**1. Introduction**

Maize is an important food crop, feed crop, and cash crop [1]. Compared with wheat and rice, maize has larger embryos, and its moisture content at harvest reaches about 30%, higher than the 25% and 22.5% of wheat and rice, respectively, which makes it more susceptible to mold infection during growth and storage [2]. Among them, *Aspergillus flavus* is the strain that most easily and commonly infects maize; aflatoxin (AFB1), produced by *Aspergillus flavus*, is extremely carcinogenic and toxic, and is the most toxic mold secondary metabolite in contaminated food [3,4]. Humans and animals eating food contaminated with AFB1 is a serious threat to life and health safety. Therefore, the early real-time detection of moldy maize has very important research significance.

Traditional detection methods for moldy maize include sensory evaluation and physical and chemical component detection [5]. Sensory evaluation is simple, time-saving, and low cost, but the evaluation results are easily disrupted by the external environment and the subjective emotions of the inspectors. Additionally, the toxic substances will also pose a threat to the health of the inspectors. The physical and chemical component determination is generally detected by high-performance liquid chromatography (HPLC), polymerase chain reaction (PCR), and enzyme-linked immunosorbent assay (ELISA) [6–8]. Although these methods can achieve more accurate measurement and qualitative analysis, they require expensive testing equipment and professional technicians, and the testing process

**Citation:** Wang, W.; Huang, W.; Yu, H.; Tian, X. Identification of Maize with Different Moldy Levels Based on Catalase Activity and Data Fusion of Hyperspectral Images. *Foods* **2022**, *11*, 1727. https://doi.org/10.3390/ foods11121727

Academic Editor: Corrado Costa

Received: 6 May 2022 Accepted: 10 June 2022 Published: 13 June 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

is complex and time-consuming [9]. Hence, using traditional methods, it is difficult to achieve simple, rapid, and non-destructive detection, which cannot meet the actual needs for identification of kernel maize.

In recent years, non-destructive detection technologies, such as electronic nose, machine vision, near-infrared, and hyperspectral imaging, have been successfully applied to the classification of moldy maize [10,11]. Electronic nose technology [12] is mainly used specific sensors to identify the level of mold based on the change of volatile organic compounds (VOCs) information. Leggieri et al. [13] used electronic nose technology to determine the concentration of AFB1 and fumonisins (FBs) in maize; the prediction model of AFB1 and FBs built by the artificial neural networks were 78% and 77%, respectively. However, the concentration of gas is easy to change in the flow state, which affected the discrimination accuracy. Machine vision detection mainly adopts machine learning algorithms to extract features from kernel images and then establish the classification model based on extracted features [14]. Shi Ying [15] extracted R-channel eigenvalues of RGB images (red, green and blue three-channel color image) and classified maize kernel samples with different levels of mold using the Back Propagation (BP) neural network. Visible-near infrared (Vis-NIR) spectroscopy technology connects the spectral information with the internal content of substances and uses the spectral curve to analyze the changes of internal components of seeds in the process of mold growth [16]. Therefore, the machine vision and Vis-NIR spectroscopy can express the external and internal changes of the target samples respectively. However, both internal quality and external characteristics of maize will change during the moldy process; neither machine vision nor Vis-NIR spectroscopy can obtain internal and external quality information at the same time. Hyperspectral imaging technology combines spectral analysis technology with image processing technology, which can simultaneously obtain the spectral data with internal component information and the image data with appearance feature information, realizing the rapid, pollution-free and non-destructive detection [17–19].

In terms of using hyperspectral imaging technology to identify moldy maize, Tao et al. [20] used random frog (RF) combined with partial least-squares discriminant analysis (PLS-DA) to qualitatively analyze the healthy maize and polluted maize inoculated with aflatoxigenic fungus at different culture days based on long wave near-infrared (LWNIR) hyperspectral images; the classification accuracy of the calibration set and verification set was 82.3% and 94.9%, respectively. Williams et al. [21] evaluated the fungal development in maize kernels using LWNIR; principal component analysis (PCA) was firstly used to remove the interference of noise, such as background, bad pixels, and shadows, from the hyperspectral images. Three distinct clusters related to the degree of infection were found in the scoring plots of PC4 and PC5. Dai et al. [22] established a classification model of moldy maize with different culture days (0 days, 2 days, 4 days, 6 days, and 8 days) based on 9 characteristic wavelengths selected from visible and short wave near-infrared (Vis-SWNIR) hyperspectral imaging using fisher discriminant analysis (FDA); the classification accuracy of the calibration set and validation set were 100% and 98.67%, respectively, illustrating that the characteristic wavelengths could represent the main information about moldy levels of maize samples. Del Fiore et al. [23] used Vis-SWNIR hyperspectral imaging combined with multivariate statistical analysis to identify maize kernels infected with fungi under different growths. The results showed that hyperspectral imaging was able to quickly distinguish between healthy and infected maize, i.e., 48 h after inoculation with mycorrhizal fungi. Previous studies have shown that both the spectral ranges of Vis-SWNIR and LWNIR can be used to distinguish the moldy level of maize; however, it has not been found that fusing the spectral information of different hyperspectral systems can construct a classification model of maize with different moldy levels. Yu et al. [24] studied the influence of Vis-SWNIR and LWNIR hyperspectral imaging systems on the prediction ability of total volatile basic nitrogen (TVB-N) content in tilapia fillets during refrigeration; the results showed that the fused spectral data of both sensors achieved a better prediction result than that of individual sensor. Meanwhile, the study of fusing

spectral data with texture data to discriminate the moldy level of maize is less extensive. Ma et al. [25] developed the classification model of fresh and frozen meats based on the spectral and texture information extracted from Vis-SWNIR hyperspectral images; the research showed that the classification model built by the feature fusion of spectra and texture was better than that of spectra and texture alone. Therefore, fusing the information of spectra and texture obtained from different hyperspectral image systems would be a new idea for constructing an accurate classification model of maize with different moldy levels.

Mold growth is uncontrollable. Hence, moldy levels may not be uniform among different samples at the same culture time and a small number of samples were not consistent with the designed moldy levels. Catalase (CAT) is a metabolite in the growth of *Aspergillus flavus* and other microorganisms [26]. Zhang et al. [27] found that the correlation coefficient between the number of mold colonies and the activity value of CAT reached more than 0.9 in various grains such as wheat, rice, and maize. In addition, CAT is the precursor product of AFB1 produced by *Aspergillus flavus*. Zhang et al. [28] analyzed the correlation between CAT activity value and AFB1 content of moldy maize and found that both of them had the same change curves. The above research showed that CAT activity value could reflect the moldy levels of maize. However, at present, it is not found that the model for discrimination of maize with different moldy levels was established based on the feature fusion of hyperspectral imaging information and CAT activity value of moldy maize.

In this study, we proposed a new method to better divide the maize with different moldy levels by monitoring the CAT activity value of maize samples infected with *Aspergillus flavus* under different culture days. The objective of this study is to examine the potential of using multi-levels data fusion of hyperspectral images to identify the maize with different moldy levels. The specific objectives of this study were to: (1) analyze the difference of spectra and texture of Vis-SWNIR and LWNIR hyperspectral images of maize samples with different moldy levels; (2) examine the ability of different pretreatment methods and classifier for identification of maize samples with different moldy levels; (3) compare the classification ability of the models based on the pixel-level fusion of spectra and different texture parameters; (4) evaluate the effects of features selected by different variable selection methods on the classification models of feature-level fusion; and (5) establish the best classification model of maize with different moldy levels by integrating the spectrum and texture data with Vis-SWNIR and LWNIR regions.
