Non-Destructive Estimation of Water Fractions by Machine Learning Models During Freeze-Drying

Prediction of water status in post-harvested agriculture products enduring drying is critical to maintain storage conditions. This study focused on the efficiency of multispectral imaging a novel nondestructive analytical tool by combining various machine-learning models such as Feedforward Neural Network (FNN), Decision Tree Regression, Support Vector Regression, and k-nearest neighbors in the prediction of water fractions during freeze-drying of mushrooms. Spectra from multispectral imaging of the Vis–NIR (405–970 nm) region were combined with machine learning models for the quantification of free water (FW), immobilized water (IM), bound water (BW) and total water (TW) during freeze-drying (FD) of shiitake mushrooms. Water distribution tests through low-field nuclear magnetic resonance demonstrated that 36 h of drying sublimates 90.55% freezable water. The modeling approach performed well, and FNN was found to be the best compared to the others. Its prediction efficiency was 97.77% and 95.95% in BW and TW, respectively. In terms of root mean square error, this model obtained the lowest prediction errors compared to the rest of the models for all water fractions. However, the FNN model prediction deviation is determined with the best bias value of 0.1312 for FW. This study provides an excellent platform in predicting the water status and food quality with a rapid and nondestructive multispectral Vis–NIR spectroscopic approach during drying. The techniques successfully handled the complex spectral data when combined with chemometrics and could be useful in the future for the detection of the chemical composition of agricultural products.