Rapid Detection Of Seedling Condition In Soybean Seedling Stage Based On Deep Learning

With the increasing of soybean planting area,it is very important to use scientific means to effectively monitor the growth of soybean in the process of multiple indicators and growth morphology,for soybean management.In large-scale soybean cultivation,soybean emergence directly affects soybean yield.Accurately identify and judge the emergence of soybean seedlings,able to timely and accurately estimate the growth trend and yield of soybean.The quality of sowing during soybean planting has a great influence on the environment of soybean in the whole growth process.If the planting density is too high,the nutrient competition between plants will be increased,while if the planting density is too low,the population protection effect will be lost.In addition,the adaptability of different soybean seedling forms to the environment is not the same,so a reasonable and accurate judgment of the current seedling stage in the field has important guiding significance for the management of water,fertilizer and medicine in the growth period.Based on the above problems,this paper summarized and analyzed the current situation of crop seedling detection at home and abroad.In order to realize the fast and accurate judgment of soybean seedling situation in the field,computer vision technology and relevant means in the field of deep learning were used to study the target detection model that can automatically detect multistage soybean seedlings.The analysis and calculation of seedling emergence rate,reseeding miss rate,seedling shape ratio,etc.The main research contents of this paper include:(1)The establishment of the detection algorithm of Seedling-Yolo network.Based on the convolutional neural network principle and the self-built soybean seedling data set,the YOLOV4 network was optimized to detect the three different seedling shapes of cotyledon,true leaf and compound leaf of soybean seedlings quickly and accurately.By replacing the main feature extraction network to realize the lightweight processing of the network,the running speed of the detection model is improved,and the accuracy of the identification and detection model is improved by adding attention mechanism.Based on the seedling data of the test set,the m AP value of the average detection accuracy of the network for cotyledon,true leaf and compound leaf seedlings reached 98.1%,and the detection rate was up to 34 FPS.The model was deployed on Jetson Nano with a running speed of 12 FSP,which met the requirements of real-time field operations.(2)Rapid detection of soybean seedling shape in seedling stage.According to Seedling-Yolo,combined with the topographical situation of the experimental field,five-point sampling method was used to collect soybean Seedling images and conduct target detection.According to the number and location information detection results of soybean seedlings in the figure,the relative location coordinates of adjacent seedlings were extracted.According to the distribution of relative coordinate data combined with agronomic standards,the detection of seedling emergence rate,the detection of reseeding missed seeding,and the monitoring of dynamic proportion of three different seedling shapes of cotyledon,true leaf and compound leaf were realized accurately.(3)Realization of soybean seedling detection software.Through analyzing the whole process of the project,the interface of soybean seedling detection upper computer was written based on Py Qt5 framework.The python language,Py Qt5 framework and QT Designer tool were used to design Windows for image reading and display,and at the same time,the trained model was automatically called to show the detection effect and save the Seedling situation result files,which realized the integration of the detection model and improved the convenience of using the model.The results show that deep learning is a fast and effective method for soybean seedling detection.Model training for soybean seedling data sets can quickly and accurately recognize the three seedling forms of soybean cotyledon stage,true leaf stage and compound leaf stage.Using edge computing equipment of deployment model to carry out field real-time detection experiment can accurately judge soybean seedling condition under the premise of ensuring running speed,and provide a reference for farmland managers.