image
Home Journals Progress in Chemistry
Progress in Chemistry

Abbreviation (ISO4): Prog Chem      Editor in chief: Jincai ZHAO

About  /  Aim & scope  /  Editorial board  /  Indexed  /  Contact  / 
Online first  /  Current issue  /  All issues  /  Top access  /  RSS

Progress in Chemistry >

2024 , Vol. 36 >Issue 12: 1874 - 1892

DOI: https://doi.org/10.7536/PC241117

Chemistry: A Century of Life-Special Edition

Research Progress and Prospects of Modern Spectral Fusion Analysis Technology

  • Jian Yang ,
  • Yu Liu ,
  • Jingyan Li ,
  • Pu Chen ,
  • Yupeng Xu ,
  • Dan Liu ,
  • Xiaoli Chu , *
Expand
  • Sinopec Research Institute of Petroleum Processing Co., LTD., Beijing 100083, China
* e-mail:

Received date: 2024-09-13

  Revised date: 2024-11-19

  Online published: 2024-12-20

Fold

Abstract

Multispectral fusion is an important research and development direction in modern spectral analysis techniques. It realizes the information complementarity and synergy of multispectral data by optimizing and integrating different types of spectra. Combined with chemometric methods, it can improve the prediction accuracy and robustness of the models. This paper systematically introduces multispectral fusion strategies and algorithms, including classic fusion strategies, fusion based on multi-block algorithms, fusion based on multi-way algorithms, and fusion based on deep learning. The application research on single-spectral fusion, two-spectral fusion, three-spectral fusion, and the fusion of spectra with other information is respectively summarized and discussed. On this basis, the advantages and disadvantages, limitations, and basic selection principles of spectral fusion methods are reviewed. Finally, the challenges faced by multispectral fusion analysis techniques and the future prospects are discussed.

Key words： spectral fusion; data fusion; chemometrics; machine learning; deep learning

Cite this article

Jian Yang , Yu Liu , Jingyan Li , Pu Chen , Yupeng Xu , Dan Liu , Xiaoli Chu . Research Progress and Prospects of Modern Spectral Fusion Analysis Technology[J]. Progress in Chemistry, 2024 , 36(12) : 1874 -1892 . DOI: 10.7536/PC241117

1 Introduction

Traditional spectroscopy is a discipline that studies atomic and molecular structures by observing the interaction of selective wavelengths of electromagnetic radiation with material objects. To date, almost all segments of the electromagnetic spectrum, from shorter-wavelength X-rays to longer-wavelength microwaves, have been utilized for spectral analysis, mainly including two categories: molecular spectra and atomic spectra. For example, ultraviolet-visible (UV-Vis) spectroscopy, molecular fluorescence spectroscopy, near-infrared (NIR) spectroscopy, mid-infrared (MIR) spectroscopy, Raman spectroscopy, terahertz spectroscopy, X-ray fluorescence (XRF) spectroscopy, laser-induced breakdown (LIBS) spectroscopy, and their imaging technologies (imaging spectroscopy). Traditional spectroscopic analysis techniques play a crucial role in multiple fields such as chemical analysis, materials science, biomedicine, and environmental monitoring, providing key methods for determining the composition and structure of substances, characterizing materials, diagnosing diseases, and monitoring the environment[1].
Modern spectral analysis techniques originated in the 1990s and have gradually developed with the rise and application of computers. One of its notable features is the use of chemometrics/machine learning methods to process spectral data, enabling non-destructive, rapid, and accurate quantitative or qualitative analysis of complex mixtures (such as petroleum, agricultural products, traditional Chinese medicine, tobacco, food, soil, minerals, coal, metals, etc.)[2]. Especially in the past decade, the rapid development of technologies such as micro-electromechanical systems (MEMS) manufacturing processes, big data, deep learning, cloud computing platforms, and the Internet of Things has played a positive role in promoting modern spectral analysis techniques. These techniques have found practical applications in areas such as high-throughput laboratory analysis, on-site rapid analysis, and industrial online analysis, gradually forming an independent new category of analytical technology[3].
Multispectral fusion is an important research and development direction in modern spectral analysis technology, which optimizes and integrates different types of spectra to achieve complementary advantages of single spectra, thereby obtaining more comprehensive, reliable, and rich characteristic data. This, combined with chemometrics methods, constructs regression or recognition models for quantitative and qualitative analysis of samples[4]. Multispectral fusion comprehensively and deeply mines information by integrating information from multiple sources, fully leveraging the complementarity and synergy among various spectra, to improve the predictive accuracy and stability of models. Compared with using only a single spectrum, multispectral fusion can enhance the classification and quantitative prediction performance of models[5].
In recent years, multispectral fusion analysis technology has been extensively and deeply studied in terms of methodology, algorithms, and applications, achieving significant progress. There are already monographs and review papers at home and abroad that have commented on the progress of research on multispectral fusion methodologies and their applications in fields such as agricultural products, traditional Chinese medicine, food, dairy products, medicine, and industrial process analysis[6⇓⇓⇓⇓⇓⇓⇓⇓-15]. Based on this, this paper systematically sorts out multispectral fusion strategies and algorithms, summarizes the new progress in the application research of multispectral fusion analysis technology over the past five years, and conducts an in-depth discussion on the challenges and future development trends of the technology, aiming to provide novel research perspectives and application ideas for R&D and application personnel in related fields, thereby further promoting the improvement of multispectral fusion analysis technology and enhancing its application maturity.

2 Spectral Fusion Strategies and Algorithms

2.1 Classic Fusion Strategies

Based on the different structures of multispectral data fusion, as shown in Figure 1, fusion strategies can be categorized into three major types: low-level fusion (data-level fusion), mid-level fusion (feature-level fusion), and high-level fusion (decision-level fusion).
图1 经典的三层光谱融合策略示意图

Fig. 1 Schematic diagram of the classic three level spectral fusion strategy

Low-level fusion, also known as spectral data layer fusion, involves arranging data from different spectral sources in a certain order into a matrix, which is the concatenation of spectral vectors (Vectors concatenation). Then, conventional chemometrics/machine learning methods are used to build the final single model. This approach is often referred to as the concatenation method, such as Concatenated PLS, etc. Moros et al.[16] proposed three additional low-level fusion algorithms: Coaddition fusion (CF), Equal rights fusion (ERF), and Outer product fusion (OPF). The outer product and outer sum operations of spectra are typically used for the fusion calculation of two types of spectra. For the fusion of multiple types of spectra, pairwise operations can be performed, or the spectral vectors can first be concatenated, followed by outer product or outer sum operations. To eliminate differences in the magnitude of different spectra and the influence of factors such as noise and background, it is often necessary to preprocess the spectra with derivatives, MSC, and normalization before fusion.
Mid-level fusion, also known as feature-level fusion, involves the extraction of spectral data from different sources, for example, using principal component analysis (PCA) scores, partial least squares regression (PLS) scores, multivariate curve resolution-alternating least squares (MCR-ALS) scores, and wavelet coefficients for one-dimensional spectra, and GRAM, parallel factor analysis (PARAFAC), Tucker3, and other scores for multi-dimensional spectra. The selected feature variables are then vectorized in a certain order[17⇓-19]. Ahmmed et al.[20] achieved more accurate content of eicosapentaenoic acid (EPA) and docosahexaenoic acid (DHA) in krill oil by fusing Raman and MIR spectral feature variables after PCA and establishing a PLS model, compared to single spectrum. Xu et al.[21] fused LIBS and NIR, using a multi-layer perceptron-PCA method to identify Huanglongbing in citrus leaves, with satisfactory results. In addition to traditional spectral feature extraction methods, deep learning approaches can be used for spectral feature extraction. Shi et al.[22] utilized a convolutional neural network (CNN) to extract deep features from an electronic nose and hyperspectral data, combined with a global extension extreme learning machine (GE-ELM), achieving high-precision recognition of rice quality differences.
In mid-level fusion, methods such as PCA or PLS can also be used to sequentially extract the scores of each spectral matrix, then combine these scores, and use feature selection methods to choose the optimal variables. A regression model is then established using traditional multivariate calibration methods. For example, the network-induced supervised learning method (Network induced supervised learning, NI-SL) adopts this strategy[23].
High-level fusion, also known as decision-level fusion, involves building separate classification or regression models for each spectral data source and combining the prediction results of each individual model to obtain the final decision. In practice, high-level fusion often includes low-level and mid-level fusion of spectral data. The method of decision fusion is key to the success of high-level fusion. For classification, common methods include Bayesian consensus based on discrete probability distributions, voting mechanisms, and Dempster-Shafer evidence theory, while for regression, common methods include weighted averaging and the Granger-Ramanathan method[24-25]. Ballabio et al.[26] used majority voting and Bayesian consensus to fuse NIR, MIR, Raman, PTR-MS, and e-nose data for honey type analysis, with results showing that the Bayesian consensus using a combination of NIR+PTR-MS+Raman performed better. Greenberg et al.[27] employed averaging and the Granger-Ramanathan method to fuse the results of predicting soil properties from mid-infrared spectroscopy and XRF spectroscopy, with the Granger-Ramanathan method yielding more robust results.
For high-level fusion, in addition to the fusion of final prediction results, high-level data (such as regression coefficients used for directly calculating the final result) can also be fused. For example, Rodionova et al[28] proposed a multi-block DD-SIMCA method using high-level data fusion, which achieves the final single-category criterion by combining the full distance of individual spectral data arrays, the total distance of multi-block data (cumulative analysis signal), and the joint critical level.

2.2 Fusion Based on Multi-Block Algorithms

In addition to the three fusion strategies based on traditional chemometrics/machine learning methods, there are also some specialized algorithms for analyzing multispectral fusion data (multiblock data), such as multiblock PCA, common dimension analysis (ComDim) and other unsupervised learning methods, as well as multiblock PLS, sequential and orthogonalized PLS (SO-PLS), and response-oriented sequential alternation (ROSA) among supervised learning methods.

2.2.1 ComDim Method

Methods for unsupervised analysis of multi-block data include multi-block PCA methods such as consensus PCA (Consensus PCA, CPCA) and hierarchical PCA (Hierarchical PCA, HPCA)[29], but in the fusion of multi-spectral data, common dimension analysis (Common dimension analysis, ComDim) is often used. The ComDim method, also known as common components and specific weights analysis (Common components and specific weights analysis, CCSW), is a commonly used unsupervised multivariate analysis method for multiple blocks. The purpose of the ComDim method is to find the common directions of sample dispersion in all spectral data blocks and assign each data block a specific weight as a measure of its contribution to that direction[30-31].
As shown in Figure 2, for the m block spectral matrix, blocks X1, X2, …, Xm, the number of samples (number of rows) is the same and is n. The number of features (number of columns) for each block is p1, p2, …, pm, respectively. The ComDim method seeks to obtain n groups of score vectors, q1, q2, …, qr, which represent the common underlying dimensions, where r is the number of common components. The ComDim method first calculates the correlation matrix XiXiT for each block of the spectral matrix, which reflects the similarity between samples. Then, it performs a weighted summation and decomposition to obtain the common component scores qj, weights λj(i), and common component loadings vj(i). Here, the scores qi can be used as sample features for pattern recognition and regression analysis, while the loadings vj(i) can be used to analyze the importance of spectral variables. The weight λj(i) represents the importance of the ith block of the spectral matrix in constructing the common component j.
图2 ComDim算法对m块光谱矩阵分解的示意图

Fig. 2 Schematic diagram of the ComDim algorithm for decomposing m-block spectral matrices

Sushkov et al.[35]utilized the ComDim method to simultaneously analyze the LIBS and Raman spectra of multiple zooplankton, deeply parsing the relationship between elemental and molecular compositions in animal tissues, providing an economical and convenient tool for studying biochemical processes in animals, and also making it possible for the rapid identification of a large number of marine zooplankton species. Williams et al.[36]fused Raman spectra and EEM spectra using three different fusion strategies for phytoplankton cell analysis, with results showing that each model successfully identified unique biomolecular and pigment components present in Riccia cells at different growth stages, which can be used to monitor changes in phytoplankton across various cellular growth phases.
Xie et al.[37] applied the ComDim method to exploratory analysis of external quality parameters, nutritional quality parameters, visible near-infrared hyperspectral and proton nuclear magnetic resonance (1H-NMR) of lemons with different storage periods, revealing the synergistic effects between multiple data blocks, providing useful insights for the quality control and storage management of lemons.
Jouan-Rimbaud et al.[38] systematically reviewed the improvement methods of ComDim, for example, replacing the PCA step in the ComDim algorithm with independent component analysis (ICA) and PLS, etc., to target different multi-block data, thereby enhancing the interpretability of analytical information and the discriminative ability of samples. Galvan et al.[39] combined the ComDim algorithm with a data-driven-soft independent modeling of class analogy (DD-SIMCA) method, proposing a one-class classifier method (DD-ComDim) for multi-block data. When identifying diesel S10 and S500 using a fusion dataset of medium-resolution nuclear magnetic resonance (MR-NMR) and time-domain nuclear magnetic resonance (TD-NMR), this method outperformed the DD-SIMCA method.
ElGhaziri et al.[40] extended the ComDim algorithm and proposed the p-ComDim method for property prediction, which involves decomposing a weighted sum of the association matrices XiXiTYYT. dosSantos et al.[41] used the p-ComDim method to develop models for predicting soil fertility properties from vis-NIR and XRF spectra. The results showed that this method can extract the weight (significance) of each spectrum and the synergy between its variables, providing a better understanding of the relationship between soil components and spectral responses, which aids in the interpretability of regression models.

2.2.2 MB-PLS Method

For the regression of multispectral data, a multiple linear regression model can be established using the scores from multi-block PCA or ComDim methods and the concentration. Alternatively, Hierarchical multiblock PLS (HPLS) or Multiblock PLS (MB-PLS) methods can be used, with MB-PLS being the more commonly adopted. As shown in Figure 3, for spectral matrix 1 (block X1) and another spectral matrix 2 (block X2), the modeling strategy of the MB-PLS method is: first, to establish partial least squares models for each block with the concentration y, extracting the corresponding PLS components (referred to as lower-level models); then, to use the PLS components obtained from each block and the concentration to build an overall partial least squares model (referred to as the upper-level model)[42]. In this process, because the number of variables in each block is much smaller than that of the whole, and each block has specific connotations, the results obtained by the multiblock partial least squares method have a stronger ability to synthesize information, providing greater explanatory power and application value.
图3 2块光谱矩阵的MB-PLS算法流程图

Fig. 3 Flowchart of the MB-PLS algorithm for two-block spectral matrices

The Serial Partial Least Squares (Serial PLS, S-PLS) method is proposed based on MB-PLS. It adopts a classical PLS algorithm using a serial strategy, by passing the concentration matrix, to sequentially establish multiple PLS models; that is, it uses the concentration residuals of the first PLS model to calculate the second model, and then uses the concentration residuals of the second PLS model to reconstruct the first model. This process iterates, updating the concentration residuals after each iteration until convergence[43]. Laxalde et al.[44] compared the advantages and disadvantages of MB-PLS and S-PLS when predicting the content of four components in heavy oil using the fusion of near-infrared and mid-infrared spectroscopy. The results showed that the prediction accuracy varies for different chemical compositions.

2.2.3 OnPLS Method

The classic latent structure orthogonal projection method (OPLS) separates the predictive information (information related to the dependent variable of interest) and non-predictive information (systematic variation information unrelated to the dependent variable) in the data. It achieves this goal by decomposing the data matrix into multiple parts, with the aim of reducing the complexity of the data, improving the interpretability of the model, and enhancing the predictive power of the model. The bidirectional OPLS method (O2PLS) is an extension of the OPLS method to two data blocks, which can obtain the common principal components of the two data matrices as well as the principal components of each block.
Based on the O2PLS method, Löfstedt et al.[45-46] proposed the OnPLS method, extending it to multi-block data arrays. OnPLS is fully symmetric, meaning that the order of the data blocks does not affect the output of OnPLS. These output matrices reveal the shared data structure of each data matrix at three levels: globally joint structure among all data blocks, locally joint structure among subsets of blocks, and unique structure within each data block. OnPLS not only extracts the minimum number of globally joint components with maximum covariance and correlation but also generates locally joint and unique components, which helps in reducing complexity and increasing interpretability while improving the accuracy of regression and classification[47-48]. Recently, Galindo-Prieto et al.[49] applied the concept of variable importance in projection (VIP) to multi-block data analysis, proposing a variable selection method for multi-block orthogonal projection variable influence, MB-VIOP. This method is a model-based variable selection approach that ranks variables according to their impact on globally joint, locally joint, and unique components based on the scores and loadings of the OnPLS model. The OnPLS method is mainly used for multi-block data processing in bio-omics, with relatively fewer studies on its application in spectral fusion.

2.2.4 SO-PLS Method

Multiblock PLS adopts a parallel correction mode, while Sequential and orthogonalized PLS (SO-PLS) uses a series correction approach. For example, for the Raman spectroscopy matrix (block X1) and the near-infrared spectroscopy matrix (block X2), the SO-PLS modeling strategy is to first establish a PLS model between block X1 and concentration y, obtaining corresponding PLS components (such as score matrix TX1 and concentration residual matrix yR, etc.). Then, TX1 is orthogonalized with block X2 to obtain the orthogonally processed spectral matrix X2orth. Next, a partial least squares model is established between the orthogonal matrix X2orth and the concentration residual matrix yR. The final prediction result is given by combining the above two calibration models. As shown in Figure 4, due to the orthogonalization processing used in the SO-PLS method, it can effectively extract additional complementary spectral information from block X2 relative to block X1.
图4 SO-PLS算法示意图

Fig. 4 Schematic diagram of the SO-PLS algorithm

The main advantage of the SO-PLS method is that when the order of data blocks is meaningful and known, it allows for modeling the important data blocks first and then the less useful ones. Mishra et al.[50] utilized the existing deep network Resnet-18 to extract features from the spatial domain of hyperspectral imaging (HIS) of pork belly and combined these with spectral information in the SO-PLS multi-block framework for predicting and analyzing quality parameters of pork belly. The results showed that, for chemical property parameters, models based on spatial features were not as effective as those based on spectral features; the SO-PLS fusion approach improved the model's ability to predict physical property parameters, indicating that the integration of spatial and spectral information mainly benefits physical properties rather than chemical ones. Baqueta et al.[51] fused near-infrared spectroscopy, mass spectrometry, and sensory data for identifying the origin of coffee. Compared to other fusion methods, the SO-PLS method achieved better recognition results, demonstrating that its sequential characteristic can provide incremental contribution information about different blocks and also more directly explain the effects brought by the synergy and complementarity of data blocks.
However, in the absence of prior knowledge, SO-PLS needs to find the most suitable arrangement by exhaustively searching all permutations. When the number of blocks increases, this becomes a very time-consuming and complex task. Campos et al[52] proposed a stepwise SO-PLS method that is fast to compute, easy to implement, and has good predictive and interpretative capabilities.
Mishra et al[53]introduced Canonical PLS into SO-PLS, proposing a method for multi-block data multi-response modeling. This method can use the meta-information of samples to achieve efficient multi-response modeling and improved subspace extraction. Biancolillo et al[54]combined SO-PLS with LDA to extend it for classification analysis of multi-block data. Orth et al[55]applied the SO-PLS-LDA method to the fusion of visible/near-infrared (VNIR) and short-wave infrared (SWIR) spectral imaging data for discriminating germinated and non-germinated barley grains, achieving good prediction results. Gomes et al[56]proposed a Sequentially Orthogonalized One-Class Partial Least Squares (SO-OC-PLS) method based on SO-PLS, used for the fusion of UV and mid-infrared spectra to identify the authenticity of wines. This method can effectively extract both common and different information from each data block, achieving an optimal balance between sensitivity and selectivity.
As the amount of data blocks being processed increases, selecting appropriate preprocessing methods and identifying information data blocks become increasingly complex and time-consuming[57], Diaz-Olivares et al.[58] introduced the selection and sorting method for preprocessing data blocks (PROSAC) into SO-PLS, which can minimize the concern over the order of preprocessing methods or the order of data blocks, and more purposefully build spectral fusion models.
Another challenge in the analysis of multi-block spectral data is the selection of input feature variables. In data fusion methods, the CovSel method has received considerable attention. Similar to the Nonlinear Iterative Partial Least Squares (NIPALS) algorithm of PLS, it iteratively extracts informative variables through maximizing covariance and orthogonalization[59-60]. To improve the computational speed of the CovSel method for handling multi-block and multidimensional data, Mishra[61] proposed a faster covariate selection algorithm, fCovSel. Biancolillo et al.[62-63] combined CovSel with SO-PLS, proposing a multi-block variable selection method, SO-CovSel, which performs CovSel in a sequential orthogonal manner; the extracted variables can be used for regression and classification. Liu et al.[64] fused near-infrared and UV-Visible spectra, using the SO-Cov method to select characteristic wavelengths and establishing a discriminant model for identifying the origin of cardamom using PLS-DA. The results showed that SO-CovSel could provide a more concise model without losing classification power, outperforming the VIP method. Based on SO-CovSel and ROSA, Mishra et al.[65] proposed a Response Oriented Covariates Selection (ROCS) method, which is insensitive to the order of data blocks and maintains the scale invariance resulting from independent modeling of each data block.
The sequential multi-block PLS algorithm (Sequential multi-block PLS algorithm, SMB-PLS) is proposed based on MB-PLS and SO-PLS. It introduces the sequential orthogonalization strategy of SO-PLS into the MB-PLS structure, not only overcoming the shortcoming of the MB-PLS method that cannot independently select the number of principal components for each block, but also visualizing the correlation and orthogonality information between blocks, exploring and explaining the complete data structure without losing information[66]. This method is currently mostly used for processing multi-block data in industrial processes and is rarely used for multi-spectral fusion.

2.2.5 PO-PLS Method

Unlike the SO-PLS method, the Parallel and orthogonalised PLS (PO-PLS) method estimates the common (i.e., overlapping) subspace among multiple data blocks and the subspace orthogonal to the common space for each block without imposing any order, in order to obtain the unique information in each data matrix. This method is achieved through PLS regression, orthogonalization, and generalized canonical correlation analysis (GCA)[67].
The first step of the PO-PLS method is to perform a separate PLS regression for each input block, and then use GCA to explore the correlation structure between these subspaces. GCA is a method for finding linear combinations of different blocks, with directions or dimensions that have sufficiently high canonical correlations being defined as the common subspace. Once the common information is confirmed, the subspaces of each block are orthogonalized with respect to the common subspace, and the remaining information is considered the unique subspace of each block. Then, a new PLS regression is established between the orthogonal subspaces and the concentration vector to obtain the unique score matrix of each block. Finally, the scores of the common subspace and the scores of the subspaces of each block are concatenated to form a new feature variable matrix, which is then regressed against the concentration vector using classical least squares regression to obtain the final regression coefficients[68].

2.2.6 ROSA Method

Response-oriented sequential alternation (ROSA) is a multi-block regression method based on PLS regression proposed by Liland et al.[69]. Similar to the SO-PLS method, ROSA is a serial algorithm where each principal component is derived from one block of data, selected from candidate components that maximize covariance, with the goal of minimizing prediction residuals. As shown in Figure 5, this method first constructs the first principal component PLS model for all blocks, selects the score and loading vectors of the block with the minimum residual, and calculates the concentration residual vector. In subsequent iterations, the score and loading vectors are orthogonalized and normalized against the previous principal component, then PLS models are constructed using the concentration residual vector and spectral data of each block, until the required scores and loadings of all principal components are obtained.
图5 3块光谱矩阵的ROSA算法流程图

Fig. 5 Flowchart of the ROSA algorithm for three-block spectral matrice

The ROSA method adopts a spectral data block competition rule, which is a forward selection approach. In each iteration, different data blocks have new opportunities to be selected, and thus, different data blocks can be used multiple times. Since only the candidate principal component scores and related prediction residuals are considered in the block selection, unlike the MB-PLS or SO-PLS methods, the order and units of the block data do not affect the results of the ROSA algorithm. Additionally, ROSA has high computational efficiency because it does not require an optimization criterion for iterative convergence; it only needs to calculate the concentration residuals rather than the residuals of all spectral data blocks. Tanzilli et al.[70] fused four online near-infrared spectral data blocks and multiple process sensor data blocks, using the ROSA method to perform online predictive analysis on two quality parameters of different ABS polymer products. This not only allowed for accurate prediction of these two parameters but also clarified the impact of specific process parameters on the final product quality. Compared with the SO-PLS method, the ROSA method demonstrated advantages in computational efficiency. However, since the ROSA algorithm uses a heuristic approach, it may get trapped in local minima, leading to its model's predictive performance being inferior to that of SO-PLS[71].
To apply the ROSA method to multi-dimensional and multi-block data arrays, Mishra et al.[72-73] proposed the Swiss knife partial least squares (SKPLS) method, which is an extension of the ROSA modeling strategy. This approach enables PLS modeling for single or multiple responses in single-block, multi-block, multi-dimensional, and multi-dimensional multi-block data.

2.3 Fusion Based on Multi-Dimensional Algorithms

Spectral fusion often encounters multidimensional spectral data arrays. A common approach is to unfold the multidimensional data array and then apply low-level spectral fusion strategies, or use multidimensional data analysis methods to decompose the multidimensional spectral array and extract score variables, followed by establishing quantitative and qualitative models through mid-level fusion strategies. As shown in Figure 12, Ríos-Reina et al.[74] used mid-infrared spectroscopy, near-infrared spectroscopy, three-dimensional fluorescence spectroscopy, and nuclear magnetic resonance hydrogen spectrum for the identification and analysis of high-quality Spanish wine vinegar. They decomposed the mid-infrared and near-infrared spectra using PCA, the three-dimensional fluorescence spectra using PARAFAC, and the nuclear magnetic resonance hydrogen spectrum using multivariate curve resolution-alternating least squares (MCR-ALS), and then fused the features to establish an identification model using PLS-DA. For the fusion of multidimensional data blocks in industrial process analysis, as shown in Figure 6, it is also possible to decompose the multidimensional data and then fuse the parameter variables at each process stage with the characteristic variables of multiple data blocks, establishing a quantitative prediction model for the final product quality[75].
图6 基于中层特征的多维多块光谱融合示意图

Fig. 6 Schematic diagram of multi-dimensional and multi-block spectral fusion based on mid-level features

Biancolillo et al.[76] proposed the SO-N-PLS algorithm for multi-dimensional and multi-block spectral data fusion based on SO-PLS and multi-dimensional PLS (N-PLS). Compared to unfolding multi-dimensional data and then fusing it using SO-PLS or MB-PLS methods, when three-dimensional data has a clear trilinear structure, the regression and classification results of the SO-N-PLS method are better, especially for small sample sizes and noisy data.
For the fusion of multiple one-dimensional spectra, it can also be constructed into a three-dimensional spectral array and then processed using multi-dimensional data analysis methods. As shown in Figure 7, Dai et al.[77] applied the N-PLS method to the fusion of NIR, FTIR, and Raman spectra. Before fusion, second-order SG convolution derivative processing was performed on the NIR and FTIR spectra, and MSC preprocessing was carried out on the Raman spectrum. Then, normalization and interpolation were conducted on the NIR, FTIR, and Raman spectra to construct a three-dimensional spectral data array. Finally, the N-PLS method was used to establish a model for predicting the conversion rate of poly-α-olefin base oils, which yielded better results than traditional fusion strategies and the SO-PLS multi-block method.
图7 基于NPLS方法的光谱融合示意图

Fig. 7 Schematic diagram of spectral fusion based on the NPLS method

2.4 Fusion Based on Deep Learning

In recent years, spectral fusion methods based on deep learning have received increasing attention and research. On the one hand, deep learning has a strong ability to extract features, capable of accurately capturing subtle characteristics, and can use different network branches for differentiated feature extraction; on the other hand, methods based on deep learning can be carefully designed with loss functions and network structures to achieve more reasonable adaptive feature fusion. The application of deep learning in spectral fusion can be categorized into three types: first, direct concatenation of spectral data into one-dimensional data or conversion into multi-dimensional data through various data transformation methods, which is then used as input to build regression or classification models using deep learning methods such as CNN; second, utilizing the strong feature extraction capability of deep learning to capture subtle features from each spectrum, followed by model building using traditional machine learning algorithms or deep learning algorithms; third, through the design of deep learning network structures, integrating spectral feature information with other information (such as spatial features in hyperspectral data or machine vision information) to enhance the predictive accuracy and reliability of the model[78]. As shown in Figure 8, Gutiérrez et al.[79] used three different deep learning fusion approaches to establish regression models for predicting the content of 15 nitrogen compounds in grapes based on visible, short-wave, and long-wave near-infrared spectra, all of which achieved better results than single-spectrum combined with classical machine learning, among which the method based on deep learning feature fusion performed the best.
图8 深度学习用于光谱融合的几种不同策略示意图

Fig. 8 Schematic diagrams of several different strategies for spectral fusion based on deep learning

Mishra et al[80]combined visible and near-infrared spectra with parallel input CNN deep learning for predicting the dry matter content of mangoes, yielding better results than single-block CNN and SO-PLS methods. Hong et al[81]used parallel input CNN deep learning to fuse VIS-NIR and MIR spectra for predicting soil organic carbon content, outperforming directly concatenated CNN and O-PLS methods. Sun et al[82]also adopted a similar deep learning framework to fuse Raman spectra with machine vision images, successfully classifying and identifying various ores. Song et al[83]fused visible-near infrared (vis-NIR) and portable X-ray fluorescence (pXRF) spectral data, combining continuous wavelet transform (CWT) with a parallel input two-dimensional convolutional neural network algorithm to improve the prediction accuracy of soil quality index (SQI), showing a significant advantage over SO-PLS methods.
Compared with VNIR data, hyperspectral imaging (HSI) contains less spectral information, but HSI has spatial information that VNIR data lacks. Li et al.[84] used LSTM to extract multi-scale spectral features from hyperspectral imaging and ConvLSTM to extract multi-scale spatial features, then fused them to establish a deep learning model for predicting soil carbon content. Compared with classical machine learning methods, this method complements the advantages of various features of VNIR and HSI, improving the accuracy and stability of soil carbon content prediction. Similarly, there are many studies on using deep learning for the extraction and fusion of spectral and spatial features in hyperspectral imaging of soil[85], food[86], crops[87], etc.
HSI contains rich spatial and spectral information, while LiDAR data includes reliable ground elevation information[88], Xu Haitao et al.[89] considering the respective advantages of the two types of data, used convolutional neural networks and Transformer networks to extract features, and designed a cross-modal feature coupling module to achieve joint extraction of hyperspectral and LiDAR features. In terms of feature fusion, they utilized a bilateral attention network module to maximize the use of their complementary information while reducing the interference of redundant information on the classification results.

3 Advances in the Application Research of Spectral Fusion

3.1 Single-Spectrum Fusion

Fusion of single-spectrum commonly involves three scenarios: one is measuring the spectrum under different parameters or methods, the second is measuring the spectrum using different types of spectral instruments or accessories, and the third is segmenting or preprocessing the spectrum measured on the same instrument and then fusing it in a multi-spectral mode.
Portable near-infrared spectrometers are less expensive than benchtop spectrometers, but they usually cannot cover the entire near-infrared spectral range. Mishra et al.[90] used spectrometers covering two spectral regions, 400~1000 nm and 1000~1700 nm, which carry complementary information, for predicting the water content and soluble solids content in pears. The use of the SO-PLS method significantly improved the prediction accuracy. Ryckewaert et al.[91] also adopted a similar research approach to accurately predict the crude protein and sugar content in sugarcane feed. Mishra et al.[92] used the SO-PL and SO-CovSel methods to fuse hyperspectral imaging data from two near-infrared spectral bands (400~1000 nm, 1000~1700 nm) and established a model for predicting the soluble solid content in grapes. Compared with a single hyperspectral camera, the sequential fusion of data from two-band hyperspectral cameras resulted in lower prediction errors.
In heterogeneous samples, the interaction of light is complex, involving transmission, absorption, and scattering, all of which affect the spectrum. Rey-Bayle et al[93]combined multi-angle spatially resolved near-infrared spectroscopy with the CCSWA method to simultaneously obtain physical and chemical information about samples during the online process. Thanavanich et al[94]measured the near-infrared spectra of mixed sugar solutions using both transmission and transreflectance methods, and then established quantitative models through multiple block fusion algorithms. The results showed that the MB-PLS algorithm had a stronger predictive ability for total sugar concentration. Additionally, information on block importance and explained variance was obtained, providing a basis for selecting the optimal detection method in near-infrared spectroscopy analysis. Awhangbo et al[95]used immersion, telemetric, and polarized near-infrared fiber optic probes to monitor the anaerobic digestion process, enhancing the complementarity of these spectra through the SO-PLS method, thus improving the understanding of the anaerobic digestion process. Casarin et al[96]fused 15 kV and 50 kV XRF spectra using the ComDim method and then used ComDim scores to establish an MLR prediction model for the content of teff flour, which was used to identify adulterated low-value cereal flours.

3.2 Two-Spectrum Fusion

3.2.1 Fusion of Molecular Spectra

For the fusion between molecular spectra, the integration of mid-infrared and Raman spectroscopy has received the most attention due to the more evident complementarity in the molecular structural information reflected by both. Liu et al.[97] combined mid-infrared and Raman spectroscopy with an SVM algorithm for the origin identification of honey samples, and the results showed that both data fusion strategies could enhance the recognition ability of honey origins, with the feature-level data fusion model having a higher prediction accuracy than individual spectral models and data-level fusion models. McKeown et al.[98] used portable mid-infrared and Raman spectroscopy to identify fentanyl precursors N-phenethyl-4-piperidone (NPP) and 4-anilino-N-phenethylpiperidine (ANPP) synthesized by four methods, and compared to single instruments, medium-level data fusion combined with the OPLS-DA method yielded better results. Leng et al.[99] utilized the complementarity of FTIR and Raman spectra of serum samples, employing a convolutional neural network-long short-term memory (CNN-LSTM) and multi-scale convolutional fusion neural network (MFCNN) to establish a multispectral fusion approach for cancer diagnosis, which improved recognition accuracy by about 10% compared to single-spectrum methods. Yu et al.[100] adopted low-level and high-level fusion strategies to use Raman and mid-infrared spectroscopy for identifying pre-mortem and post-mortem fracture categories, and the results indicated that the high-level fusion strategy based on the diversity of constructed models achieved the best outcome; to obtain model diversity, PLS-DA, SVM, RF, and GBDT models were established on both the Raman and FTIR datasets.
In the aspect of fusion between ultraviolet and mid-infrared spectroscopy, Li et al.[102]performed low-level fusion of ultraviolet and mid-infrared spectra, establishing a model for the rapid prediction of physical and chemical parameters of asphalt binders, and through the fused spectral information, identified the chemical components with the greatest impact on fatigue resistance. To improve the accuracy and stability of garlic origin prediction, Han et al.[101]developed a low-level fusion model for ultraviolet and mid-infrared spectra. The results showed that SNV-GA-ANN was the best model for processing ultraviolet spectral data, with an accuracy rate of 99.73%. SNV-GA-RF was the best model for mid-infrared spectral data, with an accuracy rate of 97.34%. After the fusion of ultraviolet and mid-infrared spectral data, the SNV-GA-SVC, SNV-GA-RF, SNV-GA-ANN, and SNV-GA-XGboost models all achieved 100% accuracy on both the training and testing sets.
In the aspect of the fusion of near-infrared spectroscopy and Raman spectroscopy, De Man et al.[103] developed an integrated fiber optic probe for near-infrared and Raman spectroscopy, used for online measurement of the mixing effectiveness in the feed frame of a rotary tablet press. The results showed that using a mid-level fusion strategy could improve the prediction accuracy of active ingredient content. Zhang et al.[104] utilized Raman spectroscopy combined with near-infrared spectroscopy to conduct real-time monitoring of the column chromatography process of traditional Chinese medicine. The results indicated that the mid-level fusion model based on the characteristic variables of both types of spectra had the best predictive ability, which is conducive to process control and improving batch-to-batch consistency.
In addition, Kandpal et al[105]fused near-infrared spectroscopy with mid-infrared spectroscopy and used the SO-PLS method to predict the chemical composition content of tuber crop powder. The prediction accuracy for amylose, starch, and cellulose was significantly improved, but the prediction effect for protein and glucose was not as good as that of the single-spectrum PLS model, which may be related to the distribution of component content in the modeling sample set. Froelich et al[106]applied the ComDim-PLS method to the fusion of EEM and synchronous fluorescence spectra of two nylon samples, and its recognition results were superior to those of traditional PCA and PARAFAC methods.

3.2.2 Fusion of Atomic Spectra

The fusion of atomic spectra mainly focuses on the combination between LIBS and XRF, as XRF is very suitable for analyzing many heavy elements such as S, Cl, and Br, as well as some transition metals. The detection limits of XRF are better than those of LIBS, but its sensitivity to lighter elements like Si, Al, and Mg is lower. In contrast, LIBS is particularly sensitive to light elements such as H, Li, Be, B, and C, making it an ideal complement to XRF. The complementary advantages of elemental information from LIBS and XRF are especially prominent in coal analysis, not only capable of measuring organic light elements in coal closely related to calorific value and volatiles, but also stably detecting inorganic ash-forming elements associated with coal ash, thereby improving the stability and accuracy of predicting coal quality[107].
Li et al.[108] proposed an XRF-assisted LIBS method, which, by combining the elemental spectral lines of both technologies, not only retains the ability of LIBS to directly analyze organic elements such as C and H in coal, but also uses XRF to compensate for the lack of stability of LIBS when measuring inorganic ash-forming elements, significantly improving the repeatability of coal calorific value measurements, meeting the needs of power plant coal blending and combustion optimization. Tian et al.[109] further developed an integrated LIBS-XRF coal quality analyzer. Although long-term fluctuations in XRF signals during power plant measurements may affect prediction accuracy, the accuracy of LIBS-XRF coal quality analysis has been significantly improved through spectral intensity correction and segmented modeling techniques.
In addition, Gamela et al[110]performed low-level fusion of XRF and LIBS, establishing a PLS quantitative model for predicting Cu, K, Sr, and Zn in cocoa beans, which showed better performance than single spectra. Andrade et al[111]used a similar low-level fusion strategy to rapidly predict the metal content on waste mobile phone printed circuit boards using XRF and LIBS.
For LIBS measurements, shock waves and sound waves are generated simultaneously with the laser-induced plasma, where the sound waves carry information related to energy conversion during the process of laser-matter interaction and plasma expansion. Additionally, plasma emission images can provide two-dimensional information about the intensity of plasma radiation, which is closely related to matrix effects, self-absorption, and fluctuations in laser intensity. Therefore, integrating LIBS with laser-induced acoustic signals and plasma emission images can enhance the signal quality and analytical performance of LIBS[112-113]. Alvarez-Llamas et al.[114] combined simultaneously acquired LIBS and laser-induced acoustic data through a mid-level fusion strategy, that is, by combining PCA scores from LIBS spectral data with acoustic feature data as input variables for distinguishing similar minerals, thereby improving the recognition capability of iron-based and calcium-based minerals. Zhou et al.[115] integrated LIBS with laser-induced acoustic signals to improve the transferability of the original SVM model between different LIBS systems using various laser beams, achieving good results with both feature-level and decision-level fusion. Zhang et al.[116] proposed a plasma imaging-spectroscopy fusion method for laser-induced breakdown spectroscopy (PISF-LIBS) to mitigate the impact of spectral fluctuations and enhance the performance of quantitative analysis in LIBS.

3.2.3 Integration of Atomic and Molecular Spectra

The integration between atomic spectra and molecular spectra is represented by Raman spectroscopy and LIBS spectroscopy, one reason being the complementary information between the two, and the other being the very similar structure of their spectral instruments, making it easy to manufacture an integrated machine. For example, Shin et al[117] designed and tested a compact Raman/LIBS integrated system, mainly for the rapid identification analysis of adulterated foods, using a mid-level fusion strategy and elastic net method. Compared with using separate spectral methods, this system combines elemental information and molecular information, improving classification accuracy by about 10%.
Sushkov et al.[118]evaluated the potential of combining LIBS and Raman spectroscopy with PCA, NMF, and ComDim for rapid classification of marine zooplankton. The results showed that, although NMF and ComDim loadings provided some interesting information, PCA based on low-level spectral fusion yielded better results. Ren et al.[119]combined LIBS and Raman spectroscopy for fish species identification, where the LIBS spectra contained emission lines of elements such as C, P, Mg, Ca, H, Na, N, K, and O, and the Raman spectra could detect molecular bands of proteins, carotenoids, lipids, and water. They established a three-level fusion classification model using SVM and CNN methods, and the results indicated that the low-level fusion CNN model achieved the highest classification accuracy. Lin et al.[120]addressed the issue of ginseng origin identification by applying PCA to reduce the dimensionality of the fused LIBS and Raman data. In the CatBoost recognition model, compared to single-spectrum LIBS and Raman, the accuracy of mid-level fusion results from LIBS-Raman improved by 11% and 6%, respectively.
Wang et al.[121] proposed a hybrid feature selection method based on analysis of variance (ANOVA) and particle swarm optimization (PSO), selecting characteristic wavelengths from the Raman and LIBS spectra of ores for fusion. In Raman, different vibration modes of CO32-, SO42-, and SiO32- were identified, while in LIBS, the elemental emission lines of Ca, Al, Na, Fe, Ba, Sr, and Mg could be detected. Then, a CNN was used to establish an ore identification model with an accuracy rate of 98%, whereas the accuracy rates using only Raman and LIBS were slightly lower, at 87.9% and 91.3%, respectively. Yang et al.[122] fused dual-pulse laser-induced breakdown spectroscopy (DP-LIBS) and surface-enhanced Raman spectroscopy (SERS) to construct LS-SVM, MLP-ANN, RBF-ANN, and PNN recognition models for assessing the degree of lead stress in wheat seedlings, and the results were superior to those of models established with single spectra.
There has been considerable research on the fusion of NIR and LIBS spectra. Donkey-hide gelatin, a traditional Chinese medicine, mainly contains collagen, proteins, amino acids, polysaccharides, trace elements, water, and other organic and inorganic components. Xia et al.[123] adopted a low-level strategy to fuse the LIBS and NIR spectra of donkey-hide gelatin, establishing a PLS-DA model for brand identification, with all samples being successfully recognized. Chen et al.[124] proposed a LIBS-VNIR fusion method based on a deep learning network (LVDLNet) to detect adulteration in Polygonatum from different origins. In the LVDLNet model, the interval normalization strategy for LIBS spectral data and the peak-valley focusing refinement strategy in VNIR data processing improved the accuracy of identifying adulterated Polygonatum. Fuentes et al.[125] combined the elemental information provided by LIBS with the molecular information provided by hyperspectral imaging to predict the mineral content in copper concentrate, demonstrating that the mid-level fusion strategy had higher accuracy than the low-level strategy.
There are various spectral fusion methods for the rapid and online analysis of soil and coal quality. Kandpal et al[126]fused MIR spectra with XRF spectra of soil, and used the SO-PLS method to establish a fusion model for predicting pH, organic carbon (OC), phosphorus (P), potassium (K), magnesium (Mg), calcium (Ca), and moisture content (MC). Compared to MIR-PLS, XRF-PLS, and low-level fusion PLS models, the SO-PLS model showed improved prediction accuracy for all soil properties, with the most significant improvements in pH, P, and Ca predictions. Song et al[127]utilized the complementary information between vis-NIR and pXRF spectra, combining feature selection with SO-PLS, to enhance the prediction accuracy of total nitrogen content in soil.
Gao et al.[128⇓-130]proposed a highly repeatable coal calorific value measurement method combining NIR and XRF. NIR can stably measure organic components such as C—H and N—H, which are positively correlated with calorific value, while XRF can stably measure inorganic elements like Na, Al, Si, Ca, Fe, and Mn, which are negatively correlated with calorific value. The combination of the two can greatly improve the repeatability of coal calorific value measurements. Based on this, a rapid coal calorific value analyzer was developed and industrial testing and performance evaluation were conducted at a coal preparation plant. It has certain application potential in industries such as coal mining, washing, power generation, and coking. Yan et al.[131]combined low, medium, and high-level data fusion strategies with Kernel Extreme Learning Machine (KELM) to utilize the information synergy between MIR and LIBS for coal quality analysis. The results showed that the advanced data fusion model, which simultaneously performs parameter optimization and variable selection, is the most effective strategy.

3.3 Tri-Spectral Fusion

The fusion of three types of spectra is often used in the food industry. Kashani et al[132]used a high-level fusion strategy to integrate the results of predicting fish fillet freshness from vis-NIR, SWIR, and fluorescence (FL) spectra, achieving a prediction accuracy of 95%, which increased the accuracy of FL, vis-NIR, and SWIR single spectra by 26%, 10%, and 9%, respectively. He et al[133]constructed a four-dimensional fluorescence spectral data array based on UV, NIR, and 3D fluorescence spectra, by adding three acid-sensitive quantum dots as additional dimensions, and established a classification model for recognizing liquor brands using intermediate data fusion, significantly improving the recognition accuracy. Gao et al[134]fused the near-infrared, mid-infrared, and Raman spectra of yam, employing an intermediate fusion strategy, and successfully identified the geographical origin of yam by establishing a grey wolf optimizer-support vector machine (GWO-SVM) model. Zhang et al[135]investigated the application potential of fusing NIR, fluorescence spectra, and LIBS for identifying edible gelatin, using competitive adaptive reweighted sampling (CARS) to extract feature variables, then building a random forest model to classify the origins of five types of edible gelatin, with its classification accuracy significantly better than that of single-spectrum results.
In rapid soil analysis, Hark et al.[136]used handheld XRF, Raman, and LIBS spectrometers to analyze soil data collected in the laboratory and in the field, employing mid-level feature fusion to improve the accuracy of soil classification. Tavares et al.[137]integrated vis-NIR, XRF, and LIBS for predicting key fertility properties of tropical soils, and through a high-level data fusion strategy, the fusion method of the three spectra provided the best prediction accuracy for clay content, cation exchange capacity, and pH. Di Raimo et al.[138]fused Vis-NIR, MIR, and XRF spectra to evaluate the characteristics of sandy soils.

3.4 Fusion of Spectral and Other Information

The integration of spectroscopy and chromatography (including hyphenated techniques) often leverages the high separation and resolution capabilities of chromatography with the abundant macroscopic information provided by spectroscopy. Crude oil, one of the most complex organic mixtures in the world, contains a large number of unique components, and Fourier transform ion cyclotron resonance mass spectrometry provides high mass accuracy, making detailed analysis of crude oil possible. Gjelsvik et al.[139] integrated mid-infrared spectroscopy, near-infrared spectroscopy, and FT-ICR-MS, using a sequential orthogonal PLSR (SO-PLSR) method after variable selection, which improved the predictive accuracy of the model. This result also indicates the potential of multi-block methods for addressing more complex problems. Zhang et al.[140] combined rapid gas chromatography electronic nose (Flash GC E-Nose) with near-infrared spectroscopy to classify the vintage of yellow wine, where the multi-source information fusion strategy, even when combined with classical algorithms such as PLS-DA, achieved 100% classification accuracy. For quality parameters of yellow wine, such as total acid, non-volatile acid, amino acid nitrogen, and total soluble solids, low-level fusion combined with deep learning yielded the best prediction results. Giannetti et al.[141] used GC-MS, MIR, and NIR, along with SO-PLS-LDA and SO-CovSel-LDA methods, to conduct an identification study on traditional Italian spirits with geographical indications, discussing the impact of different orders of multi-block data on recognition results, and showing that near-infrared spectroscopy features provided necessary information for correct classification. Li et al.[142] used near-infrared spectroscopy (NIR) and gas chromatography-mass spectrometry (GC-MS) to combine trace chemicals with high-content chemicals for multidimensional tea grade identification, demonstrating that mid-level fusion with the random forest method could better utilize feature variables, representing a promising approach for tea grade identification. Raeber et al.[143] combined dielectric barrier discharge ionization mass spectrometry (DBDI-MS) with FTIR spectroscopy, using the SO-PLS-LDA method, which significantly improved the accuracy of classification for essential oils from different origins, and compared to traditional methods, this approach has the potential for high-throughput screening.
Images obtained by color cameras contain rich information about the external quality of samples, such as color, texture, shape, and surface defects. In recent years, the integration of spectral and visual images has become increasingly close. Protein and wet gluten are important indicators that determine the quality of wheat, playing a pivotal role in evaluating its processing and baking performance. Red Green Blue (RGB) imaging easily captures the color of wheat samples, providing intuitive information for quality assessment. Zhang et al.[144] integrated near-infrared hyperspectral imaging with RGB images, extracting sensitive features at multiple levels using the Pearson-CARs-VIF algorithm, then established an analytical model for predicting the content of wheat protein and wet gluten using random forest regression. By comparing the prediction results of different input features, they selected the fusion method with the best prediction accuracy. The origin of ginseng significantly affects its nutritional value and chemical composition. Ping et al.[145] fused HSI spectral information with X-ray image texture information, constructing classification models using mid-level fusion strategies and high-level fusion strategies based on the Stacking algorithm. The results not only outperformed models built solely with HSI spectral information but also further demonstrated the advantages of fusing HSI spectral information with X-ray image texture information in the application of tracing the origin of ginseng. You et al.[146] applied the ComDim method to evaluate the hyperspectral and image texture information, NMR spectra, quality parameters, and electronic nose information of different parts of beef. Compared to the PCA method for low-level data fusion, the ComDim method could reveal relationships between different information sources, providing more specific theoretical support for the processing suitability of different parts of beef. Sheng et al.[147] used intermediate fusion to establish an LS-SVM model for predicting moisture content during the drying process of black tea based on near-infrared spectroscopy and machine vision color texture features, which has strong guiding significance for controlling the quality of black tea drying. Díaz-Romero et al.[148] proposed two deep learning models for waste aluminum classification using the fusion of LIBS and computer vision images, capable of sorting waste aluminum into commercially valuable categories with satisfactory accuracy.
Spectra, in addition to chromatography and image fusion, are also integrated with other analytical information. For example, Fu et al.[149]used near-infrared spectroscopy and acoustic emission spectra combined with data fusion strategies to perform quantitative predictive analysis of moisture content and particle size during the fluidized bed granulation process. Comparative studies showed that models based on near-infrared spectroscopy were better at predicting chemical composition moisture than acoustic emission spectra, but for physical parameters such as particle size, high-level fusion provided the best prediction results because it can compensate for system errors by adjusting the parameters of multiple linear regression, thereby reducing the contribution of weak sensors. Bodor et al.[150]explored the low-level fusion of near-infrared spectroscopy, pollen, and physicochemical analysis data, using PCA-LDA methods to classify honey from different plant and geographical origins, achieving good results, with the accuracy of plant origin classification being significantly better than that of geographical origin. Zhang Binbin et al.[151]installed three types of sensors—near-infrared, Raman, and ultrasonic—at the extruder die head, using a convolutional neural network to extract data features from the three sensors, and then establishing a prediction model for melt density using a long short-term memory network, which significantly improved the prediction accuracy.
Spectra, in addition to being fused with other analytical signals, can also be integrated with process parameters for the understanding and control of the process[152]. For example, Jul-Jørgensen et al.[153] combined Raman spectroscopy with process parameters in pharmaceutical production for process fault detection and diagnosis. In the production of traditional Chinese medicine preparations, multi-point data from various production stages such as extraction, concentration, and packaging can be collected. By using spectral and process data fusion techniques, the optimal process parameters can be identified to achieve precise control of the process. Additionally, Strani et al.[154] integrated near-infrared spectroscopy with process sensor data to predict the quality characteristics of polystyrene products.

4 Selection of Spectral Fusion Methods

When single-spectral analysis technology fails to provide satisfactory results, multi-spectral fusion technology can be considered under the premise of economic and technical feasibility. When choosing a specific fusion method, it is necessary to take into account many issues such as research objectives, data size, and type.
When choosing a fusion strategy, the following basic principles are generally followed: (1) First, use low-level fusion strategies for exploratory analysis. Due to the large sample size, low computational efficiency, and inability to directly fuse data of different magnitudes, normalization, denoising, and removal of redundant variables are usually required as preprocessing before fusion. (2) When there are significant differences in the dimensions and structures of multiple data blocks, choose mid-level fusion strategies. Mid-level fusion extracts features from different source data to form a new feature set, often resulting in better outcomes than low-level fusion strategies, making it the most commonly used fusion strategy. (3) The construction of high-level fusion strategies is more complex; however, it does not take advantage of the interlacing and complementarity of multispectral information, leading to the loss of a large amount of detailed information. Its benefit lies in processing each data block separately, reducing interference between different models, with the aim of further improving the stability of predictions through statistical integration methods, suitable for analytical scenarios requiring high robustness.
Compared with traditional classification and regression methods suitable for one-dimensional spectral data, multi-block methods for data fusion can achieve information fusion at the subspace level, thereby interpreting and understanding multi-block data through principal component loadings and scores. However, each method also has its own advantages and disadvantages. For example, MB-PLS uses a hierarchical PLS method to extract global scores and is highly sensitive to the scale of data blocks. SO-PLS models each block separately and models complementary information in the order of the data blocks, thus depending on the order of the data blocks. PO-PLS allows modeling of common and different information without considering the order of the data blocks, but it requires a large amount of computational cost to optimize the model. Although ROSA is a fast alternative in terms of computation, it may fall into local minima due to the greedy approach in extracting model components. For instance, Karami et al[155] compared the advantages and disadvantages of SO-PLS and ROSA methods for quantitative prediction analysis of soil properties using VNIR and MIR spectral fusion, and the results showed that the SO-PLS method provided better predictive accuracy.
Another advantage of using the multi-block approach is that it allows for the integration of metadata into the model, which facilitates effective extraction and enhances the predictive power of the model. Metadata refers to one or more additional variables that describe the training data, such as experimental conditions, mixing ratios, or process parameters in industrial processes. Therefore, as shown in Figure 9, the multi-block method can be used in the management and control of production processes, predicting and analyzing key qualities of each unit within the process system, to gain a deeper understanding of the causal relationships between various factors in the process system, identify critical quality control points, and stabilize and improve product quality. Due to practical conditions or economic constraints, this metadata is not easily obtainable in real laboratory or industrial production environments.
图9 多块数据处理方法用于生产工艺管控示意图

Fig. 9 Schematic diagram of multi-block data processing methods for production process control

Deep learning can capture deep associations and implicit relationships between different data sources, thereby more effectively fusing multimodal data and enhancing the understanding and analytical capabilities of the data. With the continuous growth in the volume of data, deep learning models have an advantage in processing large-scale data, especially suitable for the fusion between spectra and images. However, deep learning usually requires the design of complex network structures, including the optimization of numerous parameters and layers, which not only demands a large amount of computational resources but also does not always yield good results, particularly under conditions where data is limited or the data distribution is imbalanced. For example, Feng et al.[156] used three types of spectra, HSI, MIR, and LIBS, extracted feature spectra using PCA and autoencoders, and based on SVM, LR, and CNN, respectively established discriminative models for low-level, mid-level, and high-level fusion strategies for rice disease classification. The results showed that CNN did not achieve outstanding performance.
In most cases, multispectral fusion often yields better results than a single model, but this is not always the case, especially for the fusion between molecular spectra or atomic spectra, mainly due to the lack of strong complementarity in information or poor data quality of a certain spectrum. Robert et al.[157] used mid-infrared spectroscopy, near-infrared spectroscopy, Raman spectroscopy, and their fusion methods to predict and comparatively evaluate the performance of fatty acid composition in processed mutton. The results showed that although the fusion method improved the prediction accuracy for some fatty acids, the overall evaluation indicated that mid-infrared spectroscopy was superior. Ferreira et al.[158] compared the prediction results of elements in e-waste before and after the fusion of XRF and LIBS spectra. For Al, the fusion model degraded to the point of completely losing linearity, and for Cu and Fe, the improvement was minimal, making the data fusion strategy unfeasible from an industrial perspective. Ribeiro et al.[159] fused raw pXRF spectral data under three different measurement conditions to establish a P-ComDim model for predicting soil fertility properties, but compared to single models, the fusion model did not significantly improve the performance.

5 Conclusions and Prospects

Multispectral fusion is an important component of modern spectral analysis technology and one of the main research and development directions in recent years. It optimizes and integrates different types of spectra to obtain more comprehensive and reliable feature data, constructing regression or recognition models through chemometrics/machine learning methods, for quantitative and qualitative analysis of samples, aiming to improve the accuracy and stability of model predictions. The combination between molecular spectra, between atomic spectra, between molecular and atomic spectra, and between spectra and other information, has been widely applied in rapid quantitative and qualitative analysis of complex mixtures such as grains, food, traditional Chinese medicine, soil, and coal, via classical three fusion strategies, and fusion methods based on multi-block algorithms, multi-dimensional algorithms, and deep learning. Some multispectral analyzers based on the concept of fusion have also been successfully developed and put into practical use. In summary, multispectral fusion technology will develop towards being more precise, intelligent, practical, and efficient, bringing more innovations and applications to various fields.
Based on the current research and application progress of multi-spectral fusion technology, future work in this field should mainly focus on the following four aspects.
(1) Development of multispectral fusion spectroscopic instruments. Currently, there are commercially available or under development multispectral instruments, including: combinations of LIBS instruments and Raman spectrometers, Raman spectrometers and mid-infrared spectrometers, XRF instruments and LIBS instruments, mid-infrared spectrometers and near-infrared spectrometers, as well as various combinations of spectroscopic imaging instruments. To move towards practical applications, especially for on-site rapid analysis and industrial online analysis, the commercial development of low-cost, integrated multispectral instruments should be emphasized. In addition, the development of accompanying measurement accessories should also be valued to achieve automation and intelligence of multispectral analytical instruments.
(2) Expansion of spectral fusion information sources. The complementarity of information when combining spectra is particularly important. Sometimes, the data quality of a certain spectrum is poor or the complementarity of information is not strong, which may lead to unsatisfactory fusion results. On one hand, it is necessary to accurately judge the complementarity between different spectra and choose appropriate spectra or other analytical methods (such as chromatography, sonograms, images, etc.) for fusion. On the other hand, it is necessary to study the expansion of chemical (molecular or atomic level) information sources of samples under experimental conditions. For example, for surface-enhanced Raman spectroscopy, multiple substrate materials can be selected for the detection target to obtain Raman spectra with different enhancement signals for fusion, enabling rapid quantitative analysis of various trace substances simultaneously. In addition, information such as the physical characteristics of some samples or process parameters can also serve as information sources for fusion, further improving the accuracy and robustness of predictions.
(3) Research on fusion strategies and algorithms. Although there are various spectral fusion strategies and algorithms, the current approach mainly relies on the repeated trial-and-error optimization based on the experience of the users. How to choose and effectively use these methods still requires in-depth study. The selection of the best data fusion method is uncertain, which may be related to the number, distribution, and representativeness of samples in the training set, as well as the optimal selection of modeling parameters[160-161]. Therefore, it is necessary to conduct a deep study on the interpretability of spectral fusion models from multiple perspectives such as physics, chemistry, and algorithms, to further improve the universality of the fusion algorithms and strategies and the standardization of the modeling steps. On this basis, developing commercial software with user-friendly interfaces (such as interactive data visualization)[162], currently, chemometrics toolboxes based on Matlab and Python languages far from meet the needs of general users.
(4) The expansion and in-depth application in practice. At present, the application research of multispectral fusion technology in many fields is not deep enough. Most model developments are still in the exploratory stage under ideal laboratory conditions. Some studies do not consider the hardware costs of spectral instruments and the cost of computing resources, lacking practicality and comparability. Therefore, it is necessary to develop technically rigorous and commercialized spectroscopic databases for specific application scenarios, guided by market demand, to break through the bottleneck of applications. The emergence of the current "cloud + network + terminal" networked system platform provides conditions for the automatic optimization processing and application of multispectral fusion data, which can reduce the investment and usage costs of this technology to a certain extent and improve its application efficiency. This is of great significance for promoting the practical application and industrial development of multispectral fusion technology, helping the technology play a greater role in various fields.

6 Postscript

This article is written in commemoration of the 100th anniversary of the birth of Academician Lu Wanzhen, a renowned Chinese analytical chemist and petrochemical scientist who dedicated her life to the development and innovation of petrochemical analysis technology in China, pioneering in many fields. She laid the technical foundation for the evaluation of crude oil in China, successfully developed the flexible quartz capillary chromatographic column, and led the development of the near-infrared spectroscopy discipline in China. Her achievements were not only breakthroughs in science and technology but also exemplary of the spirit of a scientist[163-164].
In 1994, Academician Lu Wanzhen, with great foresight and strategic vision, established the applied discipline of near-infrared spectroscopy in China and assembled a research team to develop domestically produced laboratory and online near-infrared spectrometers and chemometrics software. She also created a spectral database for crude oil and petroleum products. Many of these research achievements have been industrially applied for a long time on China's refining facilities (such as gasoline blending, catalytic reforming, S Zorb, and lubricant hydroisomerization, etc.), in conjunction with advanced process control and real-time optimization control, providing support for strengthening real-time analysis and production stability optimization, achieving good economic and social benefits[163].
Academician Lu Wanzhen actively promoted the research and application paradigm of spectroscopy combined with chemometrics/machine learning. Her research directions were not only limited to near-infrared spectroscopy and petroleum analysis but also extended this research method to mid-infrared spectroscopy, ultraviolet spectroscopy, and other application fields, demonstrating her profound wisdom and foresight. Currently, modern spectroscopic analysis techniques based on chemometrics/machine learning have covered a wide range of molecular and atomic spectroscopies, including mid-infrared spectroscopy, ultraviolet-visible spectroscopy, molecular fluorescence spectroscopy, Raman spectroscopy, terahertz spectroscopy, laser-induced breakdown spectroscopy, and nuclear magnetic resonance spectroscopy. These techniques have been widely applied in various fields such as agriculture, food, pharmaceuticals, petrochemicals, metallurgy, and geology, achieving large-scale application results in some areas and contributing to the development of science, technology, and the economy.
The multispectral fusion technology reviewed in this paper is an important development direction of modern spectral analysis technology and is a natural extension of the research and application paradigm combining spectroscopy with chemometrics/machine learning. Academician Lu Wanzhen paid great attention to this research direction during her lifetime. Our research group has applied this multispectral fusion technology to predict the conversion rate of poly-α-olefin base oil[77], achieving good application results, which are innovative achievements based on inheritance. Currently, our research group is also using it for rapid analysis research and application of coal, heavy oil, and petroleum coke.

References

Publishing order | Descend order by publishing year | Descend order by cited within
[1]
Wang H P, Chen P, Dai J W, Liu D, Li J Y, Xu Y P, Chu X L. Trac Trends Anal. Chem., 2022, 153: 116648.

[2]
Huo X S, Chen P, Li J Y, Xu Y P, Liu D, Chu X L. Appl. Spectrosc. Rev., 2024, 59(4): 423.

[3]
McLean A, Veettil T C P, Giergiel M, Wood B R. Vib. Spectrosc., 2024, 133: 103708.

[4]
Smilde, Age K., Tormod Næs, and Kristian Hovde Liland. Multiblock data fusion in statistics and machine learning: Applications in the natural and life sciences. John Wiley & Sons, 2022.

[5]
Cocchi M. Data Fusion Methodology and Application, Data Handling in Science and Technology. Elsevier, Oxford, 2019.

[6]
Yao S C, Yu Z Y, Hou Z Y, Guo L B, Zhang L, Ding H B, Lu Y, Wang Q Q, Wang Z. Trac Trends Anal. Chem., 2024, 177: 117795.

[7]
Xu Y L, Zhang J Y, Wang Y Z. Food Chem., 2023, 398: 133939.

[8]
Guo M Q, Wang K Q, Lin H, Wang L, Cao L M, Sui J X. Compr. Rev. Food Sci. Food Saf., 2024, 23(1): e13301.

[9]
Azcarate S M, Ríos-Reina R, Amigo J M, Goicoechea H C. Trac Trends Anal. Chem., 2021, 143: 116355.

[10]
Mishra P, Roger J M, Jouan-Rimbaud-Bouveresse D, Biancolillo A, Marini F, Nordon A, Rutledge D N. Trac Trends Anal. Chem., 2021, 137: 116206.

[11]
Casian T, Nagy B, Kovács B, Galata D L, Hirsch E, Farkas A. Molecules, 2022, 27(15): 4846.

[12]
Hayes E, Greene D, O’Donnell C, O’Shea N, Fenelon M A. Front. Nutr., 2023, 9: 1074688.

[13]
Asachi M, Alonso Camargo-Valero M. Adv. Powder Technol., 2023, 34(7): 104055.

[14]
Dai J W, Wang H P, Chen P, Chu X L. Chinese J. Anal. Chem., 2022, 50(6): 839.

(戴嘉伟, 王海朋, 陈瀑, 褚小立. 分析化学, 2022, 50(6): 839.).

[15]
Deng Z W, Chen Z, Fu J S, Yun Y H. Chinese J. Anal. Chem., 2023, 51(01): 11.

(邓焯文, 陈喆, 付家顺, 云永欢. 分析化学, 2023, 51(1): 11.).

[16]
Moros J, Javier Laserna J. Talanta, 2015, 134: 627.

[17]
Sanaeifar A, Li X L, He Y, Huang Z X, Zhan Z H. Biosyst. Eng., 2021, 210: 206.

[18]
Martínez Bilesio A R, Puig-Castellví F, Tauler R, Sciara M, Fay F, Rasia R M, Burdisso P, García-Reiriz A G. Anal. Chim. Acta, 2024, 1309: 342689.

[19]
Ríos-Reina R, Azcarate S M, Camiña J M, Goicoechea H C. Anal. Chim. Acta, 2020, 1126: 52.

[20]
Ahmmed F, Fuller I D, Killeen D P, Fraser-Miller S J, Gordon K C. ACS Food Sci. Technol., 2021, 1(4): 570.

[21]
Xu F H, Hao Z Q, Huang L, Liu M H, Chen T B, Chen J Y, Zhang L Y, Zhou H M, Yao M Y. Appl. Phys. B, 2020, 126(3): 43.

[22]
Shi Y, Yuan H C, Xiong C N, Zhang Q, Jia S Y, Liu J J, Men H. Sens. Actuat. B Chem., 2021, 333: 129546.

[23]
Campos M P, Sousa R, Pereira A C, Reis M S. Talanta, 2017, 171: 132.

[24]
Maléchaux A, Le Dréau Y, Artaud J, Dupuy N. Talanta, 2020, 217: 121115.

[25]
Javadi S H, Mouazen A M. Remote. Sens., 2021, 13(11): 2023.

[26]
Ballabio D, Robotti E, Grisoni F, Quasso F, Bobba M, Vercelli S, Gosetti F, Calabrese G, Sangiorgi E, Orlandi M, Marengo E. Food Chem., 2018, 266: 79.

[27]
Greenberg I, Vohland M, Seidel M, Hutengs C, Bezard R, Ludwig B. Sensors, 2023, 23(2): 662.

[28]
Rodionova O, Pomerantsev A. Anal. Chim. Acta, 2023, 1265: 341328.

[29]
Smilde A. K., Westerhuis J. A., De Jong S. J F. J. Chemometrics. Society, 2003, 17(6): 323.

[30]
Hassani S, Hanafi M, Qannari E M, Kohler A. Chemom. Intell. Lab. Syst., 2013, 120: 154.

[31]
Hanafi M, Kohler A, Qannari E M. Chemom. Intell. Lab. Syst., 2011, 106(1): 37.

[32]
Qannari E M, Wakeling I, Courcoux P, MacFie H J H. Food Qual. Prefer., 2000, 111-2: 151.

[33]
Jouan-Rimbaud Bouveresse D, Pinto R C, Schmidtke L M, Locquet N, Rutledge D N. Chemom. Intell. Lab. Syst., 2011, 106(2): 173.

[34]
Nielsen J P, Bertrand D, Micklander E, Courcoux P, Munck L. J. Infrared Spectrosc., 2001, 9(4): 275.

[35]
Sushkov N I, Galbács G, Fintor K, Lobus N V, Labutin T A. Anal., 2022, 147(14): 3248.

[36]
Williams I, Matthews H, Holtkamp H U, Nieuwoudt M K, Sewell M A, Simpson M C, Broderick N G R, Novikova N I. Chemom. Intell. Lab. Syst., 2023, 243: 104985.

[37]
Xie J H, Xie W G, You Q, Lei H T, Tian X G, Xu X Y. Food Contr., 2024, 166: 110759.

[38]
Jouan-Rimbaud Bouveresse D, Rutledge D N. J. Chemom., 2024, 38(5): e3454.

[39]
Galvan D, de Andrade J C, Conte-Junior C A, Killner M H M, Bona E. Chemom. Intell. Lab. Syst., 2023, 233: 104748.

[40]
El Ghaziri A, Cariou V, Rutledge D N, Qannari E M. J. Chemom., 2016, 30(8): 420.

[41]
dos Santos F R, de Oliveira J F, Bona E, Barbosa G M C, Melquiades F L. Microchem. J., 2023, 191: 108813.

[42]
Westerhuis J A, Kourti T, MacGregor J F. J. Chemometrics, 1998, 12(5): 301.

[43]
Berglund A, Wold S. J. Chemometrics, 1999, 13: 461.

[44]
Laxalde J, Caillol N, Wahl F, Ruckebusch C, Duponchel L. Fuel, 2014, 133: 310.

[45]
Löfstedt T, Trygg J. J. Chemom., 2011, 25(8): 441.

[46]
Löfstedt T, Hoffman D, Trygg J. Anal. Chim. Acta, 2013, 791: 13.

[47]
Reinke S N, Galindo-Prieto B, Skotare T, Broadhurst D I, Singhania A, Horowitz D, Djukanović R, Hinks T S C, Geladi P, Trygg J, Wheelock C E. Anal. Chem., 2018, 90(22): 13400.

[48]
Skotare T, Sjögren R, Surowiec I, Nilsson D, Trygg J. J. Chemom., 2020, 34(1): e3071.

[49]
Galindo-Prieto B, Geladi P, Trygg J. BMC Bioinform., 2021, 22: 1.

[50]
Mishra P, Albano-Gaglio M, Font-i-Furnols M. J. Chemom., 2024: e3552.

[51]
Baqueta M R, Marini F, Teixeira A L, Goulart B H F, Pilau E J, Valderrama P, Pallone J A L. J. Food Compos. Anal., 2024, 133: 106442.

[52]
Campos M P, Sousa R, Reis M S. J. Chemom., 2018, 32(8): e3032.

[53]
Mishra P. Anal. Chim. Acta, 2023, 1250: 340957.

[54]
Biancolillo A, Måge I, Næs T. Chemom. Intell. Lab. Syst., 2015, 141: 58.

[55]
Helmut Orth S, Marini F, Patrick Fox G, Manley M, Hayward S. Microchem. J., 2023, 191: 108742.

[56]
Gomes A A, Khvalbota L, Onça L, Machyňáková A, Špánik I. Food Chem., 2022, 382: 132271.

[57]
Mishra P, Roger J M, Marini F, Biancolillo A, Rutledge D N. Chemom. Intell. Lab. Syst., 2022, 222: 104497.

[58]
Diaz-Olivares J A, Bendoula R, Saeys W, Ryckewaert M, Adriaens I, Fu X Y, Pastell M, Roger J M, Aernouts B. Anal. Chim. Acta, 2024, 1319: 342965.

[59]
Campos M P, Reis M S. Chemom. Intell. Lab. Syst., 2020, 199: 103959.

[60]
Roger J M, Palagos B, Bertrand D, Fernandez-Ahumada E. Chemom. Intell. Lab. Syst., 2011, 106(2): 216.

[61]
Mishra P. J. Chemom., 2022, 36(5): e3397.

[62]
Biancolillo A, Liland K H, Måge I, Næs T, Bro R. Chemom. Intell. Lab. Syst., 2016, 156: 89.

[63]
Biancolillo A, Marini F, Roger J M. J. Chemom., 2020, 34(2): e3120.

[64]
Liu Z M, Yang S B, Wang Y Z, Zhang J Y. Spectrochim. Acta Part A Mol. Biomol. Spectrosc., 2021, 258: 119872.

[65]
Mishra P, Metz M, Marini F, Biancolillo A, Rutledge D N. Chemom. Intell. Lab. Syst., 2022, 224: 104551.

[66]
Lauzon-Gauthier J, Manolescu P, Duchesne C. Chemom. Intell. Lab. Syst., 2018, 180: 72.

[67]
Måge I, Menichelli E, Næs T. Food Qual. Prefer., 2012, 24(1): 8.

[68]
Næs T, Tomic O, Afseth N K, Segtnan V, Måge I. Chemom. Intell. Lab. Syst., 2013, 124: 32.

[69]
Liland K H, Næs T, Indahl U G. J. Chemom., 2016, 30(11): 651.

[70]
Tanzilli D, Strani L, Bonacini F, Ferrando A, Cocchi M, Durante C. Anal. Chim. Acta, 2024, 1316: 342851.

[71]
Strani L, Vitale R, Tanzilli D, Bonacini F, Perolo A, Mantovani E, Ferrando A, Cocchi M. Sensors, 2022, 22(4): 1436.

[72]
Mishra P, Liland K H. Anal. Chim. Acta, 2022, 1206: 339786.

[73]
Mishra P, Liland K H, Indahl U G. J. Chemom., 2022, 36(10): e3441.

[74]
Ríos-Reina R, Callejón R M, Savorani F, Amigo J M, Cocchi M. Talanta, 2019, 198: 560.

[75]
Munoz Lopez C A, Lenaerts M, Peeters K, Van Impe J. IFAC-PapersOnLine, 2020, 53(2): 11722.

[76]
Biancolillo A, Næs T, Bro R, Måge I. Chemom. Intell. Lab. Syst., 2017, 164: 113.

[77]
Dai J W, Chen P, Chu X L, Xu B, Su S. Fuel, 2024, 366: 131420.

[78]
Li H, Ju W L, Song Y M, Cao Y Y, Yang W, Li M Z. Comput. Electron. Agric., 2024, 217: 108561.

[79]
Gutiérrez S, Fernández-Novales J, Garde-Cerdán T, Marín-San Román S, Tardaguila J, Diago M P. Inf. Fusion, 2023, 99: 101865.

[80]
Mishra P, Passos D. Anal. Chim. Acta, 2021, 1163: 338520.

[81]
Hong Y S, Chen S C, Hu B F, Wang N, Xue J, Zhuo Z Q, Yang Y Y, Chen Y Y, Peng J, Liu Y L, Mounem Mouazen A, Shi Z. Geoderma, 2023, 437: 116584.

[82]
Sun Y X, Tian Y, Zhang Y Y, Yu M T, Su X Q, Wang Q, Guo J J, Lu Y, Ren L H. Spectrochim. Acta Part A Mol. Biomol. Spectrosc., 2024, 318: 124454.

[83]
Song J H, Shi X Y, Wang H J, Lv X, Zhang W X, Wang J G, Li T S, Li W D. Geoderma, 2024, 447: 116938.

[84]
Li X Y, Li Z M, Qiu H M, Chen G Y, Fan P P, Liu Y. Ecol. Indic., 2024, 160: 111843.

[85]
Li X Y, Qiu H M, Fan P P. Appl. Spectrosc. Rev., 2024: 1.

[86]
Guo Z, Zhang J, Wang H F, Li S L, Shao X J, Dong H W, Sun J S, Geng L J, Zhang Q, Guo Y M, Sun X, Xia L M, Darwish I A. Postharvest Biol. Technol., 2024, 213: 112960.

[87]
Shuai L Y, Li Z Y, Chen Z A, Luo D T, Mu J. Comput. Electron. Agric., 2024, 217: 108577.

[88]
Li J X, Hong D F, Gao L R, Yao J, Zheng K, Zhang B, Chanussot J. Int. J. Appl. Earth Obs. Geoinf., 2022, 112: 102926.

[89]
Xu H T, Liu Y Z, Yan X Y, Li J J, Xue C B., J. Xidian Univ., 2024, (7):1.

徐海涛, 刘玉哲, 闫欣怡, 李娇娇, 薛长斌. 西安电子科技大学学报, 2024, (7): 1.).

[90]
Mishra P, Marini F, Brouwer B, Roger J M, Biancolillo A, Woltering E, Echtelt E H V. Talanta, 2021, 223: 121733.

[91]
Ryckewaert M, Chaix G, Héran D, Zgouz A, Bendoula R. Biosyst. Eng., 2022, 217: 18.

[92]
Mishra P, Xu J L. J. Infrared Spectrosc., 2023, 31(3): 141.

[93]
Rey-Bayle M, Bendoula R, Caillol N, Roger J M. J. Infrared Spectrosc., 2019, 27(2): 134.

[94]
Thanavanich C., Phuangsaijai N., Funsueb S., Theanjumpol P., Kittiwachana S. Asia Pac J Sci Technol., 2024, 29(02): APST-29-02-12.

[95]
Awhangbo L, Bendoula R, Roger J M, Béline F. Chemom. Intell. Lab. Syst., 2020, 196: 103905.

[96]
Casarin P, dos Santos L D, Viell F L G, Melquiades F L, Bona E. Anal. Chim. Acta, 2023, 1276: 341639.

[97]
Liu N, Chen L Z, Liu C L, Sun X R, Zhang S Z. Infrared Phys. Technol., 2024, 139: 105327.

[98]
McKeown H E, Rook T J, Pearson J R, Jones O A H. Forensic Chem., 2023, 33: 100476.

[99]
Leng H Y, Chen C, Chen C, Chen F F, Du Z J, Chen J J, Yang B, Zuo E G, Xiao M, Lv X Y, Liu P. Spectrochim. Acta Part A Mol. Biomol. Spectrosc., 2023, 285: 121839.

[100]
Yu K, Wu H, Xiong H L, Wang G J, Wei X, Liang X G, Chen R, Zhang Y Y, Zhang K, Wang Z Y. Appl. Spectrosc., 2024, 78(6): 605.

[101]
Han H, Sha R Y, Dai J, Wang Z Z, Mao J W, Cai M. Foods, 2024, 13(7): 1016.

[102]
Li J, Hou X D, Amirkhanian S N, Xiao F P. Prog. Org. Coat., 2023, 182: 107659.

[103]
De Man A, De Souter L, Shi Z Q, Mao C, De Beer T. Anal. Chem., 2024, 96(26): 10586.

[104]
Zhang S J, Zhang S, Gong X C, Qu H B. Process. Biochem., 2024, 145: 50.

[105]
Kandpal L M, Mouazen A M, Masithoh R E, Mishra P, Lohumi S, Cho B K, Lee H. Infrared Phys. Technol., 2022, 127: 104371.

[106]
Froelich N M, Azcarate S M, Goicoechea H C, Campiglia A. Applied Spectroscopy, 2024, 00037028241255150.

[107]
Tian Z H, Wang S Q, Zhang L, Zhang P H, Ye Z F, Zhu Z J, Dong L, Ma W G, Yin W B, Xiao L T, Jia S T., Acta Photon. Sin., 2023, 52(3):0352109.

(田志辉, 王树青, 张雷, 张培华, 叶泽甫, 朱竹军, 董磊, 马维光, 尹王保, 肖连团, 贾锁堂. 光子学报, 2023, 52(3): 0352109.).

[108]
Li X L, Zhang L, Tian Z H, Bai Y, Wang S Q, Han J H, Xia G F, Ma W G, Dong L, Yin W B, Xiao L T, Jia S T. J. Anal. At. Spectrom., 2020, 35(12): 2928.

[109]
Tian Z H, Li J X, Wang S Q, Bai Y, Zhao Y, Zhang L, Zhang P H, Ye Z F, Zhu Z J, Yin W B, Jia S T. J. Anal. At. Spectrom., 2023, 38(7): 1421.

[110]
Gamela R R, Pereira-Filho E R, Pereira F M V. Food Anal. Methods, 2021, 14(3): 545.

[111]
Andrade D F, de Almeida E, de Carvalho H W P, Pereira-Filho E R, Amarasiriwardena D. Talanta, 2021, 225: 122025.

[112]
Wang B B, Song W H, Tian Y, Lu Y, Li Y, Guo J J, Ye W Q, Zheng R E. J. Anal. At. Spectrom., 2023, 38(2): 281.

[113]
Lee Y N, Foster R I, Kim H, Garrett L, Morgan B W, Burger M, Jovanovic I, Choi S. Anal. Chem., 2024, 96(28): 11255.

[114]
Alvarez-Llamas C, Purohit P, Moros J, Laserna J. Anal. Chem., 2022, 94(3): 1840.

[115]
Zhou J Y, Guo L B, Zhang M S, Huang W H, Wang G D, Gong A J, Liu Y C, Sattar H. Anal. Chim. Acta, 2024, 1309: 342674.

[116]
Zhang D, Nie J F, Ma H H, Niu X C, Shi S Q, Chen F, Guo L B, Ji X Y. Anal. Chim. Acta, 2022, 1236: 340552.

[117]
Shin S, Doh I J, Okeyo K, Bae E, Robinson J P, Rajwa B. Molecules, 2023, 28(16): 6087.

[118]
Sushkov N I, Galbács G, Janovszky P, Lobus N V, Labutin T A. Sensors, 2022, 22(21): 8234.

[119]
Ren L H, Tian Y, Yang X Y, Wang Q, Wang L S, Geng X, Wang K Q, Du Z F, Li Y, Lin H. Food Chem., 2023, 400: 134043.

[120]
Lin X M, Zhen X, Lin J J, Huang Y T, Yang J F, Dai P Y, Ren Y K, Ding K. Anal. Lett., 2024: 1.

[121]
Wang Q, Xiao J T, Li Y, Lu Y, Guo J J, Tian Y, Ren L H. Anal. Chim. Acta, 2023, 1240: 340772.

[122]
Yang Z., Li J., Zuo L., Zhao Y., Yu K. J. Anal. At. Spectrom., 2023, 38(10): 2059.

[123]
Xia Z Y, Che X Q, Ye L, Zhao N, Guo D X, Peng Y F, Lin Y Q, Liu X N. Molecules, 2023, 28(4): 1778.

[124]
Chen F, Zhang M S, Huang W H, Sattar H, Guo L B. Foods, 2024, 13(14): 2306.

[125]
Fuentes R, Luarte D, Sandoval C, Myakalwar A K, Alvarez J, Yáñez J, Sbarbaro D. IFAC-PapersOnLine, 2022, 55(21): 85.

[126]
Kandpal L M, Munnaf M A, Cruz C, Mouazen A M. Sensors, 2022, 22(9): 3459.

[127]
Song J H, Shi X Y, Wang H J, Lv X, Zhang W X, Wang J G, Li T S, Li W D. Comput. Electron. Agric., 2024, 218: 108636.

[128]
Gao R, Li J X, Wang S Q, Zhang Y, Zhang L, Ye Z F, Zhu Z J, Yin W B, Jia S T. Anal. Methods, 2023, 15(13): 1674.

[129]
Li J X, Gao R, Zhang Y, Wang S Q, Zhang L, Yin W B, Jia S T. Chemosensors, 2023, 11(7): 363.

[130]
Gao R, Wang S Q, Li J X, Tian Z H, Zhang Y, Zhang L, Ye Z F, Zhu Z J, Yin W B, Jia S T. J. Anal. At. Spectrom., 2023, 38(10): 2046.

[131]
Yan C H, Su Y M, Liu Y J, Zhang T L, Li H. J. Anal. At. Spectrom., 2023, 38(11): 2424.

[132]
Kashani Zadeh H, Hardy M, Sueker M, Li Y C, Tzouchas A, MacKinnon N, Bearman G, Haughey S A, Akhbardeh A, Baek I, Hwang C, Qin J W, Tabb A M, Hellberg R S, Ismail S, Reza H, Vasefi F, Kim M, Tavakolian K, Elliott C T. Sensors, 2023, 23(11): 5149.

[133]
He M, Chen X L, Zhang J, Li J W, Zhao D, Huang Y, Huo D Q, Luo X G, Hou C J. Food Chem., 2023, 400: 134064.

[134]
Gao X, Dong W L, Ying Z H, Li G X, Cheng Q X, Zhao Z J, Li W L. Food Chem., 2024, 460: 140737.

[135]
Zhang H, Liu Z, Zhang J T, Zhang L, Wang S, Wang L, Chen J, Zou C H, Hu J D. Food Anal. Methods, 2021, 14(3): 1.

[136]
Hark R R, Throckmorton C S, Harmon R S, Plumer J R, Harmon K A, Harrison J B, Hendrickx J M H, Clausen J L. Appl. Sci., 2020, 10(23): 8723.

[137]
Tavares T R, Molin J P, Nunes L C, Wei M C F, Krug F J, de Carvalho H W P, Mouazen A M. Agronomy, 2021, 11(6): 1028.

[138]
Di Raimo L A D L, Couto E G, de Mello D C, Demattê J A M, Amorim R S S, Torres G N, Bocuti E D, Veloso G V, Poppiel R R, Francelino M R, Fernandes-Filho E I. Remote. Sens., 2022, 14(19): 4823.

[139]
Gjelsvik E L, Fossen M, Brunsvik A, Liland K H, Tøndel K. Appl. Spectrosc., 2023, 77(10): 1138.

[140]
Zhang Z T, Li Y, Bai L, Chen P, Jiang Y, Qi Y L, Guan H H, Liang Y X, Yuan D P, Lu T L, Yan G J. Microchem. J., 2024, 199: 110126.

[141]
Giannetti V, Mariani M B, Marini F, Torrelli P, Biancolillo A. Microchem. J., 2020, 157: 104896.

[142]
Li Q Q, Zhang C Y, Wang H W, Chen S F, Liu W, Li Y, Li J X. Ind. Crops Prod., 2023, 203: 117127.

[143]
Raeber J, Steuer C. Anal. Chim. Acta, 2023, 1277: 341657.

[144]
Zhang S H, Qi X H, Gao M Y, Dai C J, Yin G H, Ma D Y, Feng W, Guo T C, He L. Food Chem., 2024, 448: 139103.

[145]
Ping J C, Ying Z H, Hao N, Miao P Q, Ye C, Liu C Q, Li W L. Food Res. Int., 2024, 192: 114758.

[146]
You Q, Wang Z Y, Tian X G, Xu X Y. Food Chem., 2023, 425: 136469.

[147]
Sheng X F, Zan J Z, Jiang Y W, Shen S, Li L, Yuan H B. Optik, 2023, 276: 170645.

[148]
Díaz-Romero D, Van den Eynde S, Zaplana I, Zhou C C, Sterkens W, Goedemé T, Peeters J. Resour. Conserv. Recycl., 2023, 190: 106865.

[149]
Fu H, Teng K X, Shen Y F, Zhao J, Qu H B. Spectrochim. Acta Part A Mol. Biomol. Spectrosc., 2024, 305: 123441.

[150]
Bodor Z, Kovacs Z, Benedek C, Hitka G, Behling H. Molecules, 2021, 26(23): 7274.

[151]
Zhang B B, Chen Z Y, Zhang F, Jin G.. Adv. Eng. Sci., 2024, (6): 56.

(张彬彬, 陈祝云, 张飞, 晋刚., 工程科学与技术, 2024, (6): 56.).

[152]
Ibrahim A, Kothari B H, Fahmy R, Hoag S W. AAPS PharmSciTech, 2019, 20: 222.

[153]
Jul-Jørgensen I, Facco P, Gernaey K V, Barolo M, Hundahl C A. Comput. Chem. Eng., 2024, 184: 108647.

[154]
Strani L, Bonacini F, Ferrando A, Perolo A, Daniele, Tanzilli, Vitale R, Cocchi M. Chemical Engineering Transactions, 2023, 100: 175.

[155]
Karami A, Moosavi A A, Pourghasemi H R, Ronaghi A, Ghasemi-Fasaei R, Vidal E, Lado M. Geoderma Reg., 2024, 36: e00752.

[156]
Feng L, Wu B H, Zhu S S, Wang J M, Su Z Z, Liu F, He Y, Zhang C. Front. Plant Sci., 2020, 11: 577063.

[157]
Robert C, Bain W E, Craigie C, Hicks T M, Loeffen M, Fraser-Miller S J, Gordon K C. Meat Sci., 2023, 195: 109005.

[158]
Ferreira D S, Pereira F M V, Olivieri A C, Pereira-Filho E R. Anal. Chim. Acta, 2024, 1303: 342522.

[159]
Ribeiro J V, dos Santos F R, de Oliveira J F, Barbosa G M C, Melquiades F L. Spectrochim. Acta Part B At. Spectrosc., 2024, 211: 106835.

[160]
Tavares T R, Molin J P, Javadi S H, de Carvalho H W P, Mouazen A M. Sensors, 2020, 21(1): 148.

[161]
Dirks M, Turner D, Poole D. Chemom. Intell. Lab. Syst., 2023, 240: 104915.

[162]
Mishra P, Roger J M, Rutledge D N, Biancolillo A, Marini F, Nordon A, Jouan-Rimbaud-Bouveresse D. Chemom. Intell. Lab. Syst., 2020, 205: 104139.

[163]
Chu X L. May New Green Surpass Blue: The Biography of Lu Wanzhen. Shanghai Jiao Tong University Press, 2013.

(禇小立. 新青胜蓝惟所盼—陆婉珍传. 上海交通大学出版社, 2013.).

[164]
Chu X L, Yuan H F, Yang H H, Min Z Q. Selected Handwriting of Academician Lu Wanzhen. Beijing: Chemical Industry Press, 2024.

(褚小立. 陆婉珍院士手迹选. 北京: 化学工业出版社, 2024.).

Options
Outlines

/

〈 〉