Actually the NIR method tends to perform best when the constituent range is not extreme, so the model with data for two years has a somewhat lower precision than the model with data for one year.
The previous protein models using data for one year developed by Delwiche et al (1995) and Barton et al (2000) resulted in an SEP of 0.107% using 16 factors (range of 5–10%) and SECV of 0.14% (range of 7.03–10%), respectively.