Building Plane Contour Extraction From High-Resolution Aerial Images And LiDAR Data

Buildings play an important role in people's lives and economic development.The extraction of building contour can provide important data support for national land spatial planning and geographical situation monitoring.In recent years,aviation and satellite remote sensing technology has developed rapidly.Using multi-source remote sensing data to extract buildings can greatly reduce the cost of human and material resources.The airborne LiDAR point cloud can provide high-precision surface elevation information,which helps to distinguish ground and non-ground targets,thereby improving the accuracy of building extraction,but it cannot provide sufficient information such as building outlines and textures.Therefore,the fusion of high-resolution remote sensing image and airborne LIDAR point cloud data is conducive to the realization of high-precision building extraction.Before the methods and techniques of deep learning have not been widely used,the previous extraction methods rely more on artificially designed low and middle level image features to identify buildings,and lack the extraction and application of high-level building semantic features.The design and extraction of features come from the experts' investigation,analysis and summary of buildings,which is difficult to deal with building recognition in complex scenes.In order to solve the shortcomings of traditional classification methods,U-Net is optimized,and DRUnet is proposed.A high-precision semantic segmentation model of buildings combined with DRUnet and morphological filtering is constructed to solve the problem of building extraction in complex scenes.The main findungs are:(1)Aiming at the inconsistency of the input and output image sizes of the original UNet network and the complicated structure of the Crop and Copy channel.The DRUnet convolutional layer uses Padding to simplify the Crop and Copy channel so that it only needs to copy without cropping,which simplifies the network structure and ensures that the input and output image sizes are consistent.(2)DRUnet deepens the original U-Net network to obtain better feature learning capabilities.In order to avoid the disappearance of the gradient,reference is made to the residual learning unit in the residual network that can greatly reduce the gradient disappearance during deep network training.A residual learning unit is added to the network to ensure the forward propagation of the gradient,which greatly reduces the training difficulty of the network.(3)Using the data provided by the Data Plus Energy Analytics Group for experiments,the performance of the building extraction model in this article is tested and comparison experiments are added.Experiments show that the method in this paper can accurately and efficiently extract building areas,the accuracy is 93.23%,the kappa coefficient is 80.46%,the IoU is 78.68%,which is significantly higher than the traditional image segmentation model,and has a greater application prospect.