Identification Of Cotton Spider Mites Damage Based On Deep Learning And UAV Images

Cotton is one of the important cash crops in China,and Xinjiang is the largest cotton producing area in China.In recent years,the change of ecological environment caused by climate environment has increased the incidence of diseases and insect pests in cotton.Cotton mite is one of the three pests in cotton area of Xinjiang,which does great harm to cotton production.Effective,non-destructive and accurate monitoring of cotton mite damage can make the control of cotton mite more effective,and it is of great significance to reduce the loss of cotton yield.At present,the most widely used method of mite control in China is pesticide control.Long-term use of pesticides will lead to drug resistance and farmland environmental pollution.In this paper,three semantic segmentation models were used to recognize and classify the images of field with cotton spider mite taken by UAV,and the grid maps generated by these three algorithms were compared and classified.On the basis of reading a large number of domestic and foreign documents,this paper uses three semantic segmentation models,and the main work of this paper is as follows:(1)The DJI Phantom 3 Pro UAV was used to shoot remote sensing images at a height of20 m,and then the image Mosaic software Pix4 D Mapper was used to get the orthophoto image of the experimental cotton field.The orthophoto image was cut into small pictures,and the small pictures were labeled to divide the training set and the test set.(2)The three segmentation models used in this paper are U-Net,Deep Lab-v3+,HRNet.The three models were used to identify the occurrence image of cotton spider mite and get a unified evaluation index.The Pixel Accuracy of the three models was 87.4%,85.5% and 90.1%,respectively.Mean Pixel Accuracy was 72.8%,73.9% and 80.3%,Mean Intersection over Union was 55.4%,57.5% and 62.2%,The Frequency Weighted Intersection over Union was80.9%,79.5% and 84.3%,Through analysis and comparison,it can be seen that for the remote sensing data in this paper,the recognition effect based on HRNet network model is the best,and its PA is 90.1%,MPA is 80.3%,MIo U is 62.2%,and FWIo U is 84.3%.(3)The identification images obtained from the three models were splice to generate the identification map of the whole farmland,and the chessboard segmentation was performed on the three identification images to generate the simple distribution map of the occurrence area of cotton spider mites in the cotton field corresponding to the three models.The evaluation coefficient indexes of the three models were summarized and analyzed.The results of this study show that the method based on deep learning for UAV remote sensing image research can effectively distinguish the normal cotton field and the occurrence area of mite,and obtain a high classification accuracy.Among the models with semantic segmentation,the HRNet network model has the best recognition effect and accuracy compared with the other two models,which provides the basis for the subsequent generation of precise drug application prescription chart and the decision-making basis for precision drug application of plant protection unmanned aerial vehicle.