Journal of Chemometrics最新文献

Reducing feature dimensionality in spectroscopic data is crucial for efficient analysis and classification. Using all available features for classification typically results in an unacceptably high runtime and poor accuracy. Popular feature extraction methods, such as principal component analysis (PCA), linear discriminant analysis (LDA), and autoencoders, reduce feature dimensionality by extracting latent features that can be challenging to interpret. To enable better human interpretation of the classification model, we avoid extraction methods and instead propose applying feature selection methods. In this work, we develop an innovative PCA-based feature selection method for spectroscopic data, providing an essential subset of the original features. As an important advantage, no prior knowledge about the characteristic signals of the respective target substance is required. In this proof-of-concept study, the proposed method is initially characterized using simulated Raman and infrared absorption datasets. From the top five PCA eigenvectors of spectroscopic data, we identify a set of three top peaks each at specific wavenumbers (features). The compact set of selected features is then used for classification tasks applying a decision tree. Based on two well-defined spectroscopic datasets, our study demonstrates that our new method of PCA-based peak finding outperforms selected other approaches with regard to interpretability and accuracy. For both investigated datasets, accuracies greater than 97% are achieved. Our approach shows large potential for accurate classification combined with interpretability in further scenarios involving spectroscopic datasets.

The first article in this column was published in 2014 [1], over 10 years ago. In this intermission, we look at the importance of chemometrics to analytical chemists and the central role of PLS (Partial Least Squares), which will be the subject of the next articles.

The name chemometrics (in Swedish) was first advocated in the literature in 1972 [2] but was far from the modern day emphasis on multivariate data and primarily involved univariate curve fitting. Although a few enthusiasts communicated together during the 1970s, it was not until the early 1980s that the subject we now recognise became organised.

Most of the major building blocks were developed in the 1980s, and we will look at how the subject fitted into analytical chemistry as from 1980. A good way to study the impact is via citations. We will use Web of Science (Clarivate) with a census date of 24 July 2025 to chart the progress of major publications in this area. We look at some of the more prominent relevant journals.

Analytica Chimica Acta is one of the first general analytical journals to publish a significant number of papers in chemometrics. The classic paper by Geladi and Kowalski on PLS (Partial Least Squares) [3] in 1986 is the most cited article ever in this journal since 1980 (5746 cites). The second most cited (2483 cites) is on a subject related to chemometrics (Box–Behnken designs), suggesting the dominant impact of chemometrics in this forum. What is remarkable is the longevity of this article. Most articles decline in importance after a few years. In Figure 1, we present the citations to this article as compared to all papers together in ACA since 1986. So not only is this, by a long way, the most highly cited article ever published in this journal, its importance is increasing with time over nearly a 40-year period, compared to general articles published in 1986. This is even clearer when we look at the percentage of citations to all papers in ACA published in 1986 over time, in Figure 2. By 2024, over 80% of all citations are to this one paper. This strongly suggests that PLS continues to remain a long-term advance in analytical chemistry over four decades.

In Elsevier's Chemometrics and Intelligent Laboratory Systems, which first published in 1986, the two most cited papers by a wide margin are about PCA (9016 cites) [4] and, again, PLS (7652 cites) [5]. However, as the former paper was published in 1987 and the latter in 2001, in fact, the interest in PLS has been greater than PCA.

The journal Analytical Chemistry, although it published some early reviews in chemometrics, does not prominently feature chemometrics articles. In the decade 1980–1989, the third most cited article was focused on chemometrics and involved PLS [6] suggesting the importance of PLS to chemometrics in its early days (2434 cites). The most cited article (7820 c

O.-C. G. Engstrøm, M. Albano-Gaglio, E. S. Dreier, et al., “ Transforming Hyperspectral Images Into Chemical Maps: A Novel End-to-End Deep Learning Approach,” Journal of Chemometrics 39, no. 8 (2025): e70041, https://doi.org/10.1002/cem.70041.

In the original work, there was an apparent discrepancy between two ways of making predictions with the same PLS model. One way was predicting directly on mean spectra. The other was predicting on pixel-wise spectra and subsequently averaging the result. The discrepancy between these two ways of evaluation was solely due to the order of application of preprocessing.

In the original work, the mean spectrum was computed by taking the negative logarithm of each pixel (to transform from reflectance to pseudo absorbance), averaging all pixels, applying SNV (Standard Normal Variate) transform, and finally convolution with an SG (Savitzky–Golay) filter. However, when applying pixel-wise PLS, SNV and SG were applied to each pixel individually. As SNV is nonlinear (the standard deviation is not linear), it matters whether it is applied before or after averaging. So, this update now applies preprocessing to each pixel individually before taking the average to compute a mean spectrum. This removes any discrepancy between the results obtained by pixel-wise PLS and PLS on mean spectra. Any reference to this discrepancy has been removed from the work and all results and figures have been updated accordingly. Appendix D describes the updates in detail.

We apologize for this error.

Interactions that happen in the ingredients during the industrial process of food products, mainly in starch–protein systems, interfere directly with their typical characteristics. To investigate the effects of the individual components on textural properties blends with different soy protein isolate/gluten/starch ratios were studied. Three botanical sources of starch were also investigated by using wheat, corn, and cassava starches. The flowability, penetration force, rupture force, deformability, and hardness values of the gels, determined by texture analyzer under controlled conditions, were observed and allowed comparative study. Through chemometrics, linear, quadratic, and cubic fit models were applied to obtain polynomial equations that could adequately express the observed response surfaces for each evaluated textural parameter. The statistical significance of these models was evaluated through analysis of variance and a new experimental test, making it possible to indicate the best-fit model for each textural property studied. This study indicated that there are large correlation differences between the textural properties. The effect of soy protein isolate was greater than that of gluten and starch and a similarity of the behavior was observed from corn and wheat starches.

Industrial data have always played a pivotal role in the development, management, and optimization of processes and products. Over time, the intrinsic characteristics of data (modality, acquisition frequency, granularity, uncertainty, etc.) have evolved significantly and so have the computational methods and technology to process and integrate them, transforming these inputs into valuable insights. By integrating multiple heterogeneous data sources, data fusion provides an updated and improved status of the operations, enhancing information quality and leading to more comprehensive, accurate, and insightful analysis of the systems under examination and, therefore, better decision-making. In this work, we present a review of the data fusion methodologies from the perspective of practitioners, including engineers and data analysts, and how it can be systematically organized into classes of methods dedicated to achieving specific goals. The terminology adopted in the field, not always used consistently and without ambiguity, is also discussed and clarified. Additionally, several case studies are provided to showcase some of the applications and potential advantages of adopting data fusion frameworks.