the quality of kiwi. In this study, a multi-source data fusion model was constructed using hyperspectral imaging (HSI) and fluorescence hyperspectral imaging (FHSI) technologies, combined with chemometric methods, for the comprehensive evaluation of kiwi quality attributes. Specifically, median filtering (MF) was used for data preprocessing, and support vector regression (SVR), partial least squares regression (PLSR), and convolutional neural network-long short-term memory network (CNN-LSTM) were employed to predict the SSC, DMC, and HD of kiwi. Additionally, the whale optimization algorithm (WOA) was used to optimize the CNN-LSTM to further enhance model performance. The results demonstrated that the WOA-CNN-LSTM, based on the fusion strategy, achieved the best prediction performance. Moreover, compared to traditional models, the WOA-CNN-LSTM exhibited superior performance in handling the high-dimensional data of the fusion model, with significant improvements in prediction performance, with R² increasing by 7.29 %-25.67 %. In conclusion, the deep learning optimization model based on the multi-source data fusion strategy using HSI and FHSI offers a rapid, non-destructive, and efficient solution for predicting the quality parameters of kiwi, providing an innovative and effective approach for fruit quality monitoring.