Identification and crude protein prediction of porcini mushrooms via deep learning-assisted FTIR fingerprinting

Wild edible mushrooms are natural, green, and sustainable foods, and the key to their post-harvest safety and quality control lies in species identification, geographic traceability, and quality assessment. This study analyses the validity of partial least squares discriminant analysis (PLS-DA) model with residual convolutional neural network (ResNet) model based on Fourier transform infrared (FTIR) spectroscopy and two-dimensional correlation spectroscopy (2DCOS) for identifying species and geographic origin of porcini mushrooms. Exploring the feasibility of the partial least squares regression (PLSR) model and long short-term memory network (LSTM) model for predicting the crude protein content of porcini mushrooms. The results show that the ResNet model outperforms PLS-DA with 1.00 and 0.98 prediction results and 1.00 and 0.92 prediction accuracy for new samples, respectively. The LSTM model was more effective than the PLSR model in estimating crude protein content. Among them, LSTM model with first order derivative-multiple scattering correction-Savitzky-Golay (1D-MSC-SG) preprocessing combination was optimal with a prediction set RPD = 2.7, Sig. = 0.418. Besides, the sequential projection algorithm (SPA) feature extraction improves the prediction ability of the LSTM model. Its residual prediction deviation (RPD) was 2.9, Sig. = 0.627. The study provides a new reference for quality evaluation of edible mushrooms and other food products in the market. • Visualization information and characteristic bands about porcini were obtained. • Residual convolutional neural networks can accurately identify porcini. • Long short-term memory networks predict protein results satisfactorily. • Deep learning is more conducive to porcini identification and prediction.