| Banana is a very popular fruit that is widely cultivated in the world, and its sales volume is the second among all fruits. However, at present, quality of banana fruits cannot be assured for consumers due to that development speed of rapid detection techniques of banana quality is slower than that of banana industrialization, leading to the event of ‘Manila carcinoma’ and ‘ethephon’, making consumers being pretty worried about banana quality. This causes that banana growers are confronted with bankruptcy and heavy losses are inflicted on banana industry of china. It can improve greatly inspection efficiency of banana quality and reduce detection costs that advanced hyperspectral imaging(HSI) technique is applied to prediction of banana maturity, browning levels and nutritional ingredients for comprehensive determination of quality levels of bananas based on multiple quality parameters. This possesses important strategic significance for assuring banana quality and realizing industrialization of banana and even all fruits.In this study, quantitative analysis models of banana maturity, browning parameters, and nutritional ingredients of banana pulp were developed using different data processing methods, achieving rapid and accurate prediction and visualization of many quality parameters. Firstly, quality parameters of banana were predicted by the established prediction models. Secondly, succession Adaboost algorithm was used to analyze the significance of each quality parameter for banana quality and assign corresponding weights to each parameter. Thirdly, quality levels of bananas were determined based on prediction values of all these quality parameters and the corresponding weights. The main conclusions were as follows:Least squares support vector machine was used to establish predictive models of maturity parameters of banana based on spectral data and image features, respectively. And fuzzy neural network was applied to fuse predictive results of spectral data models and image feature models. The fused model achieved good predciton with R2 of 0.895, 0.731, 0.883, 0.891, 0.717 and 0.937 for L*, a*, b*, chlorophyll content, hardness and soluble solids content, respectively. The results indicated it was very important to fuse spectral data and image features data for improving predictive abilities of the models.BP neural network was utilized to build predictive models of bowning indicators of banana based on spectral data and image features, respectively. And then fuzzy neural network was used to fuse predictive results of detection models of spectra and image features. The performances of the fusion models were better than the spectral models based on the optimal wavelengths, and the models based on image features. The coefficients of calibration of the fusion models were 0.894, 0.726, 0.849, 0.753, 0.916 and 0.794, respectively.RBF neural network was applied to establishing predictive models of nutritional ingredients of banana based on spectral data and image features, respectively. And then fusion models of predictive results of detection models of spectra and image features were built by fuzzy neural network. Compared to the individual models only based on spectral data or image features, the fusion models were more suitable for detection of nutritional ingredients, and their coefficients of calibration were 0.949, 0.864, 0.794, 0.782, 0.752, 0.884 and 0.921, respectively.Continuous Ada Boost classification algorithm was used to carry out weight analysis of multiple quality indicators and evaluate comprehensively qualtity levels of bananas. The models achieved good prediction results with 95.4% and 93.4% classification accuracy for training set and testing set, respectively. These results indicated that it was feasible that determination of banana quality levels using HSI combined with multi-index comprehensive decision. |
PDF Full Text Request