Hyperspectral Monitoring The Damage Of Rice By Rice Leaf Folder Cnaphalocrocis Medinalis Guenee

Crop pests and diseases are one of the main factors caused the loss of agricultural production. Accurate monitoring and forecasting of pests and diseases can guide the effective management and control of pests and diseases, and reduce the economic losses. Hyperspectral remote sensing technology is accurate, timely and objective to get the ground vegetation index, and can be indirectly used to monitor the crop pests and diseases.In this paper, we focused on rice leaf folder Cnaphalocrocis medinalis Giienee (RLF), and measured the hyperspectral reflectance from the leaves and canopy of rice in plot and field at different growth stages of rice (tillering, booting and flowering stages). The spectral characteristics of rice damaged by RLF were studied. The diagnostic models for damaged levels of rice based on sensitive spectral reflectance, the first derivative value of reflectance, and reflectance’s principal component were built using the linear regression, BP neural network and the GA-BP network methods, in order to supply some effective method to monitor the damage of RLF by hyperspectral method. The main results are as followings.(1) The reflectance from leaves and canopy of rice was measured, and the correlation between reflectance and roll leaf rate was analyzed. These wavelength ranges:violet, green peak, red and near-infrared wavelengths were the sensitive ranges to reflect the number of roll leaf in combined leaves of rice at the booting stage. And at the flowering stage, the sensitive wavelengths were located at the visible and a part of the near-infrared ranges. In plot of rice, the sensitive wavelengths were the blue-violet, the green and near infrared segments at the booting stages of rice, and only a few part of the red ranges and the most of the near infrared ranges were sensitive to the scales of rice damage. In the natural rice field, the sensitive spectral wavelengths were red and near infrared ranges at tillering stage, only near infrared ranges at rooting stage, and only blue-violet and red ranges at flowering stage.Regression diagnostic models for roll leaf rate based on the reflectance at the sensitive bands of rice were constructed. These regression models were relative good for the combined leaves and rice in plot, and they were very poor for diagnosing the roll leaf rate in rice field. So the linear regression models based on the reflectance from rice canopy to diagnose the damage levels of rice by RLF were very limited. Principal components of reflectance were extracted from original spectral reflectance at bands of400～1000nm by PCA method. Each of the principal components presents different bands spectral information, and two first principal components could distinguish that the different roll leaf rate in the combined leaves and rice in plots, but they were ineffective to diagnose the roll leaf rate in natural rice filed.The diagnostic models for damage levels of rice had also been established based on the principal components of reflectance. The correction rate for predicting and simulating of the number of roll leaf in combined leaves was42.9%and95.2%, respectively at booting stage of rice; and at flowering stage, the correction rate was87.5%and90.5%, respectively. The correction rate for simulating of the roll leaf rate levels at the rooting stage and the flowering stage of rice was91.7%and62.1%, respectively. However, the regression models from the principal components were poor to diagnose the levels of roll leaf rate in rice field.(2) Diagnostic models for RLF damage of rice based on principal components of spectral reflectance were built using the BP neural network method. The correction rate for diagnosing the roll leaf in combined leaves of rice at booting stage was50%, and at flowering stage it was100%. The correction rate for simulating of the BP models at the rooting stage and the flowering stage was83.3%and60%; respectively. Moreover, the BP’s correction rate to diagnose the roll leaf rate in rice field was73.3%,60.9%, and53.6%at the tillering stage, rooting stage and the flowering stage, respectively. The accuracy of BP network models established by principal components of spectral reflectance still did not meet the need for monitor the damage of RLF by canopy spectral reflectance.(3) BP networks combined with the genetic algorithms (GA-BP) were established to improve the accuracy of diagnostic models to diagnose the damage of RLF based on the principal components of reflectance. The results showed that the GA-BP had a high correction rate to diagnose the roll leaf rate in combined leaves and plot rice both at booting and flowering stage, and the rate was near100%. Specially, the correction rate for predicting the roll leaf rate in natural rice also was close to90%at tillering, booting, and flowering stage of rice. It concluded that the GA-BP network method is an effective solution to establish the diagnostic models for predicting the damage of RLF, and it would be used to monitor the RLF in rice field.