Developing deep learning based regression approaches for prediction of firmness and pH in Kyoho grape using Vis/NIR hyperspectral imaging

Firmness and pH, the most important quality attributes of grapes, are directly associated with their quality and price. This study aimed to predict firmness and pH of Kyoho grape using hyperspectral imaging (HSI) via a deep learning approach. Stacked auto-encoders (SAE) were applied to extract deep spectral features based on the pixel-level spectra of each sample over the wavelength range of 400.68-1001.61 nm. Subsequently, these features were used as input data to construct deep learning models for assessing firmness and pH. Additionally, the successive projections algorithm and competitive adaptive reweighed sampling (as wavelength selection algorithms) as well as partial least squares (PLS) and least squares support vector machine (LSSVM) (as modeling approaches) were investigated as conventional spectra analysis approaches for comparison. The results showed that the SAE-LSSVM model achieved the optimal performance, with R2p = 0.9232, RMSEP = 0.4422 N, and RPD = 3.26 for firmness, and the SAE-PLS model yielded satisfactory accuracy, with R2p = 0.9005, RMSEP = 0.0781, and RPD = 2.82 for pH. The overall results revealed that SAE could be used as an alternative to deal with high-dimensional hyperspectral image data. Combined with HSI, it could non-destructively and rapidly detect firmness and pH in grapes; this significantly facilitates post-harvest management and may provide a valuable reference for evaluating other internal quality attributes of fruit.

Automated summary

This study aimed to predict firmness and pH of Kyoho grape using hyperspectral imaging and a deep learning approach. The SAE-LSSVM model achieved optimal performance for firmness, while the SAE-PLS model yielded satisfactory accuracy for pH.