Predictions of multiple food quality parameters using near-infrared spectroscopy with a novel multi-task genetic programming approach

Abstract

In order to meet the increasing demand for food safety and quality, new methods for simultaneous and rapid determination of multiple food quality parameters (FQPs) are urgently needed in the food industry. Incorporating near-infrared (NIR) spectroscopy and spectral prediction model for rapid, repeatable, non-destructive, and low running costs quantitative analysis of FQPs is enjoying increasing popularity in the food industry. However, most existing spectrum-based prediction models are trained under a single-task learning framework, that is, a prediction model for each quality parameter and spectrum is constructed separately. This paradigm ignores possible connections among prediction tasks of different FPQs, which may result in the performance degradation of a single FPQ prediction model. This study proposes a novel multi-task genetic programming-based approach named EM4GPO for building multiple FQPs predictions simultaneously. In EM4GPO, the multi-dimensional trees are used to encode the raw NIR spectrum to shared features of multiple FQPs; for each FQP, a least square support vector regression (LS-SVR) modeling is performed on the shared features to obtain private features and prediction model; during the optimization process, a new algorithm is developed to optimize the previously obtained shared and private features, and LS-SVR prediction models through population evolution by combining the multidimensional multiclass genetic programming with multidimensional populations optimization method with nondominated sorting method. The proposed EM4GPO model is evaluated and compared with nine popular NIR prediction models using 10 NIR spectral datasets. The experimental results showed that EM4GPO outperformed other commonly used methods in all datasets which indicates that EM4GPO is competitive and effective in solving the problem of multiple FQPs predictions using the NIR spectrum.

Introduction

As globalization proceeds, the food industry has evolved rapidly toward higher-volume, high-quality, safe food production (Nakat & Bou-Mitri, 2021). For this purpose, requirements for the fast, accurate, and non-destructive assessment of food quality have significantly increased over the past decade. Among them, near-infrared (NIR) spectroscopy has attracted considerable attention in existing studies (Cortés, Blasco, Aleixos, Cubero, & Talens, 2019; Lu et al., 2022). This technology can implement quantitative or qualitative food product detection via indirect measurements (Cai et al., 2021; Li, Jang, Li, & Liu, 2021). The general procedure of the methods is to (1) capture the spectral information of food products at different wavelengths by NIR spectroscopy, and (2) develop prediction models for determining the food quality parameters (FQP), such as chemical content and mechanical properties of food products from the corresponding spectral responses (Hernández-Hernández, Fernández-Cabanás, Rodríguez-Gutiérrez, Fernández-Prior, & Morales-Sillero, 2022; Shen et al., 2022; Tugnolo et al., 2021; Turgut, Entrenas, Taşkın, Garrido-Varo, & Pérez-Marín, 2022). However, NIR spectra are usually highly complex, such as many overlapping peaks, resulting in broad wavelengths, and wavelengths lying in the region of overtones and combination modes (Manley, 2014). Therefore, the development of FQP prediction models is one of the most critical aspects of food quality measurements using NIR spectroscopy but currently remains challenging.

Spectral prediction models are usually built based on a regression framework. Linear regression techniques, such as multiple linear regression (MLR) (Liew & Lau, 2012), partial least squares regression (PLSR) (Rossi & Lozano, 2020), etc., have achieved great success in the quantitative evaluation of food quality. In addition, several nonlinear regression approaches, such as least squares support vector regression (LS-SVR) (Li, Huang, Zhao, & Zhang, 2013), regression genetic programming (M3GPSpectra) (Yang, Wang, Zhao, Huang, & Zhu, 2021), and deep convolutional neural networks (DeepSpectra) (Zhang, Lin, Xu, Luo, & Ying, 2019), are applied for developing prediction models, to handle the complex nonlinear relationship between NIR spectral variables and FQPs. The above-mentioned methods namely single-task learning (STL) methods concentrated on single FQP estimation. Although the methods are easy to use and achieved outstanding results, they face two challenges. On the one hand, extensive sample data and its corresponding labels (FQPs) are necessary for learning methods, especially deep learning. However, the labels are usually obtained through destructive and time-consuming measuring procedures, resulting in the lack of labeled samples and unreliable prediction results (Goisser, Hey, & Kesselheim, 2020; Zhang, Ding, Wang, Guo, & Li, 2020). On the other hand, the food industry is often faced with the task of simultaneously predicting multiple FPQs. Although the task of predicting multiple FQPs can be accomplished using a combination of STL methods, the STL modeling method treats each FQP prediction as an independent task, which will block the information interaction among different FQPs and is typically not conducive to developing high-performance FQP prediction model (Mishra & Passos, 2022).

To overcome the limitations of the STL methods, much effort has been made to explore multi-task learning (MTL) methods to jointly train models for different related tasks. During the concurrent learning process, each model shares the knowledge it has learned with other models as meta-information for training them, which alleviates the negative effect of the cold-start problem caused by a few training samples on the STL methods (Feng, Liu, Wu, & Zuo, 2022). Meanwhile, these tasks restrict each other to avoid over-learning through information sharing, so that all prediction models can attain better generalization (Ng et al., 2019). MTL methods such as multi-response partial least-squares regression (PLS2) (Pedro & Ferreira, 2007) and multi-task least-squares support vector regression (MTLSSVR) (Xu, An, Qiao, Zhu, & Li, 2013) have achieved outstanding performance in the domain of food quality evaluation based on NIR spectroscopy. However, the development of the classical MTL models depends on the effective features for FPQ prediction that are difficult to craft manually (Hekler et al., 2019). The emergence of artificial intelligence technology, such as the multi-task deep learning (MTDL) method and evolutionary multitask optimization (EMTO) algorithm, greatly promotes the development of multiple parameter predictions (Sosnin et al., 2019; Zhong, Feng, Cai, & Ong, 2018). Currently, MTDL methods typically employ 1-dimensional (1D) convolution modules to extract shared features from the raw spectrum, then map these features to the final output layer using a full connection layer with a linear activation function, and finally yield multiple quality parameters, each of which corresponds to an output node (Assadzadeh, Walker, McDonald, Maharjan, & Panozzo, 2020; Tsakiridis, Keramaris, Theocharis, & Zalidis, 2020). As described above, the 1D convolution module is used to extract shared features that contain single-scale local features but no long-distance dependencies from the raw NIR spectrum, so that, FQPs mapped by these share features may be unreliable (Zhao, Wu, & Liu, 2022). Additionally, due to deep learning models being data-hungry, MTDLs easily fall into overfitting during learning prediction models (Zhang et al., 2019). In EMTO, the common information that is usually captured via the evolving population is used to optimize the concurrent related tasks. The information is transferred across tasks and updated in the transfer process, so that EMTO can learn shared information (shared features) that generalizes to all tasks. This learning paradigm is suitable for the scenario of insufficient data sources (Zheng, Qin, Gong, & Zhou, 2019). However, EMTO attaches great importance to the commonality of all tasks but overlooks the individuality of each task, and may encounter negative transfer of knowledge (Xu, Qin, & Xia, 2021).

Motivated by the aforementioned popular MTL methods, in this study, an evolutionary multitask multidimensional multiclass genetic programming with multidimensional populations optimization (EM4GPO) algorithm is proposed for the simultaneous prediction of multiple FQPs. The main contributions of this work are to (1) use a type of multi-dimensional tree for extracting multi-scale and multi-type NIR spectrum-based shared features for multiple FQPs; (2) design an LS-SVR for each FQP for learning private features and a prediction model based on the extracted shared features; (3) develop an efficient multi-task genetic programming algorithm based on the M3GP algorithm and nondominated sorting method for solving the optimization problem of shared and private features and LS-SVR prediction models; (4) validate the effectiveness of EM4GPO in ten NIR spectral datasets.