使用K-fold交叉验证方法, 通过两种支持向量机函数, 四种核函数, grid-search算法, 遗传算法, 粒子群算法, 建立对个体净光合速率预测拟合程度最高和最佳惩罚参数c的支持向量机模型。 将可见光光谱组成成分配比关系归为一个P粒子, 将叶温、 散射辐射、 气温等归为一个ε粒子。 通过信息粒子化技术对影响个体净光合速率的因子进行降维处理, 使得分析光合有效辐射、 可见光光谱组成成分和个体净光合速率之间的相关关系成为了可能。 试验结果表明, epsilon-SVR-RBF-Genetic Algorithm模型, nu-SVR-linear-grid-search模型和nu-SVR-RBF-Genetic Algorithm模型对光合有效辐射和P粒子组成预测集的拟合程度均达到97%以上, nu-SVR-linear-grid-search模型的惩罚参数c值最小, 泛化能力最强, 最终采用该模型对光合有效辐射、 P粒子和ε粒子组成的预测集进行预测分析, 拟合程度达到96%以上。

Using K-fold cross validation method and two support vector machine functions, four kernel functions, grid-search, genetic algorithm and particle swarm optimization, the authors constructed the support vector machine model of the best penalty parameter c and the best correlation coefficient. Using information granulation technology, the authors constructed P particle and ε particle about those factors affecting net photosynthetic rate, and reduced these dimensions of the determinant. P particle includes the percent of visible spectrum ingredients. ε particle includes leaf temperature, scattering radiation, air temperature, and so on. It is possible to obtain the best correlation coefficient among photosynthetic effective radiation, visible spectrum and individual net photosynthetic rate by this technology. The authors constructed the training set and the forecasting set including photosynthetic effective radiation, P particle and ε particle. The result shows that epsilon-SVR-RBF-genetic algorithm model, nu-SVR-linear-grid-search model and nu-SVR-RBF-genetic algorithm model obtain the correlation coefficient of up to 97% about the forecasting set including photosynthetic effective radiation and P particle. The penalty parameter c of nu-SVR-linear-grid-search model is the minimum, so the model’s generalization ability is the best. The authors forecasted the forecasting set including photosynthetic effective radiation, P particle and ε particle by the model, and the correlation coefficient is up to 96%.