*5.3. Non-Milled Grains*

Since very few hyperspectral analyses have been performed on non-milled grains [26], we comment briefly on those aspects here. We generally found positive results using hyperspectral data with PLS regression to predict nutrient content of non-milled grains. While the reflectance measurements for these grain samples were collected in a controlled environment (i.e., a dark room using a contact probe), the coefficients of determination were noticeably higher for the grains compared to the plant canopy measurements collected in situ.

An interesting finding emerging from this research is that in the US, several blocks of wavebands corresponding to the VIS-NIR regions were identified as important for protein prediction when using FD values (Table S1, supplementary material). We compared these findings to other studies predicting protein from hyperspectral data [26,56,60]. We do see a considerable amount of overlap in the bands identified in this study for tef, and the bands identified in those studies for forage quality, cooked hams, and cereal grains (rice, oats, and maize), particularly for the FD transformation (Table 8). However, as noted by Talens et al. [56], several of the band ranges found to be important across multiple studies for protein prediction do fall in the range impacted by O-H bonds. So, it is difficult to determine whether these consistencies are true positives or are being affected by other factors.


**Table 8.** Identification of matching bands associated with tef protein content that has also been identified as important protein indicators in past hyperspectral studies. Abbreviations: S: Savitsky–Golay; F: First derivative; B: Both; US: United States; ET: Ethiopia.
