Cotton Seedling Monitoring Based On Visible Light Image Of UAV

In the process of modern agricultural production,accurate seedling information is the key to the management of crop seedling.As the largest high-quality cotton base in China(the planting area has accounted for more than 70% of the whole country),Xinjiang obviously cannot meet the demand of current fine management by relying on the traditional seedling acquisition method of field sampling survey and manual estimation by plant protectors.How to acquire the information of cotton seedling situation quickly and efficiently and grasp the dynamics of cotton field timely and accurately are of great significance for the fine management of cotton and the improvement of cotton yield.Based on this,this paper collected high-resolution images of cotton in the 3 ~ 4 leaf period based on the remote sensing platform of uav,used image processing and machine learning technology to identify and segment cotton targets,constructed a cotton count model,and finally extracted seedling information such as cotton emergence rate,canopy coverage and growth uniformity based on this model.At the same time,the improved target detection algorithm YOLOv3 is used to identify and locate weeds in cotton field,which provides a basis for weed removal in cotton field.The main research contents and results are as follows:(1)cotton target identification and extraction.The separation of vegetation from background(soil,plastic film)is the prerequisite for obtaining the information of cotton seedling situation.After pretreatment,eight color indexes including GBDI,ExG,NGRDI and NGBDI were selected to analyze the color characteristics of the image,and the cotton target was extracted by combining with Otsu adaptive threshold method.At the same time,the method of morphology and grid line was used to remove the weed noise.The results showed that the color index(GBDI)combined with Otsu method had the best segmentation effect among the 8 color indexes.Mesh line method is better than morphological denoising method,which can effectively avoid the morphological feature change brought by denoising,and lay a foundation for the subsequent extraction of cotton target morphological feature.(2)information extraction of cotton seedling situation.Based on the previous step,the cotton target morphological characteristics were extracted,and the cotton count model was constructed by combining SVM(Support Vector Machine).The seedling emergence rate,canopy coverage and growth uniformity of cotton were extracted by combining model and image.The results showed that the classification accuracy of cotton plant count model reached 97.17%.When the model was applied to three types of plots with different area scales,the predicted seedling emergence rate errors were 5.33%,3.03% and 0.89%,respectively.At the same time,the extracted canopy coverage and canopy variation coefficient can effectively show the overall growth trend and uniformity of cotton.(3)identification and location of weeds in cotton fields.Weeding is one of the tasks that must be carried out at seedling stage.In order to remove weeds accurately,this paper proposes a fast detection method of cotton field weeds based on the improved YOLOv3 model.Aiming at the small targets of cotton and weeds in seedling stage,this paper optimizes the YOLOv3 network model by means of object box dimension clustering,scale extension,multi-scale training and other methods.At the same time,in order to explore the optimal resolution of weed identification in seedling images,data sets with three resolutions(0.10 cm,0.29 cm and 0.52cm)were constructed for verification.The study showed that the improved YOLOv3 model applied to image data of 0.29 cm resolution had the best effect on weed detection in cotton field,and could meet the needs of actual agricultural production in recognition accuracy(F1:92.12%,Recall: 90.26%)and running speed(51 frames /s).In this paper,cotton in the 3 ~ 4 leaf period was taken as the research object,and combined with uav remote sensing images and various analysis methods,the rapid and accurate acquisition of cotton seedling situation was realized.The research results could quickly and accurately obtain the information of cotton seedling situation in a relatively short time,providing technical support for subsequent cotton field management and fine plant protection.