叶绿素荧光参数Fv/Fm是探究逆境胁迫对植物光合作用影响的重要指标, 已有研究表明植被指数与Fv/Fm线性相关, 但直接将植被指数与Fv/Fm拟合存在精度不足的问题。 为实现对该参数的准确预测, 本文以茄子为研究对象, 提出一种基于可见-近红外光谱的Fv/Fm预测方法。 试验获取不同生长状态茄子叶片的可见-近红外光谱数据和荧光参数, 使用蒙特卡洛采样法（MCS）去除明显异常样本, 采取3种光谱预处理方法及5种特征波长选择算法进行光谱数据处理, 并建立偏最小二乘回归（PLSR）模型进行方法评估。 基于提取出的最优特征波长组合, 分析误差反传（BP）神经网络、 径向基函数（RBF）神经网络、 极限学习机（ELM）及回归型支持向量机（SVR）共4种机器学习算法对Fv/Fm预测模型精度的影响, 从而确定基于最优方法组合的叶绿素荧光参数Fv/Fm预测方法。 结果表明: 茄子叶片光谱反射率随Fv/Fm的增加呈明显下降趋势, 表明利用光谱信息反演Fv/Fm的可行性。 基于393组试验样本, 使用多元散射校正（MSC）、 标准正态变量变换（SNV）进行光谱预处理, 以竞争性自适应重加权采样法结合连续投影法（CARS+SPA）进行特征波长筛选的效果最优。 其中, MSC-CARS-SPA-PLSR和SNV-CARS-SPA-PLSR的测试集决定系数分别为0.896 1和0.881 2, 均方根误差为0.011 8和0.012 6, 两者精度皆高于全光谱数据对应的PLSR模型; 同时, 两方法提出的特征波长个数均为12个, 仅占全光谱波长个数（1 358）的0.88%。 该结果表明以上两种方法有效提取出了对模型预测有利的少量波长。 基于上述波长建立机器学习模型, 发现SVR建模效果最优。 以SNV-CARS-SPA-SVR的预测精度最高, 其测试集决定系数为0.911 7, 均方根误差为0.010 8。 综上, SNV-CARS-SPA-SVR建模方法提高了模型精度, 有效降低了模型复杂度, 为基于可见-近红外光谱的Fv/Fm准确预测提供了实现方法。 该方法可应用于作物生长状态的快速、 无损检测, 为农情预警提供有效手段。

Chlorophyll fluorescence parameter Fv/Fm is an important indicator to investigate the effects of stress on plant photosynthesis. Previous studies showed a high linear correlation between vegetation index and Fv/Fm. However, fitting Fv/Fm and vegetation index directly showed insufficient an accuracy. In order to achieve accurate prediction of this parameter, this research took eggplant as the research object, and proposed a Fv/Fm prediction method based on Vis-NIR Spectroscopy. The experiment obtained visible-near infrared spectrum data and Fv/Fm of eggplant leaves in different growth states, Monte Carlo Sampling (MCS) method was used to remove obvious abnormal samples. Three spectral preprocessing methods and 5 characteristic wavelength selection algorithms were adopted for spectral data processing. Partial least squares regression (PLSR) models were built to evaluate these methods. Based on the optimal characteristic wavelength combinations, Fv/Fm prediction models were established by four machine learning algorithms: back propagation (BP) neural network, radial basis function (RBF) neural network, extreme learning machine (ELM), and regression support vector machine (SVR). The effects of the algorithms on the accuracy of the Fv/Fm prediction model were analyzed. Therefore, the optimal combination of the above methods, for Fv/Fm prediction was confirmed. The results were as follows: the spectral reflectance of eggplant leaves decreased significantly with the increase of Fv/Fm, indicating the feasibility of retrieving Fv/Fm by spectral information. Based on 293 sets of experimental samples, two sets of characteristic wavelengths with optimal modeling effect were extracted, which were pre-processed by multivariate scattering correction (MSC) and standard normal variable transformation (SNV) respectively, and screened by the combination use of competitive adaptive reweighted sampling method and successive projections algorithm(CARS+SPA). Among them, the test set determination coefficient (R2) of MSC-CARS-SPA-PLSR and SNV-CARS-SPA-PLSR was 0.896 1 and 0.881 2 respectively. The root means square error was 0.011 8 and 0.012 6. Both showed higher accuracy than the PLSR model of the full spectrum data. Meanwhile, both methods selected 12 characteristic wavelengths, which only accounted for 0. 88% of the full spectrum (1 358). This indicated a small number of wavelengths conducive to model accuracy were selected. Among the machine learning models established by optimal wavelengths, SNV-CARS-SPA-SVR obtained the highest prediction accuracy, with a determination coefficient of 0.911 7 and root mean square error of 0.010 8 the test set. Thus, the SNV-CARS-SPA-SVR modeling method used in this research improved the accuracy of the model and effectively reduced the complexity of the model, providing an implementation method for accurate prediction of Fv/Fm based on the visible-near infrared spectrum. This method can be further applied in rapid and non-destructive detection of crop growth status and early warning of agricultural conditions.