*2.6. Data Fusion*

Data fusion was a process of combining information from different independent information sources, which could express the described objects or processes in more detail and complete than using a single information source alone. Generally, data fusion was divided into pixel-level fusion, feature-level fusion, and decision-level fusion according to the fusion level from low to high [36]. In the study, the spectral data and texture data obtained from hyperspectral images within Vis-SWNIR and LWNIR regions were fused at pixel-level and feature-level respectively for developing a high accuracy and robust classification model of moldy maize.

Pixel-level fusion was simply merging the data information of different sources [37], so the fused features contained more variables, which was conducive to further data processing. However, it could also input the irrelevant and redundant variables into the model. In this study, pixel-level fusion models were built by fusing the spectra matrices with texture parameters, for Vis-SWNIR and LWNIR regions. For the new matrix formed after data fusion, one row represented the characteristic information of the same sample, and one column represented the eigenvalues at a specific wavelength. The optimal combination of spectrum and texture parameters was obtained by evaluating the classification accuracy of the developed models.

Feature-level fusion was to extract features from a single data block using the variable selection method and then integrate the processed feature matrix [38,39]. Compared with pixel-level fusion, feature-level fusion could adjust the number of features from different data, especially when there were large differences between single data blocks. In this study, both spectral matrix and texture matrix obtained from Vis-SWNIR and LWNIR ranges had 389 and 112 variables, respectively, these data were often multicollinearity and redundant, especially between adjacent bands. Therefore, feature wavelength selection was commonly adopted to select the key wavelengths from full-band data, which could reduce redundant and noisy information, as well as simplify the model. The optimal combination of spectral data and texture parameters in pixel-level fusion was used as the data source of feature-level fusion. Three kinds of variable selection methods including variable combination population analysis (VCPA) [40], iteratively retains informative variables (IRIV) [41], and hybrid method mVCPA-IRIV [42] were used to select the features that carry the information of moldy maize from the spectral and texture parameters data, and then combined them into a new data matrix to build the feature-level model. The number of columns in the new data matrix was the number of features obtained from the two data.
