基于互信息的含水岩石近红外光谱特征选择

Near-Infrared Spectral Feature Selection of Water-Bearing Rocks Based on Mutual Information

传统的相关分析方法无法准确刻画含水岩石的近红外光谱与其含水量之间的非线性关系。针对这个问题,首先进行了莫高窟崖壁砾岩水分运移的室内试验,分别采集了砾岩样品3个不同位置从初始干燥状态到饱和状态的全过程,共计51条近红外光谱信息;然后采用多点平滑与基线校正相结合(NPS+B-corr)的方法对原始近红外光谱进行预处理,根据强吸收谱段1450和1950 nm处的光谱曲线特征提取峰高,半高宽,峰面积,左肩宽度,右肩宽度,左右肩比共6个初始特征变量建立初始特征集,同时对所提取的光谱特征变量进行归一化处理,根据处理之后的结果绘制各光谱特征参数与含水量变化的曲线,确定含水量级别;接着,进行初始特征集各光谱特征变量间相关性筛选,以便去掉冗余特征,将初始特征集简化为,即由峰高(Height),左肩宽度(LHW),右肩宽度(RHW)三个特征变量构成特征集;最后利用互信息作为相关程度的度量标准,分别采用BIF(best individual feature)法和MIC(maximal information coefficient)法,研究了光谱特征变量与含水量级别之间依赖关系的强弱程度,结果表明:(1)砾岩的近红外光谱在波长1450和1930 nm附近有明显的吸收峰,且吸收峰随着含水量的变化表现出较强的关联变化,说明岩石近红外光谱反射率与岩石含水量有着明显的相关性;(2)初选光谱特征变量与岩石含水量的动态规律曲线呈S形,含水量可划分为干燥、吸水、饱和状态三个级别;(3)两种信息法选取的近红外光谱特征不完全一致,基于BIF法,波长1450 nm处特征变量与岩石含水量级别之间的相关性从高到低排序为右肩宽度,峰高,左肩宽度;1900 nm处为峰高,右肩宽度,左肩宽度。基于MIC法,1450和1900 nm处的特征变量与岩石含水量级别之间的相关性从高到低排序均为左肩宽度,峰高,右肩宽度。(4)利用决策树评估MIC和BIF法的有效性,MIC法比BIF法对含水量级别的识别精度更高。

The relationship between near-infrared spectroscopic measurements of rock and its water content does not follow simple linear correlations,preventing the direct use of classical correlation analysis.In the present paper,an experiment on the water migration in cliff conglomerates from the Mogao Grottoes was performed,and collected 51 pieces of near-infrared spectra from three different positions sample.These spectra cover the whole process of the conglomerate from the initial dry state to the saturated state;then we selected a combined N point smooth and baseline correction method(NPS+B-corr)to preprocess the original near-infrared spectrum.According to the spectral curve features at 1450 and 1950 nm of the strong absorption spectrum,six initial feature variables,namely Height,Full Width at Half Maximum(FWHM),Area,Left Half Width(LHW),Right Half Width(RHW),and(LHW/RHW),were extracted to establish the initial feature set.Simultaneously,the extracted spectral characteristic variables were normalized,and the curve of each spectral characteristic parameter and the change of water content were drawn according to the result of the processing,determine the water content level.Then,the correlation among the feature variables of the initial feature set should be screened to remove redundant features.The initial feature set is simplified to three characteristic variables:Height,LHW,RHW.Finally,based on mutual information,the Best Individual Feature and Maximal Information Coefficient methods were used to evaluate the relationship between samples’spectral characteristic parameters and water content.We found that:(1)at wavelengths between approximately 1450 and 1930 nm,the near-infrared spectrum of the conglomerate has obvious absorption peaks,and the absorption peaks show a strong correlation with the change of water content,which indicates that spectral reflectance was significantly correlated with water content;(2)the relationship of primary spectral characteristic parameters with total water content can be described by an S-shaped function,water content can be divided into three states of dry,water-absorbing,and saturated;(3)The near-infrared spectral characteristics selected by the two information methods are not completely consistent.Based on the BIF method,the correlation between the characteristic variable at 1450 nm and the rock moisture content ranks from right to left as right shoulder width,peak height,and left shoulder width;at 1900 nm,the peak height,right shoulder width,and left shoulder width.Based on the MIC method,the correlation between the characteristic variables at 1450 and 1900 nm and the rock water content level from the highest to the lowest in the left shoulder width,peak height,and right shoulder width;(4)Decision tree analysis suggests that the MIC method achieves higher accuracy in identifying water content level than the BIF method.