Abstract:
Objective In order to monitor rice bacterial blight quickly and non-destructively, and guide field operations.
Method High-resolution images of rice canopy under bacterial blight stress were acquired using utilized unmanned aerial vehicles (UAVs). Color features and texture features were extracted from the images, and multiple regression models based on color features, texture features, and the fusion of color and texture features were constructed to predict the infection level of rice bacterial blight. The influence of different ground sampling distances (GSD) on the accuracy of the models was also explored.
Result The determination coefficient (R2) of the monitoring model based on color features was 85.9%, root mean square error (RMSE) was 1.43 and relative RMSE (RRMSE) was 19.1%. The R2 had increased by 2.4 percentage points and RRMSE had increased by 4.6 percentage points compared with the model based on texture features. Compared with single-feature models, the prediction model based on the fusion of color and texture features (R2=89.6%, RMSE=1.06, RRMSE=15.1%) exhibited significant improvement in accuracy. By constructing models with different GSDs, it was found that when the GSD was 0.2, 0.5 or 0.8 cm, the models achieved higher accuracy with R2 all above 80%.
Conclusion The color and texture features extracted from low-altitude remote sensing images captured by UAVs can be used for monitoring rice bacterial blight. The results can provide effective technical support for UAV remote sensing monitoring of rice bacterial blight.