Research On Building And Road Extraction Technology In High-resolution Satellite Image

Extracting ground object from remote sensing images is a widely studied topic in the fields of computer vision,pattern recognition,and remote sensing.It has a lot of application scenarios,such as the construction of geographic databases,urban construction and planning,and disaster assessment and response.However,due to the variability of ground object shape,the complexity of color and texture,and the occlusion of objects,the automatic extraction technology is still facing too many problems to be a real usable applications.In recent years,deep learning has been widely used in object recognition or semantic segmentation tasks for remote sensing images.Our work here is to apply the convolutional neural networks to the tasks of building extraction and road extraction in remote sensing images,and study the specific problems in the process of building and road extraction.Our work is mainly consist of the following three parts:1.A model combining attention mechanism and feature extraction network is proposed to segment buildings and roads in remote sensing images.Local receptive field of convolution operations lacks the ability to modeling remote context spatial relations.In addition,the channel information of feature maps also plays a key role in deep convolutional networks.In this work,we combined two simple and effective network module,spatial relation module and channel relation module,to extract the long and short range context information in the dimension of channel and spatial,and to learn and reason ablout the weights of different channel features and different regional features.In addition,we designed the strip convolutional multi-branch module according to the roads distribution features.Both the qualitative and quantitative results show that the channel attention module and spatial attention module can improve the performance of the original model,and can adaptively extract effective features from images and supress useless context information.2.Based on the feature extraction capability of convolutional neural networks,building extraction has now achieved high detection accuracy.However,many supervised learning models do not impose constraints on the building boundary information,resulting in less regular shapes of the segmented buildings.For this reason,a directional branching learning module is introduced to generate direction vector field,which will be used in the active contour model to evolve the position of building polygon vertices.By this way,a more regular building boundary segmentation is achieved.3.To solve the problems of road disconnections in the road extraction of remote sensing images,we discuss a connectivity-oriented loss.The main idea is to measure the connectivity of the road network mediately according to the separation caused by the road segment between the background regions.Obviously,if the extracted road is disconnected,the resulting gap will cause the connection of the background regions which should lie on both sides of the road.The loss function is designed to suppress such wrong connectivity between background regions and thus suppress the disconnection of the road network.In addition,the loss function penalizes false disconnections of the background region,thus reducing the generation of false positive examples in the predictions.We demonstrate this loss function improves the road network connectivity and it’s enough to be skeletonize it to generate maps comparable to current most used networks.The loss function can be conviniently combined with any semantic segmentation network.