为建立基于高光谱的苎麻叶片含水量估测模型, 在大田栽培条件下, 采集了360个苎麻叶片高光谱数据和相应的叶片含水量。用高杠杆值排除异常样本, 用浓度梯度法划分样本集。采用多种光谱预处理方法, 建立并比较各预处理方法的PLSR(partial least squares regression)模型效果, 其中OSC(orthonormal signalcorrection)预处理方法最佳, 预测集R2=0.8503, RMSEp=0.0235。为了减少变量个数, 通过OSC_PLSR模型的回归系数RC(regression coefficient)选择特征波段EB(effective bands)作为输入变量。随后, 为了进一步降低计算量, 本研究提出的一种新的特征提取方法: 在基于RCEB建立的PLSR模型中, 再次提取RC特征波长EW(effective wavelength)。由建模结果可知: 与全波段相比, 2种特征提取方法的变量个数均大幅减少(全波段为2 031个, RCEB为508个, RCEB_EW为16个); RCEB_PLS模型预测集指标最佳(R2=0.8546, RMSEp=0.0232); 与RCEB_PLS模型相比, RCEB_EW_PLSR模型预测集指标略低(R2=0.8499, RMSEp=0.0234), 但这种方法变量个数最少, 因此综合评价效果最优。研究探讨了叶片高光谱与含水量之间的量化关系, 建立基于高光谱的叶片含水量预测模型, 对作物栽培中水分的实时监测和精确诊断具有实际指导意义。

In order to establish a hyperspectral estimation model of water content in ramie leaves, the objective of the experiments is to develop a key method for fast and nondestructive monitoring leaf water content in ramie.We collected 360 hyperspectral data on ramie leaves and corresponding water content under the condition of field cultivation.The leverage were used to exclude the outliers, and the sample sets were divided with the concentration gradient method. Moreover, the PLSR models were respectively built for many spectral preprocessing methods and the effects of these models were compared. Among these methods, the OSC preprocessing method has the best effect, the prediction sets used here are R2=0.8503 and RMSEp=0.0235. In order to reduce the number of variables, the Effective Bands(EB)were selected as the input variables through the Regression Coefficient(RC)of the OSC_PLSR model. Then, this study presented a new feature-extracting method, which extracts RC’s effective wavelength(EW)again in the PLSR model built on the basis of RCEB, to further reduce the computation. What can be observed from the modeling results involve:the variables of the two feature-extracting methods are reduced significantly, the number of variables used in the full-wave band, the RCEB and the RCEB_EW is respectively 2 031, 508 and 16; the RCEB_PLS model has the best indicators of prediction sets(R2=0.8546, RMSEp=0.0232); when compared with the RCED_PLS model, the indicators for the prediction sets of the RCEB_EW_PLSR model is slightly lower(R2=0.8499, RMSEp=0.0234), but this method has the fewest variables, and thus has the best comprehensive effect.This study explored the quantitative relationship between hyperspectral and water content of leaves, and established the hyperspectral leaf’s water content model,which has practical significance for real-time monitoring and accurate diagnosis of water in crop cultivation.