Abstract: In order to achieve rapid and nondestructive monitoring of leaf water content in the canopy of Lyceum barbarum tree, Ningqi No.7 was taken as the research object to measure the spectra and water content of Lyceum barbarum canopy leaves. Two mathematical transformations (first‑derivative and continuum removal) were carried out on the original spectra. Based on the correlation analysis between the original spectrum (OS), first‑derivative spectra (FDS), continuum removal spectrum (CRS) and water content, sensitive wavelengths  were selected. Random forest regression models (RFRM), partial least squares regression models (PLSRM), ridge regression models (RRM) and univariate regression models (URM)  were constructed. Subsequently, the accuracy of these models was tested and evaluated. The results indicate that the fitting degrees of the FDS‑based models and CRS‑based models range from 0.716 to 0.938 and from 0.710 to 0.920, respectively, while OS‑based models range from 0.710 to 0.874. It is evident that FDS‑based and CRS‑based models have a higher fitting degree than OS‑based ones. From the analysis of model types, the random forest regression models (RFRM) exhibit the best fitting degree at 0.874-0.938 , followed by partial least squares regression models (PLSRM) at 0.826-0.866 and ridge regression models (RRM) at 0.737-0.889, then the univariate regression models (URM) at 0.710-0.730 with the worst fit . A comprehensive analysis reveals that the random forest regression model based on first‑derivative spectra (FDR+RFRM) has the best prediction effect. The fitting degree of the training datasets and the test datasets are 0.938 and 0.893, respectively, and the R2, RMSE, MAE and RPD of validation datasets are 0.872, 0.561, 0.466 and 2.156, respectively. It is concluded that a hyperspectral detection model with high prediction accuracy is developed by combining spectral transformation with machine learning, which is suitable for monitoring the water content of Lyceum barbarum canopy leaves. This provides a suitable and efficient method for monitoring the water content of Lyceum barbarum canopy.

Key words: water content, Lyceum barbarum, hyperspectral, partial least squares regression model, random forest regression model, ridge regression model

摘要： 为实现枸杞冠层水分信息的快速无损监测，以“宁杞7号”枸杞为试验对象，测定枸杞冠层叶片光谱和叶片含水率，对原始光谱进行一阶微分和连续统去除2种数学变换，将获取的原始光谱(OS)、一阶微分光谱(FDS)及连续统去除光谱(CRS)与含水率进行相关性分析，筛选出敏感波长并构建预测含水率的随机森林回归模型(RFRM)、偏最小二乘回归模型(PLSRM)、岭回归模型(RRM)及一元回归模型(URM)，最后对模型的精度进行检验与评价。结果表明：从敏感波长分析，基于FDS构建的模型，其拟合度为0.716~0.938；基于CRS构建的模型，其拟合度为0.710~0.920；基于OS构建的模型，其拟合度为0.710~0.874；可见，基于FDS和CRS构建的模型，拟合度均高于基于OS构建的模型。从模型类型分析，RFRM的拟合度最高(0.874~0.938)，其次为PLSRM(0.826~0.866)和RRM(0.737~0.889)，URM的拟合度最低(0.710~0.730)。综合分析，基于一阶微分光谱构建的随机森林回归模型(FDS+RFRM)预测效果最优，其训练集和测试集的拟合度分别为0.938和0.893，检验集[R2、RMSE、MAE及RPD]分别为0.872、0.561、0.466和2.156。研究将光谱变换与机器学习相结合，开发一套适用于枸杞冠层叶片含水率的且预测精度很高的高光谱探测模型，为枸杞冠层含水率的监测提供适宜高效的方法。

关键词: 含水率, 枸杞, 高光谱, 偏最小二乘回归模型, 随机森林回归模型, 岭回归模型