A Comparative Study of Representative Subset Selection for NIR Model Updating

Xiaping Fu,Yibin Ying,Danting Yang
DOI: https://doi.org/10.13031/2013.37411
2011-01-01
Abstract:Because of the uncertain changes (instrumental and non-instrumental), near infrared (NIR) calibration models for fruit quality determination tend to become invalid because of non-modeled variations in the new data. In order to avoid repetition of the whole calibration procedure and solve the model updating problem in NIR spectroscopy, several algorithms related to subset partitioning (Duplex algorithm and Rank-selection method) and calibration updating (Kennard-Stone (KS) algorithm and random selection) were tested with the objective of comparing them in terms of their performance. The issue of pear soluble solid content (SSC) model robustness over time (within different years), and the use of model updating procedure using small number of samples from the new population were examined in this study. The results showed that models established using subsets partitioned by Duplex algorithm performed better than those using Rank-selection method. KS algorithm is more effective for update sample selection. With the addition of approximately 15 samples, models can be successfully updated to improve prediction of new samples.
What problem does this paper attempt to address?