Research On Hierarchical Road Extraction Method From Remote Sensing Image With Upper And Lower Road Structure

At present,the methods of extracting roads from remote sensing images can be divided into three categories: pixel-oriented,object-oriented and deep learning,but this kind of image cannot fully represent the road information in the remote sensing image.For example,in the image s containing the upper and lower road structure,the binary road network image cannot distinguish such road intersections from road intersections on the same horizontal line and connected to each other.These methods are currently limited to extracting a binary road network image.In response to the above problems,the thesis proposes a method for layered extraction of roads in images containing upper and lower road structures,which consists of the following parts:(1)Based on the Massachusetts road data set,the data set for this thesis is divided into two parts: the data set for training the model of remote sensing image road network extraction,and the data set for road hierarchical extraction,both of which contain training set and test set.When building the dataset,the author first cut the original dataset to 144 × 144,then filter the dataset and perform data enhancement operations.The thesis establishes the classification standard of labeling the upper and lower roads,and labels the upper and lower roads in the image s respectively.(2)This thesis proposes a road network extraction model of remote sensing image based on improved dual U-Net network model,and adds a context feature extraction module based on Dense ASPP.The results of the upsampling part of the first U-Net are substituted into the module,and the results are connected with the pooled results of the second U-Net.(3)The thesis proposes an improved hierarchical road extraction method for remote sensing images with upper and lower road structure,which mainly consists of two aspects.The first part is to adjust the network input.After receiving the road network images by using the road network extraction model,the authors adjust the input image and extract hog features.Then,the original image,road network image and feature map are input.In the second part,the loss function of the proposed model is composed by Io U loss with the logcosh method and weighting that with the binary cross-entropy loss.(4)In this thesis,all the improvements are verified by experiments.The first part of the experiment is the extraction of roads from remote sensing image,which includes comparisons of studies under three conditions,the use of images enhanced by data,different loss functions,and different network models;The second part is the experiment of road hierarchical extraction from remote sensing images with upper and lower road structures.The comparative experiments are carried out in three aspects –– whether to include feature map as input,whether to use different loss functions and different network models –– and the best weight parameters of loss function are obtained.Experimental results show that the accuracy,intersection ratio and F1 score of the proposed method can reach 0.9535,0.6172 and 0.7548 in the upper road extraction,and 0.9547,0.6567 and 0.7874 in the lower road extraction,which is better than other models.