基于多核支持向量机的小麦条锈病遥感监测研究

为提高小麦条锈病的遥感探测精度,依据日光诱导叶绿素荧光和冠层反射光谱数据在小麦条锈病遥感探测中的优势及其与病情严重度之间的映射关系,在运用独立分量分析法对光谱数据降维的基础上,利用核学习算法分别确定冠层光谱特征和日光诱导叶绿素荧光特征反映小麦条锈病病情严重度的最优核,同时针对冠层光谱与叶绿素荧光特征组,建立基于不同特征最优核映射的多核学习支持向量机模型,并与基于特征直接拼接的模型结果进行对比。结果表明,对于冠层光谱而言,采用高斯核构建的支持向量机模型可较好估测小麦条锈病病情指数,而日光诱导叶绿素荧光指数则是采用多项式核的效果更优;采用直接拼接法融合叶绿素荧光指数和冠层光谱特征能够在一定程度上改善小麦条锈病病情指数估测精度,决定系数(r~2)最高为0.847,而单独利用冠层光谱信息或者叶绿素荧光信息时,r~2最高仅为0.802;对日光诱导叶绿素荧光和反射光谱特征分别利用其最优核进行映射构建的多核学习支持向量机模型精度最高,r~2为0.915,RMSE为0.090,优于基于特征直接拼接构建的支持向量机模型精度。

Remote Sensing Monitoring of Wheat Stripe Rust Based on Multiple Kernel SVM

In the remote sensing detection of wheat stripe rust,reflectance data can better determine crop biochemical parameters,while chlorophyll fluorescence has obvious advantages in photosynthetic physiological diagnosis. In order to make better use of the advantages of solar-induced chlorophyll fluorescence and canopy reflectance spectroscopy data,a multiple kernel learning support vector machine model based on multi-feature fusion for monitoring wheat stripe rust disease was proposed. Based on the dimension reduction of spectral data by using independent component analysis method,the kernel learning algorithm was used to find the optimal kernel which could better map the severity of wheat stripe rust and canopy spectral characteristics or solar-induced chlorophyll fluorescence characteristics. Then,a multi-kernel learning support vector machine model based on feature-optimal kernel mapping was established by using the combined canopy spectrum and chlorophyll fluorescence feature. Finally,the accuracy of this model was compared with the model based on feature direct splicing. The results showed that Gaussian kernel could better simulate the correlation between canopy spectral characteristics and different disease index of wheat stripe rust,but for chlorophyll fluorescence index,the effect of polynomial kernel was obviously better than Gaussian kernel. In the case of single kernel modeling,the model using the fusion of chlorophyll fluorescence index and canopy spectral features as the feature space had a higher accuracy than that using them alone. Furthermore,the accuracy of the multi-kernel support vector machine model constructed by the different feature component groups mapping by their optimal kernels was better than that of the single kernel support vector machine model constructed by the direct splicing of different feature groups,which r was 0.915,and the RMSE was only 0.090. The study had significance for improving the accuracy of remote sensing detection of wheat stripe rust,and provided a new idea and theoretical reference for remote sensing detection of other crop diseases.