Research On Cotton Boll Detection Technology Based On Multi-scale Image And Development Of Cotton Yield Estimation Platform

As the most common fiber crop in the world today,cotton is of great significance to rural economic and social construction and industrial development.The prediction of large-scale crop yield is the main basis for formulating agricultural policies,promoting production and making plans.Cotton yield estimation can provide important information for future import and export planning and may affect market price trends.Through the use of ground images and unmanned aerial vehicle remote sensing images taken at different stages of cotton boll opening period,combined with depth learning and migration learning methods,cotton boll identification and cotton yield estimation at different stages can effectively solve the problems of low efficiency,untimely,inaccurate and destructive traditional sampling yield estimation methods in cotton production.Therefore,this thesis studies the cotton boll detection technology based on multi-scale images,and designs and develops the cotton boll detection and cotton yield estimation platform,which provides application value for cotton precision agriculture.The main research contents and conclusions are as follows:(1)The detection effect of cotton boll image data set on different target detection basic networks is studied.First of all,the cotton boll data images at the ground and UAV scales were collected,and the cotton boll ground image data set and UAV remote sensing RGB image data set during the boll opening period were established,and the image data set was preprocessed to improve the quality of the data set;Secondly,four basic network models for target detection,Faster R-CNN,YOLOv5,YOLOv7 and Center Net,are built to train and migrate on the ground and UAV image data sets.Through the research on the comprehensive performance of cotton boll detection,YOLOv7 is finally selected as the basic network for cotton boll detection.(2)An improved YOLOv7 cotton boll detection model method is proposed.First,in order to obtain more shallow image features,the CA attention mechanism is introduced into the feature extraction layer to improve the detection efficiency of the network;Secondly,the MPConv module is improved.Through the separation and merging operation,the feature map retains all features and reduces the size by half.Through comparative experiments,the performance of the improved target detection network for cotton boll detection of different scales is studied.Compared with the basic network,the improved YOLOv7 detection model improves the accuracy,recall rate,F1 score and AP50 of open cotton boll detection for ground images by 2.3%,2.4%,2.4% and 2.9% respectively,and the accuracy,recall rate,F1 score and AP50 of open cotton boll detection for UAV RGB images by 2.7%,2.9%,2.8% and 2.3% respectively.Compared with YOLOv7 basic network,the improved YOLOv7 is more suitable for cotton boll recognition and detection in complex field environment.(3)The improved cotton boll counting performance of YOLOv7 was studied and the yield prediction model was constructed.On the UAV RGB cotton boll data set,the counting performance of the improved YOLOv7 model is tested and the results show that the detection model R~2 It is 0.85,RMSE is 13.36,RRMSE is 12.42%,and the counting result is good.The cotton yield prediction model is constructed by taking the number of cotton bolls,the nine spectral indexes of UAV image RGB and the four texture features extracted as the input variables of KNN regression algorithm,random forest algorithm and extreme random forest algorithm respectively.Through experimental research,it is concluded that the cotton yield prediction model constructed by random forest algorithm has the best generalization compared with other models,r~2 It is 0.84,RMSE is 4.14,and the linear regression equation between predicted output and actual output is y=1.5177x-36.381,R ~2 Is 0.8473.It can be used to predict cotton yield in actual production environment.(4)The cotton boll detection and cotton yield estimation platform is designed and implemented.The Flask micro-frame is used as the back-end part,and Vue.js is used as the front-end framework for development.The detection and yield estimation model is deployed in the back-end,and the front-end displays the detection results and yield estimation results.Finally,the system test is carried out.The results show that the platform is easy to operate,intuitive and easy to use.The cotton yield prediction results can effectively assist farmers and agricultural researchers in cotton insurance,storage requirements,cash flow budget,fertilizer Make reasonable decisions on water and other input factors.