Road Extraction Of Satelite Image Based On Neural Network

With the rapid development of remote sensing technology,the research of high resolution satellite image road extraction technology has become a hot spot in the field of remote sensing.However,road extraction of high-resolution satellite images is still a challenging task due to some special reasons,such as the interference of non-road features and the slender characteristics of the road itself.This paper studies how to accurately extract roads from high-resolution satellite images.Based on u-net neural network,a satellite image road extraction algorithm based on improved u-net neural network is proposed.It improves the precision of road extraction and provides a new idea for subsequent satellite image road extraction.The specific work of this paper is as follows:(1)The road extraction network model is studied.Firstly,the main structures of neural networks and their respective functions are introduced.Then,introduce several neural network models which are outstanding in the field of semantic segmentation,and analyze their advantages and disadvantages in road extraction.Finally,aiming at these shortcomings,an end-to-end network model integrating multi-scale and multi-level feature information is designed.(2)The improved u-net network model is implemented.Firstly,the training data are preprocessed,Then make data set and enhance the data set;Finally,the experimental results are given and analyzed.(3)The transfer learning method was introduced into the improved u-net network model to improve network performance and reduce the risk of over-fitting.Specific implementation: the convolutional neural network on the contraction path of the original model was replaced by the resnet-34 residual network pre-trained in the ImageNet data set.In this way,the characteristics of residual learning and u-net multi-scale fusion were integrated,and the accuracy of model extraction was improved.(4)This paper introduces several main learning algorithms in detail,analyzes and compares the main learning algorithms,and USES the best performance learning algorithm.