Flowering Classification Of Aerial Images Of Brassica Napus In Large Fields Based On Improved Object Detection Model

As the world’s leading agricultural country,China has been facing major challenges in terms of domestic food production and exports as the glob al population has grown over the last decade.Food crops are an important material for the country’s livelihood and are fundamental to solving the problem of feeding the entire population.Therefore,in the new era of s ocialism with Chinese characteristics,it is more important to create new technologies to enhance the production of food crops and to use adv anced equipment to promote innovative and high-quality development in the field of intelligent agriculture,to achieve the goal of enhancing people’s well-being and improving their quality of life.Rich in high quality vitamins,carotenoids,and many kinds of dietary fiber,brassica napus is an important food crop and oil crop in China.With its excellent nutritional and economic value,Chinese people have been cultivating rape on a large scale,and it is the largest oil crop in China,occupying a leading position in national crop production.As mentioned above,it is an urgent and important agricultural task to pay attention to the growth condition of rape crop,bad condition detection,and breeding research to develop better quality rape varieties.Aerial images of large fields of brassica napus were previously captured by UAV equipment,and although the aerial images were able to capture the brassica napus fields completely,the unavoidable problem was that som e useless areas were acquired in the images,such as greenhouse farming sheds in the fields,irrigation canals,and house buildings and roads at the edge of the fields.These distracting areas will cause interference and inconvenience to the flowering classification study of field brassica napus images,and the accuracy of recognition is not satisfactory.The traditional method of agricultural researchers is to m anually crop the brassica napus field areas in the aerial images through image editing software and subsequently perform flowering identification.Obviously,this method faces a huge time cost problem,and the use of manual cropping is obviously not suitab le if the number of brassicas napus fields in the image is too large,unevenly distributed,and the image information is complex.Therefore,it would be wise to use image object detection techniques in the field of deep learning to solve the above problem.The first innovation of this paper is to optimize and improve the network structure of the original object detection algorithm model YOLO-V5,especially for the original neck structure and head structure,to make the performance of the network model more excellent as far as possible.While the model has a high detection rate,it ensures that the network complexity has a small change to adapt to the hardware experimental platform with a lower configuration.Then preprocess the original field brassica napus aerial images as a training data sample.Train the original YOLO-V5 and the improved network models based on it,compare the performance parameters one by one,and find the b est optimization strategy.After many comparative experiments,the object detection model yolov5-brassica of the field brassica napus image in this paper is determined.Its main performance evaluation index is comprehensively ahead of the original YOLO-V5.The mean average precision of the model has reached 98.55%,with a 2.5% increase,and the average loss value is only 0.01259.The second innovation of this paper is to design an automatic classification model that can cover more rape flowering periods.Using the object detection model yolov5-brassica in this paper,the original aerial images of field brassica napus are pre-processed to obtain images containing only rape regions,which can be used as the training dataset for brassica napus flowering stage classification model.Absorbing the method of transfer learning and combini ng it with deep learning,this paper uses three classic deep learning networks,Inception-V3,Res Net-50,and Mobile Net-V3,as the basis of the flowering classification mo del.According to the experimental brassica napus images and the flowering stage of brassica napus,adjust the parameter configuration of the original network and continue training,and finally determine the flowering classification model of brassica napus by comparing the three training results.The experimental results show that with the help of the yolov5-brassica object detection model in this paper,the recognition accuracy of the Mobile Net-V3-based brassica napus flowering stage classification model can reach a maximum of 97.1%.At the same time,the model can support the classification of seven different types of flowering stages,which has a better recognition accuracy and more classifications than the previous flowering stage classification studies.