Research Of Improved Double-Path Convolutional Neural Network Model In Landmark Classification Of Remote Sensing Images

High spatial resolution remote sensing image(hereinafter referred to as highresolution image)can quickly record the geometry,texture,shape,color and other information of terrain objects.It is an important basic data in the fields of land survey,urban planning,national defense security and so on.The classification of terrain features is the basic work to realize the value of high-resolution image data.Although the research of land feature classification methods has been very deep,including decision tree,support vector machine,random forest and so on,the problem of "big difference between the same kind of land features and small difference between different types of land features" still poses great challenges to land feature classification.Therefore,it is necessary to study the classification of terrain features from three aspects: calculation time,recognition accuracy and operation difficulty.At present,Convolutional Neural Network(CNN)has made breakthroughs in the field of image recognition and detection,and its accuracy is far superior to traditional machine learning algorithms.However,due to the limited number and area of open remote sensing image data sets,the application of CNN technology in high-resolution image classification is relatively rare,especially in the field of high-resolution image classification.Therefore,CNN technology is deeply explored and an improved Double-Path Convolutional Neural Network Model(DP-CNN)is proposed to promote the development of terrain classification.The main work of this paper includes the following points:(1)This paper systematically summarizes the main traditional methods of remote sensing image classification,combs the development of CNN technology and its application status in remote sensing field,points out the problems existing in the application of CNN technology in the direction of high-resolution image classification,and points out the research content of this paper.(2)Based on the principle of CNN and using Inception structure and ResNet residual structure for reference,the connection mode of U-Net model is reformed.The original single serial connection structure is replaced by the double-path connection structure of four parallel connection and residual branch parallel computation.The receptive field range of single layer feature extraction is increased,which is more conducive to extracting network features and reducing the parameters of semantic segmentation model in CNN technology.Number,speed up the calculation.(3)This paper introduces the Python platform built by DP-CNN model,and shows the main third-party libraries used in the construction.After the completion of the construction,the exploratory experiment of DP-CNN model based on Inria Aerial Image Dataset data set is carried out.By comparing with the indexes of common semantics segmentation models,the simplicity,rapidity and stability of DP-CNN model are verified.Especially when the number of training rounds reaches 40,the correct recognition rate of DP-CNN model is stable at about 98%.(4)On the high-score image of Taiyuan City,a simple DP-CNN model migration learning application experiment was designed.Based on the open highscore image of Google,a two-class data set of six types of objects is made.Then,the migration training of DP-CNN model is carried out on the data set.Finally,the model prediction and prediction results of the three selected scenes are evaluated.