含水率是影响光合作用的重要因素之一，为了构建效果更好、更具普适性的油菜叶片含水量(leaf water content，LWC)定量监测模型，以蕾薹期、初花期油菜叶片为研究对象，采用自然风干法去除叶片水分，同步采集叶片质量和光谱信息。为了降低干扰以及消除噪声，采用标准正态变量变换、Savitzky-Golay卷积平滑算法(SG平滑)、多元散射校正、一阶求导和二阶求导5种方法对光谱数据进行预处理，并结合偏最小二乘法(partial least squares，PLS)分析选取最优预处理方法;采用连续投影算法(successive projections algorithm，SPA)筛选预处理后的光谱特征变量，获得对水分含量变化敏感的特征波长;利用支持向量机(support vector regression，SVR)和BP神经网络(back-propagation neural network，BPNN)方法，以特征波长建立的光谱指数为自变量建立油菜叶片水分含量估算模型。结果表明:采用多元散射校正预处理综合表现最好，2个生育期预测集相关系数均达到0。71以上;通过SPA法选择特征变量，分别筛选出特征波长，其中蕾薹期6个，初花期7个;在蕾薹期和初花期叶片水分含量预测模型中，基于SVR模型和BPNN模型建立的模型预测集决定系数(R2)均在0。800以上，均能实现油菜叶片水分含量的精准监测，其中SVR模型预测效果优于BPNN模型，R2分别为0。857和0。827，RMSE分别为1。791和1。521。因此，利用油菜叶片高光谱建模反演油菜叶片含水率能准确监测油菜叶片含水率，可为精准农业水分管理提供理论参考。
Hyperspectral Estimation of Rape Leaf Water Content Based on Machine Learning
Leaf water content(LWC)is an important factor affecting the photosynthesis of rapeseed.In order to establish a quantitative monitoring model for LWC with better monitoring effect and universality,rapeseed leaves at the budding and initial flowering stages were selected as the research objects.The leaves were subjected to natural air-drying to remove water,and the mass and spectral information were collected simultaneously.To reduce interference and eliminate noise,5 methods were used to preprocess the spectral data including standard normal variable transformation,Savitzky-Golay convolution smoothing algorithm(SG smoothing),multiple scattering correction,first-order derivative,and second-order derivative,and the optimal preprocessing method was selected by combining with partial least squares(PLS)analysis.Successive projections algorithm(SPA)was used to select sensitive feature wavelengths for water content changes from the preprocessed spectra.Support vector regression(SVR)and back-propagation neural network(BPNN)were used to establish LWC estimation models based on spectral indices using the selected feature wavelengths as independent variables.The results showed that the multiple scattering correction method performed the best,and the correlation coefficients of the prediction sets for both growth stages were above 0.71.SPA selected 6 and 7 feature wavelengths for the budding and initial flowering stages,respectively.In the LWC prediction models for the 2 growth stages,the models based on SVR and BPNN had determination coefficients(R2)above 0.800 for the prediction sets and could achieve accurate monitoring of LWC in rapeseed leaves.The SVR model had better prediction performance than the BPNN model,with R2 values of 0.857 and 0.827 and RMSE values of 1.791 and 1.521,respectively.Therefore,using high-spectral modeling to invert LWC in rapeseed leaves can accurately detect LWC and provide theoretical reference for precision agriculture water management monitoring.
rapeleaf water contenthyperspectralmachine learning
国家科技期刊平台
NSTL
万方数据
