Development of deep learning method for predicting firmness and soluble solid content of postharvest Korla fragrant pear using Vis/NIR hyperspectral reflectance imaging

Abstract

Abstract The objective of this research was to develop a deep learning method which consisted of stacked auto-encoders (SAE) and fully-connected neural network (FNN) for predicting firmness and soluble solid content (SSC) of postharvest Korla fragrant pear (Pyrus brestschneideri Rehd). Firstly, deep spectral features in visible and near-infrared (380–1030 nm) hyperspectral reflectance image data of pear were extracted by SAE, and then these features were used as input data to predict firmness and SSC by FNN. The SAE-FNN model achieved reasonable prediction performance with R2P = 0.890, RMSEP = 1.81 N and RPDP = 3.05 for firmness, and R2P = 0.921, RMSEP = 0.22% and RPDP = 3.68 for SSC. This research demonstrated that deep learning method coupled with hyperspectral imaging technique can be used for rapid and nondestructive detecting firmness and SSC in Korla fragrant pear, which would be useful for postharvest fruit quality inspections.