Intelligent Delineation And Grading Of Leaf Blast Area In Rice

As one of the most important cash crops extensively cultivated in China,rice is the widely consumed food for a large part of the Chinese population.Therefore,its growth and yield are directly related to China's food security.Leaf disease is one of the main diseases in rice production.The extraction of its characteristic information such as disease spot position and cultivation area is of great significance for the rice leaf blast disease monitoring.Traditional rice leaf blast disease monitoring is based on leaf loss rate calculated by agricultural technology personnel in field sampling survey as well as on the proportion of rice disease spot area with high cost and low efficiency.The situation can be effectively improved by using unmanned aerial vehicle(UAV)as the monitoring platform and image processing technology to delineate the diseased area in the canopy image of the rice field,and monitoring and grading the disease distribution in the region based on the delineation results.The monitoring results are more objective and accurate,which has great research value.With the rapid development of deep learning technology,convolutional neural network has obvious advantages in the field of image processing.In recent years,researchers have proposed various deep network structures for different application scenarios and achieved good segmentation or classification results.This research was carried out in an experimental field in Hubei province.In the field environment,the rice images of tillering stage and disease spot annotated samples taken by UAV at low altitude were taken as datasets,and the leaf blast delineation model was trained based on the structure of Linknet deep convolutional neural network.This paper has mainly achieved the following research results:(1)The image delineation model of leaf blast was established and trained based on the Linknet deep convolutional neural network.On the verification set,the accuracy and recall rates are above 97% on average,and the false alarm rate is about 2%.It is proved that the Linknet structure is suitable for the delineation of leaf blast spot in complex scenes.(2)The original Linknet model was improved based on Focal Loss to make it more suitable for small target delineation tasks in complex scenes by combining with the distribution characteristics of leaf blast spot in the image.The experimental results show that the improved network can significantly improve the accuracy rate,recall rate,false alarm rate and other evaluation indexes.(3)The improved Linknet network was used to delineate the canopy images with different degrees of disease,and the grading model of leaf blast was established by combining the two characteristics of disease spot area and its number in the results.Results from the cross-validated verification experiments and artificial sampling revealed that,the Kappa coefficient is 0.775,which shows that the leaf disease classification model established in the research is highly consistent with the results of traditional artificial investigation,and that UAV for monitoring platform,based on rice leaf spot disease model and hierarchical model for rapid monitoring of rice disease,has certain feasibility.