融合GA与SVR算法的小麦条锈病特征优选与模型构建

为提高小麦条锈病遥感监测精度，综合利用反射率光谱在作物生化参数探测方面的优势和叶绿素荧光在光合生理诊断方面的优势，构建了冠层日光诱导叶绿素荧光(Solar induced chlorophyll fluorescence，SIF)协同反射率光谱吸收参量的初始特征集合，并基于融合遗传算法(Genetic algorithm，GA)和支持向量回归(Support vector regression，SVR)算法对初始特征集合与SVR参数进行联合优选，确定遥感监测小麦条锈病严重度的敏感因子，建立基于GA-SVR算法的小麦条锈病遥感监测模型，并将其与相关系数(Correlation coefficient，CC)分析法提取特征参量构建的CC-SVR模型精度进行对比。小区试验数据验证结果表明，融合GA和SVR算法优选特征参量构建的GA-SVR模型精度优于CC-SVR模型，3个样本组中GA-SVR模型预测病情指数(Disease index，DI)与实测DI间的决定系数R2比CC-SVR模型至少提高了2.7%，平均提高了17.8%，均方根误差(Root mean square error，RMSE)至少减少了10.1%，平均减少了32.1%。大田调查数据进一步验证了利用GA-SVR算法对小麦条锈病遥感监测的敏感因子进行优选及模型构建能够提高小麦条锈病遥感监测精度，研究结果为实现大面积高精度遥感监测作物健康状况提供了思路。

Feature Selection and Model Construction of Wheat Stripe Rust Based on GA and SVR Algorithm

Scientific and accurate prediction of the incidence of wheat stripe rust is of great significance for its precise control. Reflectance data can detect crop biochemical parameters, while chlorophyll fluorescence has obvious advantages in photosynthetic physiological diagnosis. In order to improve the detection accuracy of wheat stripe rust and determine the sensitive factors and suitable algorithms for detecting the severity of wheat stripe rust by remote sensing, two feature selection algorithms, filters and wrappers were used to select solar-induced chlorophyll fluorescence and visible light absorption features of wheat stripe rust of different severity. Firstly, the absorption features and SIF data were calculated. Then, the genetic algorithm (GA) and support vector regression (SVR) wrapping method were used to select sensitive features of wheat stripe rust. For comparison, the correlation coefficient method of filter method for feature selection was also used. Finally, GA-SVR model and CC-SVR model for predicting the severity of wheat stripe rust were established by using the characteristics selected by the two methods. The results showed that the GA-SVR model constructed with the combined features of GA and SVR algorithms had better accuracy than that of the CC-SVR model. The verification results of the plot experiment data showed that the determination coefficient between the predicted disease index (DI) and the measured DI of the GA-SVR model in the three sample groups was at least 2.7% higher than that of the CC-SVR model, and the root mean square error was at least 10.1% lower than that of the CC-SVR model. The field survey data verification results also confirmed that using GA-SVR algorithm to optimize the sensitive factors for wheat stripe rust remote sensing detection and model construction can improve the accuracy of wheat stripe rust remote sensing detection. The research results provided a new idea for further realizing large-scale high-precision remote sensing monitoring of crop health status.