Pixel-level aflatoxin detecting in maize based on feature selection and hyperspectral imaging

Advancing detection of fungal and mycotoxins contamination in grains and oilseeds: Hyperspectral imaging for enhanced food safety

Grains and oilseeds, including maize, wheat, and peanuts, are essential for human and animal nutrition but are vulnerable to contamination by fungi and their toxic metabolites, mycotoxins. This review provides a comprehensive investigation of the applications of hyperspectral imaging (HSI) technologies for the detection of fungal and mycotoxins contamination in grains and oilseeds. It explores the capability of HSI to identify specific spectral features of contamination and emphasized the critical role of sample properties and sample preparation techniques in HSI applications. Additionally, it reveals the challenges posed by the voluminous HSI data generated and discusses the application of sophisticated data processing techniques, including chemometrics methods and machine learning algorithms. The review highlights future research directions focused on refining HSI applications for practical use. Ultimately, this review underscores the potential of integrating HSI with advanced technologies to significantly enhance food safety and quality assurance.

Novel decision-level fusion strategies combined with hyperspectral imaging for the detection of soybean protein content

Soybeans are used for human consumption or animal feed due to their abundant protein content. In this study, visible-near infrared (VNIR) hyperspectral imaging (HSI) and short-wave infrared HSI combined with three-levels data fusion methods were employed to detect the protein content of soybean seeds, including measurement-level fusion, feature-level fusion, and decision-level fusion. Additionally, three novel decision-level fusion methods were proposed, including binary linear regression, feature-based multiple linear regression (MLR), and model-based MLR. An IVISSA-SPA-MLR model based on decision-level fusion demonstrated the best predictive performance, with a residual prediction deviation value of 3.6796. The results suggested that the IVISSA-SPA-MLR achieved accurate predictions, effectively enabling precise detection of soybean seeds protein content. Decision-level fusion proved to be an accurate and efficient quantitative detection technique, enhancing the predictive performance of regression models. This research provides a novel method for protein content detection in food products and introduces new strategies for data fusion.

Hybrid wavelength selection strategy combined with ATR-FTIR spectroscopy for preliminary exploration of vintage labeling traceability of sauce-flavor baijiu

The allure of substantial profits has perpetuated the illicit trade of counterfeit vintage labels for baijiu. While various approaches have been employed to intelligently ascertain the vintage of baijiu, many of them are both cost-intensive and time-consuming. This work pioneered the use of Attenuated total reflection Fourier transform infrared (ATR-FTIR) spectroscopy, coupled with chemometric analysis, offering a non-destructive and economically viable method for discriminating sauce-flavor baijiu across different aging periods (1-, 2-, and 3-year). In this research, principal component analysis (PCA) was first conducted to explore clustering trends among distinct vintage groups. Subsequently, the effect of spectral pre-processing on modeling performance was explored. For wavelength selection, four wavelength selection methods (ReliefF, random forest variable importance (RFVI), variable importance in projection (VIP), and Venn) were first used to identify the subset of candidate features that potentially best mapped the vintage labels. Immediately following this, to explore the possibility of further improving the identification capabilities of the model as well as to reduce the redundant data that may still be present, sequential backward selection (SBS) was utilized for secondary feature reduction within the subset of candidates. The amalgamation of these two techniques is termed a "hybrid wavelength selection strategy." Additionally, the dimensionality reduction effects of PCA and kernel principal component analysis (KPCA) were compared to demonstrate the robustness of the proposed method. Finally, classification models such as partial least squares discriminant analysis (PLS-DA), random forest (RF), and grasshopper optimization algorithm-based support vector machine (GOA-SVM) were developed. The results show that the spectral data need not be pre-processed, and the proposed hybrid wavelength selection strategy can further improve the identification ability of the model. Among the many models developed, ReliefF-SBS-GOA-SVM emerged as the most proficient classification model, yielding accuracy, sensitivity, and specificity rates of 94.44%, 95.23%, and 94.44%, respectively. This method not only holds promise for the discrimination of baijiu class attributes such as brand, origin, flavor, and vintage but also exhibits potential applicability in other non-targeted identification studies involving spectroscopy methodologies.

Pixel-Level Recognition of Trace Mycotoxins in Red Ginseng Based on Hyperspectral Imaging Combined with 1DCNN-Residual-BiLSTM-Attention Model

Red ginseng is widely used in food and pharmaceuticals due to its significant nutritional value. However, during the processing and storage of red ginseng, it is susceptible to grow mold and produce mycotoxins, generating security issues. This study proposes a novel approach using hyperspectral imaging technology and a 1D-convolutional neural network-residual-bidirectional-long short-term memory attention mechanism (1DCNN-ResBiLSTM-Attention) for pixel-level mycotoxin recognition in red ginseng. The "Red Ginseng-Mycotoxin" (R-M) dataset is established, and optimal parameters for 1D-CNN, residual bidirectional long short-term memory (ResBiLSTM), and 1DCNN-ResBiLSTM-Attention models are determined. The models achieved testing accuracies of 98.75%, 99.03%, and 99.17%, respectively. To simulate real detection scenarios with potential interfering impurities during the sampling process, a "Red Ginseng-Mycotoxin-Interfering Impurities" (R-M-I) dataset was created. The testing accuracy of the 1DCNN-ResBiLSTM-Attention model reached 96.39%, and it successfully predicted pixel-wise classification for other unknown samples. This study introduces a novel method for real-time mycotoxin monitoring in traditional Chinese medicine, with important implications for the on-site quality control of herbal materials.

Qualitative and quantitative detection of aflatoxins B1 in maize kernels with fluorescence hyperspectral imaging based on the combination method of boosting and stacking

The most widespread, toxic, and harmful toxin is aflatoxins B1 (AFB1). The fluorescence hyperspectral imaging (HSI) system was employed for AFB1 detection in this study. This study developed the under sampling stacking (USS) algorithm for imbalanced data. The results indicated that the USS method combined with ANOVA for featured wavelength achieved the best performance with the accuracy of 0.98 for 20 or 50 μg /kg threshold using endosperm side spectra. As for the quantitative analysis, a specified function was used to compress AFB1 content, and the combination of boosting and stacking was used for regression. The support vector regression (SVR)-Boosting, Adaptive Boosting (AdaBoost), and extremely randomized trees (Extra-Trees)-Boosting were used as the base learner, while the K nearest neighbors (KNN) algorithm was used as the meta learner could obtain the best results, with the correlation coefficient of prediction (R_p) was 0.86. These results provided the basis for developing AFB1 detection and estimation technologies.

Quantitative Monitoring of Leaf Area Index in Rice Based on Hyperspectral Feature Bands and Ridge Regression Algorithm

Leaf area index (LAI) is one of the indicators measuring the growth of rice in the field. LAI monitoring plays an important role in ensuring the stable increase of grain yield. In this study, the canopy reflectance spectrum of rice was obtained by ASD at the elongation, booting, heading and post-flowering stages of rice, and the correlations between the original reflectance (OR), first-derivative transformation (FD), reciprocal transformation (1/R), and logarithmic transformation (LOG) with LAI were analyzed. Characteristic bands of spectral data were then selected based on the successive projections algorithm (SPA) and Pearson correlation. Moreover, ridge regression (RR), partial least squares (PLS), and multivariate stepwise regression (MSR) were conducted to establish estimation models based on characteristic bands and vegetation indices. The research results showed that the correlation between canopy spectrum and LAI was significantly improved after FD transformation. Modeling using SPA to select FD characteristic bands performed better than using Pearson correlation. The optimal modeling combination was FD-SPA-VI-RR, with the coefficient of determination (R2) of 0.807 and the root-mean-square error (RMSE) of 0.794 for the training set, R2 of 0.878 and RMSE of 0.773 for the validation set 1, and R2 of 0.705 and RMSE of 1.026 for the validation set 2. The results indicated that the present model may predict the rice LAI accurately, meeting the requirements of large-scale statistical monitoring of rice growth indicators in the field.

Quantitative detection of Aflatoxin B1 by subpixel CNN regression

Aflatoxin is a highly toxic substance dispersed in peanuts, which seriously harms the health of humans and animals. In this paper, we propose a new method for aflatoxin B1(AFB1) detection inspired by quantitative remote sensing. Firstly, we obtained the relative content of AFB1 at the sub-pixel level by subpixel decomposition (endmember extraction, nonnegative matrix decomposition). Then we modified the transfer learning models (LeNet5, AlexNet, VGG16, and ResNet18) to construct a deep learning regression network for quantitative detection of AFB1. There are 67,178 pixels used for training and 67,164 pixels used for testing. After subpixel decomposition, each aflatoxin pixel was determined to contain content, and each pixel had 400 hyperspectral values (415-799 nm). The experimental results showed that, among the four models, the modified ResNet18 model achieved the best effect, with R² of 0.8898, RMSE of 0.0138, and RPD of 2.8851. Here, we implemented a sub-pixel model for quantitative AFB1 detection and proposed a regression method based on deep learning. Meanwhile, the modified convolution classification model has high predictive ability and robustness. This method provides a new scheme in designing the sorting machine and has practical value.

Identification of Typical Solid Hazardous Chemicals Based on Hyperspectral Imaging

The identification of hazardous chemicals based on hyperspectral imaging is an important emergent means for the prevention of explosion accidents and the early warning of secondary hazards. In this study, we used a combination of spectral curve matching based on full-waveform characteristics and spectral matching based on spectral characteristics to identify the hazardous chemicals, and proposed a method to quantitatively characterize the matching degree of the spectral curves of hazardous chemicals. The results showed that the four hazardous chemicals, sulfur, red phosphorus, potassium permanganate, and corn starch had bright colors, distinct spectral curve characteristics, and obvious changes in reflectivity, which were easy to identify. Moreover, the matching degree of their spectral curves was positively correlated with their reflectivity. However, the spectral characteristics of carbon powder, strontium nitrate, wheat starch, and magnesium–aluminum alloy powder were not obvious, with no obvious characteristic peaks or trends of change in reflectivity. Except for the reflectivity and the matching degree of the carbon powder being maintained at a low level, the reflectivity of the remaining three samples was relatively close, so that it was difficult to identify with the spectral curves alone, and color information should be considered for further identification.

Classification of Aflatoxin B1 Concentration of Single Maize Kernel Based on Near-Infrared Hyperspectral Imaging and Feature Selection

A rapid and nondestructive method is greatly important for the classification of aflatoxin B1 (AFB1) concentration of single maize kernel to satisfy the ever-growing needs of consumers for food safety. A novel method for classification of AFB1 concentration of single maize kernel was developed on the basis of the near-infrared (NIR) hyperspectral imaging (1100–2000 nm). Four groups of AFB1 samples with different concentrations (10, 20, 50, and 100 ppb) and one group of control samples were prepared, which were preprocessed with Savitzky–Golay (SG) smoothing and first derivative (FD) algorithms for their raw NIR spectra. A key wavelength selection method, combining the variance and order of average spectral intensity, was proposed on the basis of pretreated spectra. Moreover, principal component analysis (PCA) was conducted to reduce the dimensionality of hyperspectral data. Finally, a classification model for AFB1 concentrations was developed through linear discriminant analysis (LDA), combined with five key wavelengths and the first three PCs. The results show that the proposed method achieved an ideal performance for classifying AFB1 concentrations in a single maize kernel with overall accuracy, with an F1-score and Kappa values of 95.56%, 0.9554, and 0.9444, respectively, as well as the test accuracy yield of 88.67% for independent validation samples. The combinations of variance and order of average spectral intensity can be used for key wavelength selection which, combined with PCA, can achieve an ideal dimensionality reduction effect for model development. The findings of this study have positive significance for the classification of AFB1 concentration of maize kernels.

Aflatoxin rapid detection based on hyperspectral with 1D-convolution neural network in the pixel level

Aflatoxin is commonly exists in moldy foods, it is classified as a class one carcinogen by the World Health Organization. In this paper, we used one dimensional convolution neural network (1D-CNN) to classify whether a pixel contains aflatoxin. Firstly we found the best combination of 1D-CNN parameters were epoch = 30, learning rate = 0.00005 and 'relu' for active function, the highest test accuracy reached 96.35% for peanut, 92.11% for maize and 94.64% for mix data. Then we compared 1D-CNN with feature selection and methods in other papers, result shows that neural network has greatly improved the detection efficiency than feature selection. Finally we visualized the classification result of different training 1D-CNN networks. This research provides the core algorithm for the intelligent sorter with aflatoxin detection function, which is of positive significance for grain processing and the prenatal detoxification of foreign trade enterprises.

Machine learning techniques for analysis of hyperspectral images to determine quality of food products: A review

Non-destructive testing techniques have gained importance in monitoring food quality over the years. Hyperspectral imaging is one of the important non-destructive quality testing techniques which provides both spatial and spectral information. Advancement in machine learning techniques for rapid analysis with higher classification accuracy have improved the potential of using this technique for food applications. This paper provides an overview of the application of different machine learning techniques in analysis of hyperspectral images for determination of food quality. It covers the principle underlying hyperspectral imaging, the advantages, and the limitations of each machine learning technique. The machine learning techniques exhibited rapid analysis of hyperspectral images of food products with high accuracy thereby enabling robust classification or regression models. The selection of effective wavelengths from the hyperspectral data is of paramount importance since it greatly reduces the computational load and time which enhances the scope for real time applications. Due to the feature learning nature of deep learning, it is one of the most promising and powerful techniques for real time applications. However, the field of deep learning is relatively new and need further research for its full utilization. Similarly, lifelong machine learning paves the way for real time HSI applications but needs further research to incorporate the seasonal variations in food quality. Further, the research gaps in machine learning techniques for hyperspectral image analysis, and the prospects are discussed.

•

Artificial neural network has been intensively used for Hyperspectral image (HSI) analysis.

•

Support vector machines and random forest techniques are gaining momentum for HSI analysis.

•

Deep learning applications has potential for implementation in real time HSI analysis.

•

Lifelong machine learning needs further research to incorporate the seasonal variations in food quality.

Hyperspectral imaging coupled with multivariate methods for seed vitality estimation and forecast for Quercus variabilis

In this study, the feasibility of estimation and forecast of different vitality Quercus variabilis seeds by a hyperspectral imaging technique were investigated. Artificially accelerated aging was conducive to achieve the division of four vitality levels. Hyperspectral data in the first 10 h of germination were continuously collected at one-hour intervals. The optimal band was selected for the original and pre-processed spectra which were treated by multiple scatter correction (MSC) and the Savitzky-Golay first derivative (SG 1st). Five characteristic wavelength methods were compared: successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS), genetic algorithm (GA), variable important in projection (VIP), and random frog (RF). Partial least square-discriminant analysis (PLS-DA) and K-nearest neighbor (KNN) built the vitality estimation model based on different data sets, and GA + PLS-DA constructed the optimal model with the highest accuracy. According to the weight coefficient and reflectance of the characteristic band extracted by the GA, the reflectance curves of different levels over time were plotted. The data of 0 h was employed to establish the vitality forecast model. The forecast model had a high recognition rate, with PLS-DA exceeding 99% and KNN exceeding 85%. This indicated that hyperspectral imaging of seed germination processes could achieve non-destructive estimation of Q. variabilis seed vitality, and accurate prediction in a shorter time is feasible.