Crop Classification And Identification Based On UAV Remote Sensing Image With Deep Learning

Agriculture is the source of food and clothing and the basis of production and survival.With the degradation of arable land,water pollution and other major resources and environmental problems,the current situation of the world's shortage of agricultural resources has become increasingly serious.Precision agriculture has practical significance for realizing the important strategies of sustainable development of agriculture such as high quality,high yield,low consumption and environmental protection.The crop information is the core part of precision agriculture,and how to accurately and quickly obtain crop classification and identification information is the basis of precision agriculture.At present,the classification and identification of crops mainly adopt the methods based on aerial satellite remote sensing technology and lowaltitude UAV remote sensing technology,and the aerial satellite remote sensing technology is often used for the identification of large-scale crop types.The advantages of UAV remote sensing technology,such as high efficiency,flexibility,and ease of operation,make it of great application value and research significance in monitoring crop growth in small areas.UAV remote sensing images lack rich spectral information,and there is a phenomenon of "homogeneous images"(that is,the phenomenon of the same kind of ground objects caused by the rich geometric texture of the ground objects caused by high spatial resolution has a large difference in image characteristics),but the traditional remote sensing image classification method has a higher degree of manual participation.Deep learning can automatically learn the deep-seated features of the image to make accurate classification decision,and bring new opportunities for getting better classification results of high-resolution remote sensing image obtained.Based on the theory of Deep Learning,this thesis selects and trains two kinds of deep learning neural network models after UAV data acquisition and Orthophoto Image Acquisition in the research area.The following work is carried out to classify crops based on UAV remote sensing images in this thesis.:(1)UAV data acquisition and processing.The author chose DJi phantom 4pro UAV as the professional equipment with the advantages of the small size,light weight,stable performance,easy to use and high-precision.According to the basic requirements of UAV aerial photographing safety operation issued by the State Bureau of Surveying and mapping,low altitude digital aerial photographing specifications and the field situation of the research area,the author collects UAV remote sensing image in Tang Chang Town,Pidu District,Chengdu City as the research area.Meanwhile,according to the fixed focal length of the camera of the CMOS sensor carried by the UAV and the same effect on the distortion of each image,the image distortion is corrected by the image point coordinate distortion difference,and the external orientation elements of the original image are calculated by the Pix 4d mapper automatic air three,which automatically corrects the image according to the area network adjustment technology,and obtains the UAV Orthophoto Image of the research area.(2)Crop classification model training based on the theory of Deep Learning.For the purpose of reducing the model parameters and calculation amount,the convolution method is changed to depth separable convolution,based on Python language and python deep learning framework,two deep learning models of crop classification are built: googlenet model based on image block classification and u-net model based on image semantic segmentation;sample labels of real ground objects are carried out,and training sample databases based on rice,corn,Loropetalum rubrum and balsam pear are constructed in combination with image enhancement.Input the sample images of the training set and the verification set in the model and use the cross entropy function to calculate the loss value between the feature map of the model training and the ground real label,and use the Adam optimization algorithm to continuously adjust the model training parameters until its loss value convergence,so as to obtain the GoogLeNet optimal model and U-Net optimal model based on crop classification and recognition.(3)Crop classification experiment based on the theory of deep learning and its result evaluation.According to the trained GoogLeNet optimal model and U-Net optimal model based on crop classification,input the classification image to be verified to get the crop classification results of the two models;use the accuracy evaluation index of the model overall accuracy and the confusion matrix to evaluate the classification results of the two models from the classification accuracy.The results show that the classification accuracy of GoogLeNet model is 88.85%,while that of UNet model is 95.78%.In terms of training efficiency,the training efficiency of GoogLeNet model is higher than that of U-Net model.;Comparing the classification results with the traditional Support Vector Machine(SVM)classification method,it is concluded that the classification accuracy based on the GoogLeNet network model and the U-Net network model in this study is higher than the SVM classification accuracy(86.41%).(4)Preliminary analysis of the effect of network parameters on accuracy: In the U-Net model training process of crop classification and recognition scenarios,different initial learning rates and batch sizes are used for model training,and their effects on classification accuracy and efficiency are analyzed.The analysis shows that the network parameters set in this experimental model training have certain advantages.