**4. Conclusions**

As an important grain crop, maize is susceptible to mold infection during growth and storage due to its large embryo area and high moisture content. Therefore, it is essential to distinguish the moldy sample from healthy groups to prevent the spread of mold and avoid huger economic losses. 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. In this study, the hyperspectral images of maize with different moldy levels were collected within Vis-SWNIR and LWNIR regions, and the spectra and texture information were extracted and used to establish the classification model with the methods of pixel-level and feature-level fusions. The results showed that data fusion strategies at both levels achieved better classification results than spectra alone. For pixel-level data fusion of spectral and texture information, the energy and contrast achieved positive effect on improving the classification model, with prediction accuracy of 90% and 90% for Vis-SWNIR and LWNIR regions, respectively. The improvement in model detection accuracy is not very apparent, as some irrelevant variables are introduced along with useful information. For featurelevel data fusion of spectral and texture information, the variables selected by VCPA significantly increase the classification accuracy, with prediction accuracy of 93.33% and 90% for Vis-SWNIR and LWNIR regions, respectively. Feature-level fusion models based on the key variable combination of two hyperspectral systems were best for the classification of maize with different moldy levels, with an overall prediction accuracy of 95.00% for each moldy level.

This paper mainly focused on the identify of moldy maize; it should be noted that the data fusion strategies presented in this study are generally suitable to the quality detection of other grain crops such as wheat, rice, and peanut. Although this study had shown that the great feasibility of using hyperspectral imaging technology and multi-source data fusion method to discriminate the maize with different moldy levels, there must be some differences between naturally and artificially moldy maize samples, and we will use the data fusion strategies to classify moldy maize under natural growth in our future work. Additionally, AFB1 is a metabolite of mold with high toxicity; therefore, the growth monitoring of mold plays an important role in the early warning of AFB1 pollution in maize. CAT is a precursor product synthesized by AFB1, its dynamic activity reflects the level of mold activity, and it has a significant relationship with the content of AFB1. Therefore, we will develop a model for warning the AFB1 contamination based on the relationship between CAT of mold and AFB1 under different moldy conditions, which provides theoretical basis and technical guarantee for safe storage of maize.

**Author Contributions:** Conceptualization, W.W. and X.T.; software, W.H.; validation, W.W. and W.H.; data curation, W.W. and X.T.; writing—original draft preparation, W.W.; writing—review and editing, W.W., H.Y. and X.T.; project administration, H.Y. and X.T.; funding acquisition, H.Y. and X.T. All authors have read and agreed to the published version of the manuscript.

**Funding:** This study was supported by National Natural Science Foundation of China (Grant No. 31901402) and Young Elite Scientists Sponsorship Program by CAST (Grant No. 2019QNRC001).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** The data that support the findings of this study are available upon request from the authors.

**Conflicts of Interest:** The authors declare no conflict of interest.
