Estimation of Soluble Solids Content for Dangshan Pear using Hyperspectral Image Reconstruction Based on Deep Learning

In this study, hyperspectral image reconstruction of Dangshan pear was realized based on SSR technique by extracting different number of channels from the original HSI images as input to the SSR model. Subsequently, the spectra of 79 important wavelengths obtained by CARS algorithm was reconstructed by proposed lightweight SSR model at the optimal number of input channels. The model built by partial least squares (PLS) and reconstructed spectra is used to predict the SSC of pears with a Rp2 of 0.668 and EMSEp of 0.366. The experimental results show that increasing the input channels can improve the reconstruction quality of the SSR model, and the reasonable selection of output channels helps to remove the redundant information and realize the accurate prediction of SSC, which illustrated that the importance of rationally selecting the input channels of the SSR model as well as the output channels to the SSR results.