| Popis: |
Being rich in anthocyanin is one of the most important physiological traits of mulberry fruits. Efficient and non-destructive detection of anthocyanin content and distribution in fruits is important for the breeding, cultivation, harvesting and selling of them. This study aims at building a fast, non-destructive, and high-precision method for detecting and visualizing anthocyanin content of mulberry fruit by using hyperspectral imaging. Visible near-infrared hyperspectral images of the fruits of two varieties at three maturity stages are collected. Successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS) and stacked auto-encoder (SAE) are used to reduce the dimension of high-dimensional hyperspectral data. The least squares-support vector machine and extreme learning machine (ELM) are used to build models for predicting the anthocyanin content of mulberry fruit. And genetic algorithm (GA) is used to optimize the major parameters of models. The results show that the higher the anthocyanin content is, the lower the spectral reflectance is. 15, 7 and 13 characteristic variables are extracted by applying CARS, SPA and SAE respectively. The model based on SAE-GA-ELM achieved the best performance with R2 of 0.97 and the RMSE of 0.22 mg/g in both the training set and testing set, and it is applied to retrieve the distribution of anthocyanin content in mulberry fruits. By applying SAE-GA-ELM model to each pixel of the mulberry fruit images, distribution maps are created to visualize the changes in anthocyanin content of mulberry fruits at three maturity stages. The overall results indicate that hyperspectral imaging, in combination with SAE-GA-ELM, can help achieve rapid, non-destructive and high-precision detection and visualization of anthocyanin content in mulberry fruits. |
