نوع مقاله : مقاله پژوهشی
عنوان مقاله English
نویسندگان English
The button mushroom (*Agaricus bisporus*) is one of the most important edible mushrooms globally, and its protein content serves as a key indicator of nutritional value, harvest timing, storage conditions, and processability. Conventional methods for protein measurement are time-consuming and costly, rendering them unsuitable for rapid, online monitoring. This study investigated the feasibility of non-destructively predicting the soluble protein content of button mushrooms using visible/near-infrared (Vis/NIR) spectroscopy (400–1100 nm). Two algorithms—Decision Tree and Random Forest—were evaluated, alongside Partial Least Squares Regression (PLSR) as a chemometric reference model. Reflectance spectra were acquired from 300 button mushroom samples, followed immediately by the determination of their soluble protein content using the Bradford assay. To prevent data leakage, a nested, repeated cross-validation approach was employed, incorporating eight spectral pre-processing techniques. For the selection of effective wavelengths, three methods—CARS, SPA, and GA—were compared within the nested cross-validation framework; CARS was identified as the superior method, selecting a median of 9 features. The results demonstrated that the Random Forest model—achieving an R² of 0.8975, an RMSE of 0.1575, and an RPD of 3.13—outperformed both the Decision Tree and PLSR models. Furthermore, permutation testing, response randomization, and bootstrap confidence intervals confirmed the significance and stability of all three models. Accordingly, the combination of Vis/NIR spectroscopy, feature selection via CARS, and a Random Forest model is presented as an efficient, non-destructive, and interpretable approach for estimating the soluble protein content of button mushrooms.
کلیدواژهها English