Rapid species authentication and protein prediction of porcini mushrooms using FTIR-2DCOS coupled with deep learning

Edible fungi are highly valued worldwide for their unique culinary and therapeutic properties. Nevertheless, the pronounced morphological similarities among commercially dehydrated species constantly provoke market misclassification, leading to severe fluctuations in consumer safety and product consistency. To address this challenge, the present research employed Fourier transform infrared (FTIR) spectroscopy integrated with chemometric algorithms to simultaneously authenticate porcini mushroom species and forecast their crude protein levels. By applying a tri-step infrared analytical framework, we successfully mapped the chemical fingerprints of these fungi, elucidating the dynamic response sequences of specific functional groups across target spectral regions and isolating 12 pivotal characteristic variables. Notably, when two-dimensional correlation spectroscopy (2DCOS) imaging was fed into a residual convolutional neural network (ResNet), the hybrid system yielded a species classification accuracy of 100.00%. Concurrently, predictive models for both taxonomy and protein quantification were constructed utilizing diverse feature screening approaches: 2DCOS, the Successive Projections Algorithm (SPA), and Competitive Adaptive Reweighted Sampling (CARS). While 2DCOS outperformed other individual extraction techniques, the synergistic fusion of 2DCOS, CARS, and SPA maximized the overall predictive efficacy. Ultimately, this work highlights 2DCOS-based feature extraction as a highly potent mechanism for evaluating porcini quality, laying a solid technological foundation for the standardization of mushroom commodities.