An Approach Of Road Extraction From High Resolution Remote Sensing Images Based On Vehicle GPS Data Learning

Road traffic network is an important geographical element.It has a large amount of updated data and fast update speed.At the same time,its accuracy and perfection directly affect people's travel navigation,disaster emergency and other aspects of life.However,in actual production and production,there is still a large amount of road drawing work done by field mapping or manual image interpretation.The drawing cycle is long and a lot of manpower and material resources are invested.With the development of GPS positioning technology,remote sensing science technology,computer vision technology,etc.,road extraction has also changed from manual depiction to semi-automatic and automated extraction.Comprehensive use of a variety of scientific techniques and means,intelligent and efficient road information extraction research,has important significance for the improvement of map data information.From the common data sources extracted from the road,the vehicle GPS trajectory data can directly reflect the road geometry,but the information contained is too single and susceptible to GPS accuracy,which may affect the integrity of the extraction.However,although the remote sensing image data is interfered by various noises,the road extraction may affect the accuracy of the results,but the ground features are detailed and comprehensive,and the two have their own strengths,which can complement each other to a certain extent.This thesis proposes a new method for automatic extraction of roads based on complex terrain and multi-level roads in different scenes,using floating vehicle trajectory data and high-resolution remote sensing image data.The main research contents of this thesis are as follows:(1)According to the road grade,according to the degree of density of the track points,the GPS trajectories are rasterized separately,and the trajectory raster images are combined to obtain a reasonable image processing flow by using mathematical morphology related techniques.Road information such as the centerline and key points of the road.(2)In this thesis,the problem of artificially marking the deep learning sample set is huge.According to the road information extracted by the floating vehicle trajectory,the road in the remote sensing image is automatically marked according to the format required by the network training.The completed sample data set includes a training set and a test set for training and testing of the road extraction model.(3)Based on the image semantic segmentation neural network model LinkNet,this thesis constructs the D-LinkNet model,and obtains the road extraction neural network by learning the training sample set containing the trajectory data annotation.After inputting the test set data,the road extraction results of the test area are obtained.In order to verify the validity of the research method,this thesis will learn the road results of the trajectory data model with good extraction accuracy and completeness,compared with the results of no learning trajectory model extraction and floating vehicle trajectory road extraction results.The overall results of the road extraction results are better,indicating the effectiveness of the research method in this thesis.