基于地面高光谱数据的油茶炭疽病病情指数反演

Inverse disease indices of Colletotrichum gloeosporioides based on ground hyper-spectral data

使用FieldSpec HandHeldTM地物光谱仪采集不同发病程度的油茶冠层光谱数据,并实地调查油茶炭疽病病情指数,将光谱数据进行一阶微分与滑动平均滤波相结合的预处理,提取与病情指数相关性较高的敏感波段,并采用主成分分析法(principal component analysis,PCA)对敏感波段的光谱数据进行降维,分别以敏感波段和PCA降维处理后的敏感波段作为输入变量建立了病情指数的BP神经网络反演模型。两种建模方法建立的BP神经网络模型计算出的预测值与观测值之间的决定系数(R2)均达99%以上。精度检验证明,以PCA降维所得到的前10个主成分作为输入变量建立的10-7-1三层BP神经网络模型预测精度更高,模型计算出的预测值与观测值之间的决定系数(R2)和均方根误差(RMSE)分别为0.998 6和0.814 8。该研究表明,利用地面高光谱数据结合主成分分析和BP神经网络算法反演油茶炭疽病病情指数是一种有效的方法。

The Colletotrichum gloeosporioides disease index（DI） was available through field investigation,and the spectral data of canopy with different disease severities was collected by using the FieldSpec HandHeldTM spectrometer.The first order differential of spectral data combined with moving average filter was pretreated.The dimensions of spectral data in sensitive band ranges highly related to DI were descended by PCA.The BP neural network models were built by using sensitive bands and dealt with PCA as input variables separately.The results showed that both determination coefficients between predictive values calculated by the above two models and observed values were over 99%.The accuracy test demonstrated that the prediction precision in the three layers 10-7-1 BP neural network model built by using top 10 principal components from the dimension reduction as input variables was higher.The R2 and the RMSE between predictive values and observed values were 0.998 6 and 0.814 8,respectively.This study showed an effective way to retrieve C.gloeosporioides DI by exploiting ground hyper-spectral data combined with PCA and BP neural network.