Deep learning driven methodology for the prediction of mushroom moisture content using a novel LED-based portable hyperspectral imaging system

This study proposes a deep-learning driven methodology for the analysis of mushroom moisture content (MC) datasets acquired using a novel portable hyperspectral imaging (HSI) system. One-dimensional convolutional neural network (1D-CNN) was developed and validated to process the raw HSI data of white button mushrooms (Agaricus bisporus) for MC prediction. For comparison purposes, state-of-the-art machine learning algorithms, i.e., support vector machine regression (SVMR) and partial least squares regression (PLSR) were also investigated for the model development based on five spectra pre-processed methods using two different lighting systems i.e., enhanced light-emitting diode (LED) and tungsten halogen (TH). Overall, the predictive models based on the HSI data acquired using the LED lights (Rp2 of 0.977, RMSEP of 4.27 %, and RPDp of 6.89) exhibited better performances on the prediction of mushroom MC than those models developed using the TH-HSI data (Rp2 of 0.868, RMSEP of 10.69 %, and RPDp of 2.75). Specifically, the 1D-CNN model based on the raw LED-HSI data (Rp2 of 0.972, RMSEP of 4.70 % and RPDp of 6.29) and the SVMR model based on multiplicative scatter correction (MSC) pretreated LED-HSI data (Rp2 of 0.977, RMSEP of 4.27 %, and RPDp of 6.89) achieved exceptional predictive accuracy for mushroom MC. This finding highlights the effectiveness of the 1D-CNN model in the analysis of HSI data, which performed similarly to the SVMR model without requiring complex data preprocessing steps. In addition, the feasibility of employing a novel LED illumination system in conjunction with a portable HSI camera for the precise MC monitoring of button mushrooms was demonstrated in the present work.