Improving the Robustness and Stability of Partial Least Squares Regression for Near‐infrared Spectral Analysis

Abstract

Partial least-squares (PLS) regression has been presented as a powerful tool for spectral quantitative measurement. However, the improvement of the robustness and stability of PLS models is still needed, because it is difficult to build a stable model when complex samples are analyzed or outliers are contained in the calibration data set. To achieve the purpose, a robust ensemble PLS technique based on probability resampling was proposed, which is named RE-PLS. In the proposed method, a probability is firstly obtained for each calibration sample from its residual in a robust regression. Then, multiple PLS models are constructed based on probability resampling. At last, the multiple PLS models are used to predict unknown samples by taking the average of the predictions from the multiple models as final prediction result. To validate the effectiveness and universality of the proposed method, it was applied to two different sets of NIR spectra. The results show that RE-PLS can not only effectively avoid the interference of outliers but also enhance the precision of prediction and the stability of PLS regression. Thus, it may provide a useful tool for multivariate calibration with multiple outliers.

Research advances in nondestructive determination of internal quality in watermelon/melon: A review

“…Although NIR spectroscopy technique has been widely used in measuring SSC of various fruits such as apples, pears, and peaches (Crowe and Delwiche, 1996;Slaughter, 1995;Lammertyn et al, 1998;Liu and Ying, 2004;Zhang et al, 2006;McGlone et al, 2007;Liu et al, 2008;Shi et al, 2008;Fan et al, 2009;Camps and Christen, 2009;Bobelyn et al, 2010), there are still several potential limitations in NIR assessment including variation in SSC (Long and Walsh, 2006), fruit sampling location (Peiris et al, 1999;Guthrie et al, 2006;Long and Walsh, 2006), and robustness of calibration models (Golic and Walsh, 2006;Preys et al, 2008;Shao et al, 2009). Dull et al (1989a) used two wavelengths to assess SSC of slice cantaloupes (913 nm and 884 nm) and intact cantaloupe (896 nm and 860 nm), the result of SSC of slice cantaloupes was much better than that of intact cantaloupes (slice cantaloupes: r c = À0.968, SEC = 0.56, SEP = 1.56; intact cantaloupes: r c = À0.60, SEC = 1.67, SEP = 2.18).…”