随着人们对近红外光谱分析技术了解的深入, 人们发现通过剔除近红外光谱中的冗余变量不仅可以简化近红外光谱分析模型, 提高模型的可解读性, 通常还可以提高模型的预测效果及稳健性。 变量选择的有效性已经在各种近红外光谱应用体系中得到了广泛的验证, 发展成为了近红外光谱分析建模过程中一个越来越重要的步骤。 为此, 化学计量学家们近些年来开发了大量原理不同的新型变量选择算法, 基于各种原理的衍生算法也层出不穷。 为了让近红外光谱分析研究人员能够较为迅速地对这些算法的特点有所认识, 　对目前常见的各种变量选择算法的算法原理和优缺点进行了梳理。 根据各种算法依据的原理不同, 将目前近红外光谱领域常见的变量选择算法大致分为基于偏最小二乘模型参数, 基于智能优化算法, 基于连续投影策略, 基于模型集群分析策略和基于变量区间等五类。 在梳理的过程中, 我们发现变量选择算法的发展趋势目前主要集中在以下两点: 第一, 算法的复杂程度不断提高; 第二, 不同变量选择算法之间的联用开始逐渐增多。 此外, 作者结合自身在应用变量选择算法时的体会和思考, 还总结了变量选择算法在应用层面上存在的一些问题。 例如光谱预处理方法对变量选择算法使用效果的影响, 以及部分算法存在的稳定性较差, 选择变量的可靠性存疑等。

Researchers begin to realize that near infrared spectroscopy analysis model can be simplified by removing some redundant variables from the full-spectrum with the growing understanding of near infrared spectroscopy. It is obvious that the simplified model constructed with retained informative variables can be interpreted more easily. Moreover, both prediction performance and robustness of calibration model can be improved wi hvariable selection, which has been proved in numerous applied examples. Therefore, variable selection has become a critical step in the process of constructing near infrared spectroscopy analysis models, and various kinds of variable selection algorithms and their derivative algorithms have been developed by chemometrics scientists. In order to help the researchers in near infrared spectroscopy analysis field to have a fast overview on variable selection algorithms, we try to review some variable selection algorithms commonly used in near infrared spectroscopy area in this article, including their main rationales and characteristics. These variable selection algorithms are divided into five categories according to their different features. These algorithms are based on parameters of partial least squares (PLS) model, intelligent optimization algorithms, successive projections strategy, model population analysis strategy, and spectral intervals respectively. During the process of carding literatures, we find that the development trends of variable selection algorithms mainly focus on two points: firstly, complexity of new proposed algorithms increaces continually; secondly, the combination of different algorithms becomes more and more popular. Furthermore, we also summarized several specific applied problems that may be occurred when variable selection algorithms are applied in near infrared spectroscopy analysis area. For example, how do different spectral pretreatment methods affect the performance of variable selection algorithm? How to address the poor stability and reliability of some variable selection algorithms?