Research On Road Extraction Method Of High Resolution Remote Sensing Image Based On Deep Learning

Aiming at the problem of low accuracy of road extraction in remote sensing images,using the powerful feature extraction capabilities of neural networks,combined with the rich spatial information of remote sensing data,Research on how to improve the accuracy of road segmentation from high-resolution remote sensing images with complex backgrounds.This article will combine deep learning technology to solve the road extraction task,conducting experimental verification on the Massachusetts road dataset and Deep Globe global satellite image road extraction competition dataset,and two road extraction methods based on deep learning framework have been proposed.The main research work is as follows:(1)Aiming at the problem of blurred road prediction in high-resolution images and improving the accuracy and completeness of road segmentation,a road extraction method based on prediction network and refinement network is designed.First,the prediction network performs initial prediction.A multi-branch dilated convolution model and a multi-kernel pooling module are added to the prediction network.The purpose is to increase the context information and semantic information of road features and improve the integrity of road segmentation.Secondly,the refinement network will optimize the output of the prediction network,and to improve the problem of blurred road boundaries.Finally,the obtained training model can be accurately classified,and the segmented result has a complete road structure.(2)Aiming at the problem of low road extraction accuracy,a road extraction method based on encoder-decoder structure is also proposed.Considering the phenomenon of overfitting in training.First of all,the data expansion operation is performed on the training images in the data set to improve the robustness of the data and the generalization ability of the model.The designed network is a symmetrical structure,and an dilated convolution model based on the feature pyramid model is added at the bottom of the network to extract more semantic information,the purpose is to reduce the loss of road information.Secondly,each layer of encoder and decoder uses skip connections for information transfer,which is used for fusion of road feature information.Finally,each layer of decoders uses a transposed convolution operation to restore the feature map to its original size,and finally trains an accurate road extraction model.Experimental results show that the above designed method has a significant improvement over the mainstream algorithms in terms of accuracy rate,recall rate,F value,and accuracy rate index,and it has obtained higher road extraction accuracy and has certain application value.