Research On Road Automatic Extraction Of Remote Sensing Image Based On Deep Learning

Remote sensing technology is widely used in agriculture,urban planning,earthquake prevention and disaster resistance,smart transportation,military strikes and other important fields.The remote sensing images with different resolutions increase the possibility of achieving super-large scale surface surveying and centimeter-level positioning of ground objects.Roads are the carriers of pedestrians and daily transportation,and are very important in economic development and the progress of all walks of life in society.With the rapid progress of satellite technology,extracting roads from remote sensing images automatically updates GIS databases and maps,urban road planning,and intelligent vehicle navigation have extremely important economic value and scientific significance.How to separate roads accurately and quickly from remote sensing images has important theoretical and practical significance for the construction of basic geographic information data.Because the high-resolution remote sensing image of the road contains a large number of other objects,the completeness,accuracy and speed of road extraction are not satisfactory.Therefore,this thesis improves the road segmentation semantic segmentation algorithm for automatic road extraction of remote sensing images,combined with the latest research results in the field of computer vision.Mainly divided into the following parts:(1)Create a road extraction data set of remote sensing images with road labels.Introduces the main process of manual annotation of data sets,data set augmentation methods,image normalization processing,and data set division(2)Through the learning and analysis of the concept of fully convolutional neural network,the use of MobileNetV2 network and a large number of 1*1 convolution kernels as the semantic segmentation network feature extraction part enhances the network feature extraction ability.Based on the residual module,the hollow convolution method is used to increase the receptive field of the neuron.The deepwise separable convolution and other structures accelerate the calculation speed while optimizing the network,and use the porous space pyramid pooling to increase the receptive field.Therefore,it is more conducive to extract large-scale objects.Experiments show that the pyramidal pooling in porous space has greatly improved the extraction of global road features.connection method is used to enhance the network feature extraction ability.Based on the residual module,the hollow convolution method is used to increase the receptive field of the neuron.The deep separable convolution and other structures accelerate the calculation speed while optimizing the network,and use the porous space pyramid pooling to increase the receptive field.Therefore,it is more conducive to extract large-scale objects.Experiments show that the pyramidal pooling in porous space has greatly improved the extraction of global road features.(3)According to the requirements of this thesis,Python is used as the development language,combined with the PyQt graphical interface library and the PaddlePaddle deep learning framework,a deep learning road extraction system is designed.The main functions of the system are three modules of data set training,testing and visualization.Improve the stability and completeness of the system,and also add data preprocessing,data set inspection,data list generation,GPU acceleration and other functions.(4)Three sets of comparative experiments are designed.Experiment 1 combines the hardware resources of the experiment in this thesis.Under the condition of estimating the accuracy of the experiment,the number of remote sensing images fed to the neural network in each batch during training is determined;experiment 2 does not use porous space In the case of the pyramid pooling feature fusion module,five network structures such as FCN8 s,U-Net,and different encoder and decoder combinations are used,and the extraction experiment IoU reaches a maximum of 72.8%,which has a certain degree compared to the classic segmentation network.Improvement;Experiment 3 uses improved porous space pyramid pooling on the basis of Experiment 2,which is an average increase of 3.8% over the original porous space pyramid pooling IoU.