Prediction Of Internal Quality In Apples Using Hyperspectral Scattering Images

As the living standard of people improving increasingly, how to detect the internal quality of agricultural products nondestructively has been become a new field of scientific research. Hyperspectral image technology combines spectroscopy and image technology advantages, which is a valuable research direction on non-destructive testing of internal quality of agricultural products. But hyperspectral scattering images contain a great of redundant information and they will affect the performance and efficiency of the detection. Wavelength selection of hyperspectral scattering images is a basic approach for improving efficiency and precision. The main idea of this paper is how to select wavelengths of hyperspectral scattering images of apple samples and use them to develop prediction model for predicting apple firmness and soluble solids content (SSC). It provided the technical reference to realize hyperspectral imaging real-time online non-destructive testing for the agricultural products. The main contents of this dissertation are as follows:1. Partial least squares (PLS) projection analysis combined with uninformative variable elimination (UVE) was used to select optimal wavelengths from apple hyperspectral scattering images. The selected feature wavelengths were set as inputs of PLS model. Compared with full wavelengths, the performance parameters of the prediction model were improved. The results showed that the method can eliminate the redundant information of hyperspectral scattering images effectively and there was not shortcoming in genetic algorithm (GA) such as random parameter selection.2. This dissertation studied the wavelength selection of hyperspectral scattering image using propagation neighbor clustering algorithm (AP). On the basis of AP, a fusion model was then developed using UVE and AP models coupled with back propagation neural network. The model fusion technology combined the advantages of single wavelength selection algorithm and minimized the weaknesses of each algorithm. Finally, the ideal prediction models were yielded.3. Further studies of the AP were analyzed for the wavelength selection in the dissertation. Traditional AP for high-dimensional space would appear instability in the performance problems. This dissertation adopted a new similarity measuring function and the similarity matrix was adjusted by semi-supervised strategy. This new semi-supervised clustering algorithm was applied to the feature wavelengths selection of the hyperspectral scattering images for apple firmness and SSC prediction. The simulation results showed that the apple firmness and SSC prediction models had been improved as desired. It provided a new and effective technical support for agricultural products real-time online and non-destructive testing using hyperspectral scattering image technique.