Research On Corn Growth Monitoring Based On UAV

As one of the three most widely planted grains in the world,corn has very important economic value,and monitoring crop growth is of great significance for crop cultivation,seed cultivation,yield estimation and any other agricultural task.However,since the current observation methods for corn crops in the field are mainly manual observations,which for corn,a high-plant intensive crop,are difficult low timeliness,and is not conducive to the management of corn planting,also the traditional empirical planting management model is lack of accurate agricultural information data as a guide for planting management,thus is low production efficiency and not suitable for agricultural informatization and large-scale development,so automatic and intelligent monitoring of field corn has become an urgent need for current corn production.Therefore,this thesis chooses the drone remote sensing platform to realize the accurate collection of field images,and uses computer vision technology and deep learning technology to complete the intelligent monitoring of corn growth.The specific research content is as follows:(1)Path planning based on drone energy consumption.The monitoring of corn growth in the later stage has certain requirements for image quality,thus the path planning of the drone have to take image quality and energy consumption factors into consideration,which means minimizing the energy consumption of the UAV while ensuring the image quality.According to the requirements of experimental image quality and the accuracy of the camera carried by the drone,the range of flight altitude and speed are determined.The UAV energy consumption model proposed by Cabreira suggests that the main factors affecting flight energy consumption are turning angle,path length and flight speed.These three factors are studied separately,and first the optimal flight route is determined by using the minimum turn number algorithm combined with the improved grid coverage method.Then,the optimal flight speed is selected by using the relationship between the path length and the energy consumption of flight speed.Finally,the route planning method based on energy consumption optimization is realized by setting flight path and flight speed.The experiment selects a convex polygon area with an area of 111187.5m~2for simulation.When the ground sample distance is set to 0.5cm,the energy consumption of path planning of this thesis is reduced by 29.291k J over the energy consumption of path planning based on convex polygons.(2)For the identification and classification of corn development period,convolutional neural network was selected for study.First of all,according to the morphological characteristics of corn development and important agricultural activities arrangements,the development period is divided into seedling period,extraction period and maturation period,and the images are optimized through Histogram Equalization,noise cancellation filtering,data enhancement and any other processing technology.Three advanced network models Res Net-50,Net-121 and Efficient Net-B0 are used to identify the corn development period,and according to the experimental results,the accuracy of the identification and classification of Res Net-50,Dense Net-121 and Efficient Net-B0 is 87.15%,90.01%and91.13%,respectively.It can be seen that Efficient Net-B0 classification accuracy is the highest,so Efficient Net-B0 is used for the identification and classification of corn development.(3)According to the characteristics of maize growth at seedling stage and jointing stage,a set of evaluation methods for maize growth at seedling stage and jointing stage is put forward.Firstly,the super green algorithm is improved by using the color features of maize plants in seedling and jointing stages,and used to enhance the difference between healthy plants and other backgrounds in the image.Then,the segmentation binary graph of healthy plants is obtained by using threshold segmentation method,and the coverage of healthy plants in the binary graph is counted.Different coverage thresholds are set at seedling stage and jointing stage respectively to judge whether there is phenomenon of absent plants at young seedling stage and withered plants at jointing stage,and the health evaluation criteria at jointing stage were established.If it is determined that there is phenomenon of absent plants in the seedling stage,the image of absent plants is further processed,the seedling skeleton is extracted,the stem position is determined,the crop line fitting line is obtained according to the stem position,and the estimation model is established to estimate the number of plants absent.The experiment shows that the estimation error of the number of absent plants is close to the true value,and therefore,the model can be used to estimate the number of absent plants in seedling stage.(4)In the mature stage of maize,there is a great correlation between maize yield and spikes and the yield can be estimated by counting spikes.However,the spikes in the field often have no fixed size and complex background,so they are difficult to count.Therefore,a method of estimating and counting with Grab Cut algorithm and Tassel Netv2 network model is established.First,the data set of mature stage is preprocessed.Grab Cut algorithm is used to extract maize as foreground images from original images in order to reduce the influence of illumination and complex background on the estimation of counting.Then Tassel Netv2 network model based on local regression counting is used to count corn in a local field,and the loss function of Tassel Netv2 network model is improved.The experimental results show that the average accuracy of the improved Tassel Netv2 on the test set reaches 94.3%,which is 3.2%higher than that of the original model on the test set,the mean absolute error is 4.539,which is 0.805 lower than that of the original model on the test set,and the root mean square error is 5.257,which is 0.668 lower than that of the original model on the test set.It can be seen that the improved Tassel Netv2 has improved the counting accuracy and the robustness of the model.The evaluation index determination coefficientR~2 of the model fitting effect is 0.948,0.028 higher over the original model,which proves that the regression fitting effect of the improved Tassel Netv2 model is better than the original.Finally,the improved Tassel Netv2 combined with the field corn total number estimation model is used to realize the field corn yield estimation.