Study On Remote Sensing Monitoring Method Of Maize Tassel Based On Deep Learning

The tassel of maize grows at the top of maize plant,which is the key sign to judge whether maize enters the heading stage.Relevant studies have pointed out that the size of maize tassel is directly related to maize yield.The accurate identification of maize tassel in farmland has reference significance for maize emasculation and maize variety breeding.Based on the visible light UAV remote sensing technology,this paper obtains the single background corn tassel data set and the complex background corn tassel data set respectively.According to the phenotypic characteristics and background complexity of corn tassels in different data sets,the improved Center Net corn tassel recognition model under single background and the improved YO LO x-s corn tassel recognition model under complex background are established respectively.The main research contents and conclusions are as follows:(1)A single background maize tassel data set and a complex background maize tassel data set were established.Based on visible light UAV remote sensing technology,the images of maize tassels in farmland environment in 2020 and 2021 were obtained.The target phenotypes of maize tassels were analyzed,such as fuzzy target,incomplete target,occlusion,different bifurcation and so on;It was analyzed that there were interference factors such as test sundries,white film,black film,straw,weeds,shrubs,bare land and so on.It provides data basis for the subsequent establishment of tassel recognition model.(2)A corn tassel recognition model based on improved Center Net is proposed,and its detection performance is verified on a single background corn tassel data set.Based on Center Net target detection model without anchor frame,an improved corn tassel recognition model is proposed by analyzing the size distribution of corn tassel,simplifying the backbone feature extraction network and adding location coordinates.Firstly,according to the characteristics of small tassel size,the feature extraction module of image scale reduction in Center Net network is removed to improve the detection speed while reducing the model parameters;Secondly,the location information is added to the Center Net feature extraction model to improve the positio ning accuracy and reduce the missing rate of tassel;Finally,the experimental results show that compared with the anchor frame Yolo V4 and faster r-cnn models,the improved Center Net tassel detection model has a recognition accuracy of 92.4% for corn tassels in UAV remote sensing images,which is 26.22% and3.42% higher than faster r-cnn and Yolo V4 models respectively;The detection speed is 36F/ s,which is 32 f / s and 23 f / s higher than fast r-cnn and Yolo V4 models respectively.The proposed method can accurately detect the smaller corn tassels in the UAV remote sensing image,and provide a reference for the recognition of small and medium-sized target corn tassels in a single background.(3)A corn tassel recognition model based on improved YOLO x-s is proposed.The accuracy and robustness of the model are verified in the corn tassel data set with complex background and different size targets.Based on the YOLO x-s target recognition model,the iterative attention aggregation module is replaced by the series method of down sampling in the path aggregation network,so as to better realize the fusion of high-resolution and low-resolution feature layers.The experimental results show that the average accuracy of the improved YOLO x-s model is improved by 2.15 percentage points compared with the original model.Among them,in the detection of tassels of large target corn,the accuracy was increased by 1.36 percentage points,the recall rate was 6.3 percentage points higher,and in the detection of tassels of small and medium-sized targets,the accuracy was increased by 2.94 percentage points and the accuracy was increased by 5.26 percentage points.At the same time,by analyzing the recognition effect on the test set,it is proved that the improved model has better recognition effect under the background of wet bare soil,black plastic film,white plastic film and weeds.In addition,the counting verification is also carried out on the public data set maize tassel counting datasets.The average error and root mean square error of the model are 3.6 and 4.3 respectively,which are lower than 4.8 and5.8 of tasselnet recognition model,which verifies the robustness of the model.The improved YOLO x-s corn tassel recognition model has higher recognition accurac y and better robustness.