Abstract：Soluble solid content (SSC) is a key factor to evaluate the flavor and quality of fruits. The feature extraction of hyperspectral images provides the data basis and method path for the non-destructive estimation of the solid soluble content. Previous studies have shown that fruit internal quality evaluation based on multi-spectrum, fluorescence spectrum, near-infrared spectrum, and electronic nose has achieved good results. However, the lack of multi-feature fusion limits the accurate estimation of fruit quality. Therefore, this study proposed a model based on stacked autoencoder-particle swarm optimization-support vector regression (SAE-PSO-SVR) to predict the solid soluble content of fresh peaches. Firstly, hyperspectral images extracted spectral information, image pixel information corresponding to different bands, and fusion information. Secondly, a universal stacked autoencoder (SAE) was set up to extract the deep features of spectral information, spatial information, and space-spectrum fusion information. Finally, the deep features were used as the input data of the particle swarm optimization-support vector regression (PSO-SVR) model to predict the solid soluble content of fresh peaches.Among them, three hidden layer network structures were designed for the SAE model with spectral information as input data, including 453-300-200-100-40, 453-350-250-150-50 and 453-350-250-100-60. Three network structures of hidden layer nodes were designed forthe SAE model with image information as input data, including 894-700-500-300-50, 894-650-350-200-80 and 894-800-700-500-100. Three hidden layer network structures were designed forthe SAE model with fusion information as input data, including 1347-800-400-200-40, 1347-750-550-400-100 and 1347-700-500-360-150.The experimental results show that the models with SAE structures of 453-300-200-100-40, 894-800-700-500-100 and 1347-750-550-400-100 have the better estimation effect for spectral information, image information and fusion information as input data of the SAE model, and the prediction accuracy of the model based on the deep features of the fusion information was significantly better than that of the model based on spectral features or image features. The SAE model with the structure of 1347-750-550-400-100 was used to extract the deep features of the fusion information to estimate and visualize the solid soluble content of different varieties of fresh peaches. The results show that the prediction performance based on the SAE-PSO-SVR model was the best (R2=0.873 3, RMSE=0.645 1). Therefore, the SAE-PSO-SVR model proposed can improve the estimation accuracy of solid soluble content of fresh peaches, which provide technical support for detecting other components of fresh peaches.
Key words：Solid content; Hyperspectral image; Deep feature; Support vector regression; Peach
[1] CHEN Dong-jie, JIANG Pei-hong, GUO Feng-jun, et al（陈东杰, 姜沛宏, 郭风军, 等）. Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2020, 40(6): 1839.
[2] Li J, Xue L, Liu M H, et al. Advanced Materials Research, 2011, 186: 126.
[3] Penchaiya P, Bobelyn E, Verlinden B E, et al. Journal of Food Engineering, 2009, 94(3-4): 267.
[4] Zhang W，Pan L，Zhao X, et al. Int. International Journal of Food Properties，2016, 19(1)：53.
[5] Pu Y Y, Sun D W, Riccioli C, et al. Food Analytical Methods, 2017，(11): 1.
[6] Fan S, Zhang B, Li J, et al. Biosystems Engineering, 2016, 143：9.
[7] Li J, Chen L. Computers & Electronics in Agriculture, 2017, 142: 524.
[8] Li S，Yu B，Wu W, et al. Neurocomputing，2015, 151：565.
[9] Blaschke T，Olivecrona M，Engkvist O, et al. Molecular Informatics，2018, 37：1700123.
[10] SUN Jun, JIN Hai-tao, LU Bing, et al（孙 俊, 靳海涛, 芦 兵, 等）. Transactions of the Chinese Society of Agricultural Engineering(农业工程学报), 2019, 35(15)：295.
