Using deep learning to reduce nonlinearity effects in near-infrared spectroscopy for accurate quantification of tobacco leaf pectin concentrations

In the near-infrared (NIR) spectroscopic data of complex sample systems, such as tobacco leaves, nonlinearity is fairly significant between the absorbance and concentration. This nonlinearity severely degrades the quantitative results of traditional methods, such as partial least squares regression (PLS), which can be used to construct linear models. The problem was addressed in this study by using deep learning (DL). We employed three different DL models: a one-dimensional convolutional neural network (1D CNN), a deep neural network (DNN), and a stacked autoencoder with feedforward neural networks (SAE-FNNs). By carefully selecting and tuning the architectures and parameters of these models, we were able to find the most suitable model for dealing with such nonlinear relationships. Our experimental findings reveal that both the DNN and the SAE-FNN models excel in addressing the nonlinear issues of pectin concentration in tobacco, surpassing the performance of the classic linear model (PLS). Specifically, the DNN model stands out for its low average root mean squared error of prediction (RMSEP) value and small standard deviation (SD) of RMSEPs, leading to a tighter and more centered distribution of residuals in the prediction set. These DL models not only proficiently identify complex patterns within NIR data but also boast high prediction accuracy and fast implementation, demonstrating their effectiveness in analytical applications.