Non-destructive Detection of Water Content in Porphyra Based on Near-infrared Spectroscopy and Deep Learning

In order to explore the feasibility of combining near-infrared (NIR) spectroscopy and deep learning network for quantitative moisture detection, the dried Porphyra was divided into 479 groups, which detected the NIR spectra and moisture content. Four traditional quantitative moisture prediction models and a convolution neural networks (CNN) deep-learning moisture prediction model were finally established at full spectrum by preprocessing and analyzing the experimental data. After comparing the prediction results of the five models, it was found that the CNN model established by the preprocessing method of S-G smoothing combined with the second derivative had the best prediction effect. Its root-mean-square error of prediction (RMSEP) value was 0.456 and the coefficient of determination of prediction (R_p²) value was 0.990. After optimization, the RMSEP value of the model was reduced to 0.342 and the R_p² value could reach 0.994 (>0.8). At the same time, the ratio of performance to deviation for validation (RPD) was 6.155 (>3), which proved the possibility of practical application of the model in agriculture and food industry. The CNN model could predict the moisture content quickly, accurately, and non-destructive, improve the efficiency and accuracy of moisture detection, and provide an important reference for the quality control of related dry aquatic products.