Determination of watermelon soluble solids content based on visible/near infrared spectroscopy with convolutional neural network

Abstract

The soluble solids content（SSC）of fruits is an essential indicator of fruit taste and flavor. The near-infrared spectroscopy (NIRS) is widely used in fruit SSC detection. A common method to deal with such predictive modelling is the partial least-squares (PLSR). Wavelength selection algorithms are often employed to eliminate redundant wavelengths before constructing the model. However, the wavelength selection algorithms can not only increase the model complexity but also may lose valuable information. In this work, 1D-CNN was used to improve the prediction model of the watermelon SSC. First, the best parameter combination for 1D-CNN was found. The root mean square error of prediction (RMSEP) was 0.21 and the determination coefficient of prediction (Rp2) was 0.97. Then, the performance of PLSR and backpropagation neural network (BPNN) are compared. The results show that the Rp2 of 1D-CNN is improved by 14.1% and 6.6%, respectively. Finally, a gradient-based method for visualizing features of CNN regression models, namely regression activation mapping (RAM), was proposed. The results show that the features of 1D-CNN are consistent with the feature peaks of the spectrum, while the features of the BPNN and PLSR are not obvious.