Research On Urban Road Recognition By Deep Learning Method Based On Multi-source Data Fusion

Digital road information is an important basic geographic element,and it is a key component of China's basic geographic information.The accurate acquisition of digital road information is of great significance to traffic management,urban planning,road monitoring,GPS navigation and map updating.At present,there are two main ways to obtain digital road information: GPS trajectory data and remote sensing images.The road targets in remote sensing images have significant features in space,radiation,topology and texture.With the popularity of deep learning,the method of using the deep learning method to extract roads from remote sensing images has become the main research direction of road extraction research.The GPS trajectory data has the advantages of wide coverage,fast update,easy acquisition,low cost,and strong current situation.Compared with roads in remote sensing images,trees are easily obscured by trees,buildings,etc.GPS trajectory data can reflect road network well.The structure and location information,after rasterizing the GPS track data,the road network is clearly visible.Therefore,this paper proposes a method for road extraction based on GPS trajectory data using a deep learning tool.Considering the complementarity between remote sensing image and trajectory raster image,this paper also designs three road extraction schemes that combine trajectory data with remote sensing image.In this paper,the trajectory data features are analyzed and extracted.The trajectory feature raster map is obtained by rasterization technology.Combined with road vector data and remote sensing images,GeoServer is used to quickly construct the data set needed for deep learning.Based on the data set,the appropriate convolutional neural network structure and loss function are used to learn the semantic segmentation of the trajectory feature raster map,and the structures such as roads and intersections are extracted.Secondly,combining trajectory data with remote sensing images,three different road extraction schemes are proposed:(1)Fusion of trajectory data and remote sensing image at the data level,splicing Multi-channel image,then input network to learn road and intersection information;(2)For the same network,design two different encoders to extract the features of trajectory data and remote sensing image,and trajectory in the decoder stage The remote sensing feature map is fused,unified by the decoder for image restoration,and then the road and intersection information is extracted;(3)two different networks are designed to extract roads and intersections from the trajectory raster map and the remote sensing image respectively.The extracted results are fused by a weighted summation or a maximum value to obtain a final road and intersection extraction result.This paper selects three regions of Wuhan,and uses the 10-day taxi GPS trajectory data of Wuhan City and the remote sensing image of the region to carry out related experiments,and validates the effectiveness of the proposed method.Experiments show that the method of road extraction based on GPS trajectory data using deep learning tools can obtain a relatively complete road structure.The scheme of trajectory data and remote sensing image fusion for road extraction can also effectively improve the accuracy of road extraction.