Road Segmentation Of Urban Remote Sensing Image Based On Convolutional Neural Network

Remote sensing image processing technologies such as image processing,classification,fusion and understanding have developed rapidly since the United States launched the Earth's first land observation satellite in 1972.High-resolution remote sensing image is widely used in urban development planning,basic geographic information mapping,environmental monitoring and assessment,precision agriculture and public information services.The main application goal of remote sensing images is to extract the information of interested object and identify it to complete image understanding.Road extraction from high-resolution remote sensing satellite images is not only challenging,but also of great research value.Road is the backbone and basic mode of transportation,which provides much support for the development of human civilization.The research of road extraction is of great significance to urban planning,road monitoring,GPS navigation,map revision and image registration.In this thesis,the main research is the use of convolutional neural network for high-resolution remote sensing image road segmentation and extraction.The model of convolutional neural network was improved and the neural network was trained with the road data set of Hohhot city remote sensing image.The main work of this thesis is as follows:(1)he classical convolutional neural network model has some disadvantages,such as reduced resolution of output image,too much information loss on image,insufficient information extraction and too many parameters.So we introduced deconvolution layers,removing the full connection layers of the neural network,replacing it with convolutional layers,and using the deconvolution layers after the convolutional layers.The existence of deconvolution can reduce the number of parameters,retrieve the missing information,and make the output image resolution as large as the original image.(2)High resolution remote sensing image automatic road extraction based on convolutional neural network:In this thesis,we made a remote sensing data set of urban roads in Hohhot,Inner Mongolia,China,trained our improved convolutional neural network,and then extracted urban roads.(3)During the experiment,we studied the influence of the learning rate,batch size and other super-parameters on the model accuracy and training time.Then select the optimal model experimental results can be known:The accuracy of the convolutional neural network model improved in this paper reaches 93.2%in the training set and 90.1%in the verification set.(4)This model is used to extract remote sensing image roads and compared with other model methods.(5)In this thesis,the image feature extraction process of neural network is understood through the visualization of convolutional layer image features in the process of neural network training.