Research On Rice Lodging Region Recognition Based On Semantic Segmentation

The output of rice ranks the third in the world grain output and is also the main grain crop in our country.Lodging of rice is a common phenomenon in the growth process of rice plants,which has a negative impact on the yield and quality of rice,and will also cause difficulties in grain harvesting and reduce the efficiency of mechanization during harvesting.Therefore,it is very important to strengthen the research on rice lodging and enhance the lodging resistance of rice so as to cultivate excellent varieties.In recent years,with the development of UAV remote sensing technology and deep learning,the related research on using UAV to acquire near-earth remote sensing images and conducting deep learning on them is developing rapidly.However,there are relatively few researches on semantic segmentation algorithm in rice lodging research.Therefore,this paper carries out research on using UAV to acquire remote sensing image data sets in the study area based on semantic segmentation algorithm.The main research contents of this paper are as follows:(1)UAV was used to obtain remote sensing images in the study area,and the data set was processed to obtain the rice lodging data set,which could provide data resources for subsequent agricultural insurance claims and system applications in the agricultural field.Data clipping and splicing and data enhancement operations were carried out on the original data of the study area to construct the rice lodging data set of this study.Then,the extraction effect of rice lodging area was studied by using the method of deep learning,and the extraction of rice lodging area provided data support for crop lodging research based on UAV platform.(2)Aiming at the problem of low accuracy of rice lodging image recognition,this paper proposes a semantic segmentation network model,which is an improvement on the structure of UNet network model.The Dense Net model with hollow convolution is used to extract shallow features of remote sensing images.U-Net extracts deep features,and adds attention module at the jump connection,so as to focus attention on the target area.Transposed convolution was used to restore the original size,so as to improve the recognition effect of rice lodging.In this paper,through the research on high-resolution rice lodging segmentation from UAV images,the scope of application of deep learning semantic segmentation is expanded.(3)In this paper,through a series of comprehensive analysis of comparative experiments,lodge Net has a high comprehensive effect in identifying the research area of rice lodging.The accuracy of the model is 97.3%,the PA is 0.9532,and the m Io U is 0.9009.In order to further optimize the label pixel value of the prediction result graph,this paper uses the fully connected conditional random field to modify the pixel value of the prediction result of lodge Net network model.After the correction,the precision,recall and F1 score values of each category are improved.(4)The network optimized model proposed in this paper is applied to the actual lodging scene of the farm.By using the remote sensing image taken by the UAV to input into the network model,the label map with the same size as the original image is output,and the pixel of each plot in the farm is calculated statistically to calculate the area of each category.Thus,the area of lodging rice area,semi-lodging rice area,and normal rice area were calculated.To provide data support for agricultural insurance claims.The lodge Net network proposed in this paper extracts the features of the image by first extracting the shallow features and then extracting the deep features,and adds the attention mechanism in the up-sampling,so that the attention is focused on the area of the current task,which has a good effect in extracting the lodging area of rice.The research on high-resolution rice lodging segmentation from UAV images is of great significance in the fields of rice breeding and agricultural disaster insurance claims.