*4.3. Combined Samples for Multiple Environments*

Following the single environment analytics, we tested the performance of a PLSR for the two environments (US and ET) combined. We re-ran the analytics with the same methodologies for the PLS-Full and PLS-Wave models using the combined samples (named USET), and found this generally resulted in better model fits compared to the single sites for both calibration and validation for both the PLS-Full (Table 6) and PLS-wave models (Table 7). Calibration and validation coefficients of determination (*R*2) were all greater than 0.9 for all plant and grain nutrients using both SG and FD data and for both the PLS-Full and PLS-Wave models. RMSE values from the validation were slightly higher than those from the calibration, which is expected, and RMSEp values were all within 11 percent, so the models do not appear to be overfit.


**Table 6.** Partial least squares regression (PLS) results for the full spectrum model (PLS-Full) for plant and grain using the combined United States and Ethiopia (USET) dataset. SG: Savitsky–Golay; FD: First derivative; NLV: Number of latent variables; std: Standard deviation.

**Table 7.** Partial least squares regression (PLS) results for the waveband selection model (PLS-Wave) for plant and grain using the combined Untied States and Ethiopia (USET) dataset. SG: Savitsky–Golay; FD: First derivative; NLV: Number of latent variables; std: Standard deviation.



**Table 7.** *Cont.*
