2.4.1. Construction of Texture Indices

Hyperspectral image has a gray image in each of its spectral bands, so there is a large amount of redundant information. Based on the gray-level co-occurrence matrix (GLCM) of the gray image in each selected spectral band, eight texture features were acquired for each gray image [32], as listed in Table 1. In addition to the texture features, a new texture-based index called the normalized difference texture index (NDTI) was proposed, which follows the conventional definition of normalized difference vegetation index. A new NDTI was constructed with all possible two-texture feature combinations at all sensitive wavelengths with eight GLCM-based texture features of each gray image. The texture index NDTI has been used to classify sun and shade leaves of crops and plant biomass estimation [33]. For example, six wavelengths were selected in this research with 8 texture features at each wavelength, thus 48 texture features were acquired. NDTI can be calculated with the combination of two different texture features from all 48 texture features. A total of 2256 combination pairs were obtained, and the specific formula used for NDTI calculation was NDTI = (T1 − T2)/(T1 + T2), where T1 and T2 are the texture features of the selected sensitive wavelength.


**Table 1.** Texture features based on gray-level co-occurrence matrix.

Notes: *<sup>P</sup>*(*<sup>i</sup>*, *j*) = *<sup>V</sup>*(*<sup>i</sup>*, *j*)/ ∑*Gi*=<sup>0</sup> <sup>∑</sup>*GJ*=<sup>0</sup> *<sup>V</sup>*(*<sup>i</sup>*, *j*) where *<sup>V</sup>*(*<sup>i</sup>*, *j*) is the value in the cell at row of *i* and column of *j* and G is the number of rows or columns.

Here, the sensitive wavelengths for each texture feature were selected by SPA. SPA is a new statistics-based feature selection algorithm that combines a random-forest permutation method and model-based cluster analysis. The main aspects of SPA are the establishment of sub-models based on samples and variable sub-windows, and the subjection of parameters of interest to strict statistical analysis. The detailed implementation of SPA can be seen in Li et al. [26].
