An improved method for predicting soluble solids content in apples by heterogeneous transfer learning and near-infrared spectroscopy

Predicting fruit soluble solids content (SSC) is a hot topic in non-destructive detection. Biological variability of fruit decreases the accuracy of the prediction model. Therefore, modeling methods that can reduce the negative effect of biological variability are necessary. In this paper, an improved modeling method based on the partial least squares (PLS) regression was proposed, using convolutional autoencoder and heterogeneous transfer learning for feature extraction. The dataset used for calibration and prediction contains spectra and corresponding SSC values of apples collected from 2012 to 2018. In comparison with the traditional PLS method, the proposed method performed better for long-term SSC prediction of apples with biological variability. The correlation coefficients calculated on validation set were 0.934, 0.940, 0.915, 0.899, 0.901, respectively. The root mean square errors calculated on validation set were 0.736 °Brix, 0.694 °Brix, 0.674 °Brix, 0.571 °Brix, 0.620 °Brix, respectively. Besides, the proposed method could still achieve relatively satisfactory results when the quantity of calibration samples was less.