Crop Aerial Image Recognition Based On Deep Learning

Traditional agriculture is gradually entering the era of big data,and the concepts of agricultural informatization,internet agriculture,and precision agriculture are increasingly being embodied and applied.The demand for agricultural data information is also knowledgeable,refined and real-time.Among them,crop planting status,planting variety,planting area and growth status are often the premise of agricultural decision-making and agricultural information application.How to collect agricultural planting data,and how to quickly and accurately identify crops are the topic worthy of further discussion.To this end,scholars have proposed collecting agricultural information from the air to easily obtain agricultural big data information.In the past,crop image collection and identification from high altitude were mainly based on remote sensing technology,but this implementation has problems such as complicated calculation and low recognition efficiency,which is limited.In recent years,image recognition and classification have been widely applied under the promotion of deep learning technology.However,the application of deep learning technology to crop remote sensing images obtained by aerial photography has its own unique difficulty.Such as the characteristics of high-altitude features,insufficient detail information,and background diversity interference,how to form an effective multi-classification method is a very worthwhile question.This paper focuses on the method of deep learning in aerial crop image recognition.The main research contents are as follows:(1)In view of the limited public dataset of aerial crop images,this paper uses network crawling and consumer-level drones to capture aerial images of crops.Moreover,in order to make better use of the aerial crop image information and the amount of augmented image data,a method of comprehensive preprocessing of the image by image gradation,data cropping and the like is introduced.(2)In order to reduce the tedious work on the aerial image of the drone aerial photography,the usability of the migration learning technology in the deep network model is analyzed,and the methods of migration learning and data enhancement are discussed.Combining the migration learning technology to realize the image recognition of aerial crops,the method of model migration is proposed.The experimental results prove the effectiveness of the method,improve the speed of training images and reduce the problem of model over-fitting.(3)Based on the original model VGG16,combined with the unique characteristics of aerial imagery,some of its network architectures are improved to construct a new model,which makes it more consistent with the identification and processing of aerial crop images.A regularization method is proposed to reduce the parameter values of the cost function,perform feature adjustment and minimize the model parameters to optimize the model.Based on the Flowers data set,the variation of the three convolutional neural networks before and after the data set doubling is quantitatively analyzed.The model requirements for aerial crop image recognition are proposed,and the regularization method is verified by experiments to improve the image recognition rate.(4)In addition,by constructing a comparative experimental scheme on multiple sets of data sets,the performance and performance of the new model and the original model on the aerial crop dataset were verified.The experiment proved that the aerial image recognition result based on the original model VGG16 was unstable.The improved model is more reliable,and the crop identification accuracy rate can reach 98%,which provides a useful supplement for agricultural wisdom decision-making.