Journal of Chemometrics最新文献

The growing complexity and information content of the data, together with the need to understand both the complex structures, relationships, and phenomena present in these data spaces, compounded with the emerging need to understand the results produced by AI tools used to analyze the data, requires development of novel, effective data visualization tools. Much of the growing complexity is reflected in the increasing dimensionality of data spaces, where extended reality (XR) naturally emerges as a candidate to help extend our capability for higher dimensional understanding. However, humans often understand lower dimensionality representations more effectively. Still, XR offers an opportunity for a seamless integration of simulated traditional data displays within the three-dimensional virtual data spaces, leading to more intuitive and more effective data analytics. In this paper we present an overview of the benefits of seamlessly integrated two-dimensional and three-dimensional interactive visual representations embedded in XR spaces, and present three case studies that leverage these approaches for more efficient data analytics.

One of the particularities of spectral imaging when compared to single point spectroscopy is the importance of spectral pretreatments to minimize environmental and sample-related effects, such as shadows, shapes, or variations in illumination. However, the influence of spectral pretreatments on distance metrics is rarely considered in any great depth. This work explores and discusses the effects of combining different spectral pretreatments with the most commonly used distance metrics. A case study on the classification of recyclable materials is used as an example. MATLAB code scripts and functions are provided and referenced through the article to allow the reader to follow and implement the process step by step. The theoretical basis for the calculation and choice of different pretreatments and distance metrics is explained in depth, and the classification performance of the different combinations of pretreatments and distance metrics is discussed in the light of these.

Results show that distance metrics based on angle or correlation (i.e., Spectral Angle Mapper or Spectral Correlation Mapper) could be used with or without pretreatments without significantly impacting the classification results. Classification based on Euclidean and Cityblock distances was the most computationally efficient but was also the most affected by multiplicative effects in the spectra and thus benefited the most from pretreatments such as standard normal variate (SNV) or from combining SNV and second derivative. Lastly, Mahalanobis distance showed the best classification performance for nonpretreated spectra but showed the worst performance for SNV pretreated spectra, illustrating the importance of assessing spectral similarity between calibration and validation datasets on pretreated spectra when using Mahalanobis distance, particularly when applying spectral pretreatments.

This work provides practical insights into the effects that the parameters used have on the results of distance-based classification in terms of the performance of classification models. Considering this issue can greatly improve classification performance when assessing the potential of hyperspectral imaging systems for a particular application.

In the process of spectral modeling, the representativeness of samples to the overall space determines the modeling efficiency. A commonly used unsupervised sample selection method is based on the maximum–minimum distance of selected samples. However, this approach selects samples based solely on a single distance metric, which introduces a degree of randomness. Inspired by the spatial field strength distribution law of multiple point charges, we propose a novel sample selection method based on field strength. In this method, each sample point is treated as a point charge that generates an electric field in its vicinity. The field strength at any given position is the sum of the contributions from all point charges at that location, with a higher field strength indicating that the point is already well represented. By calculating the total field strength exerted by each selected sample on the candidate points and incorporating the point with the minimum field strength into the calibration set, the method maximizes the coverage of field strength in the calibration space. Sequentially selecting and adding points with the lowest field strength yields a highly representative sample set. This approach enables efficient and unsupervised selection of modeling samples.

This paper proposes a novel quantitative modeling and prediction approach for near-infrared (NIR) spectroscopy, combining a stacked target-related autoencoder with an extreme learning machine (STAE-ELM). The STAE performs hierarchical pre-training using multiple improved target-related autoencoders (TAEs) to extract deep spectral features highly correlated with target values. Crucially, the top-level structure of the STAE is replaced by the ELM, which serves as the final prediction model. This integration streamlines training by reducing parameters and steps while simultaneously enhancing performance through optimized initialization of the ELM's weights and biases. Compared to conventional feature selection methods and stacked autoencoders, the STAE-ELM extracts more comprehensive and target-relevant deep features from spectral data, mitigating overfitting risks. The method's efficacy was validated on five open NIR datasets, benchmarking against three approaches: feature selection modeling, SAE-based feature extraction modeling, and backpropagation-based deep network modeling. Results demonstrate that calibration models built with STAE-ELM achieved average reductions in RMSEP of 18.48%, 5.74%, and 12.14%, respectively compared to these benchmarks. Furthermore, modeling efficiency was significantly improved over the backpropagation-based deep network approach.

Vis–NIR spectroscopy is increasingly widely used for soil property analysis due to its rapid, cost-effective, and nondestructive advantages. In particular, deep learning models perform very well when working with large sample data. In this study, we propose a deep learning model based on three-channel ECAM-ConvNeXt. Firstly, the method applies three window functions, Bartlett, Gaussian, and Blackman, in the short-time Fourier transform to convert a one-dimensional spectral sequence signal into three different two-dimensional spectrograms. Next, we perform multichannel feature fusion and use the resulting triple-channel spectrograms as model inputs. This method fully preserves the temporal information and spectral characteristics of the spectral sequence, thereby improving the performance of the model. Secondly, this study introduces the Efficient Channel Attention Module in the ConvNeXt model. This module combines the advantages of the Convolutional Block Attention Module and Efficient Channel Attention Network, further enhancing the expressive ability of the network by highlighting useful information and suppressing irrelevant information. Finally, we also validate the effectiveness of multichannel inputs by deep learning models (AlexNet18, ResNet50, MobileNet-V3, EfficientNet, VIT) and compare them with existing techniques reported in the literature. The results indicate that the root-mean-square error (RMSE) of the TriCH-ECAM-ConvNeXt model in predicting soil nitrogen content (N (g/kg)), organic carbon content (OC (g/kg)), cation exchange capacity (CEC (cmol(+)/kg)), pH, clay content (%), and sand content (%) was reduced to 0.9847, 19.7347, 6.3380, 0.3812, 5.1537, and 12.9706, respectively, and the coefficient of determination (R²) increased to 0.9307, 0.9544, 0.7999, 0.9206, 0.8493, and 0.7526, respectively.

To achieve rapid and comprehensive evaluation of the quality of ginger, a rapid multi-indicator quality evaluation method based on genetic algorithm and near-infrared spectroscopy technology is proposed to detect the content of multiple compounds (6-gingerol, 8-gingerol, 10-gingerol, 6-shogaol, and zingerone). First, the near-infrared spectra of ginger samples is collected. Then, the spectra is preprocessed to reduce the noise. Next, features of the spectra are extracted through the genetic algorithm where the population initialization and fitness function methods are designed. Finally, the prediction model is generated through regression. Experimental results demonstrate that the proposed method achieves higher R values (0.9052, 0.9107, 0.9269, 0.9843, 0.9030) compared to the traditional PLSR model (0.6666, 0.51, 0.4358, 0.9248, 0.4846) for 6-gingerol, 8-gingerol, 10-gingerol, 6-shogaol, and zingerone, respectively. Therefore, the proposed method can reduce prediction errors and improve the performance of near-infrared spectroscopy quantitative analysis model for ginger.

Rapid, noninvasive quantification of canopy nitrogen (N) and chlorophyll (Chl) content is critical for precision nitrogen management in maize cultivation. Although near-infrared spectroscopy (near-infrared spectroscopy, NIRS) offers a viable approach for biochemical component analysis, conventional machine learning models often fail to capture the complex nonlinear relationships inherent in spectral data and lack interpretability, limiting their robustness for real-time inversion tasks. To address these limitations, this study introduces a hybrid deep learning architecture combining convolutional neural networks (CNNs) and gated recurrent units (GRUs), augmented by a convolutional block attention module (CBAM), integrated with explainable artificial intelligence for accurate biochemical content inversion. Preprocessing of hyperspectral images from 200 maize canopy samples via sequential Savitzky–Golay smoothing (SG), standard normal variate (SG-SNV), and SG transformations enhanced mean test set R² by 0.016 units. Subsequent dimensionality reduction via the successive projection algorithm (SPA) and competitive adaptive reweighting sampling (CARS) significantly reduced spectral features from 176 to 10 and 22 bands, respectively. The core predictive model synergistically combines CNNs and GRUs, augmented by a CBAM to enhance feature extraction and temporal dependency modeling. Comparative evaluation demonstrates the superior performance of CNN-GRU-CBAM over traditional machine learning and alternative deep learning models. For the test set, it achieved R² values of 0.934 (N) and 0.788 (Chl), with corresponding root mean square error (RMSE) values of 1.940 and 0.216. Model interpretability was rigorously validated using Shapley Additive Explanations (SHAP), identifying key spectral regions driving predictions. This work innovatively bridges high-performance deep learning with explainable artificial intelligence, enabling precise, nondestructive estimation of maize foliar biochemical constituents. The framework provides a transferable approach for biochemical content inversion in diverse crops.