Research On Building Extraction In Multiple Scenes Based On Deep Learnin

With the rapid development of real 3D construction in China,the requirement of building information extraction is getting more and more important.Remote sensing images have rich ground-object information,building information extraction is one of the important contents of remote sensing image interpretation.It is important to efficiently and accurately extract building information from remote sensing images.For high-resolution remote sensing image data,buildings in different scenarios have differences in ground features and spectral characteristics.The current building information extraction methods include unclear building outlines,clustered boundaries,and small target buildings cannot be effective extraction,etc.Based on the theory and algorithm of deep learning,this paper adopts the WHU building change detection data set,Inria aerial image data set and UAV image data.On the basis of image enhancement and normalization of remote sensing images,building data sets are made,building a network to extract building information in different scenes,and conducting experimental comparative analysis with other conventional models to carry out research on building information extraction based on deep learning.The main conclusions of this paper are summarized as follows:(1)Aiming at the problems of gradient explosion and location information being easily ignored in building extraction in complex scenes,this paper constructs an improved U-Net network model to improve the accuracy of building information extraction.The improved model firstly adds the residual structure Res Net50 to the encoder part of the U-Net network to alleviate the gradient explosion problem caused by the deepening of the network layer;and then adds the Coordinate Attention Module(CAM)to use this module.Considering the relationship between channels and the location information of the feature space,the ability of the model to collect location information is improved,so that the encoder of the deeper convolutional neural network can obtain rich image feature information.In this paper,the WHU building change detection data set is used to conduct experimental research on the improved model,and compare and analyze it with the U-Net model,the Seg-Net model,and the PSP-Net model.The experimental results show that compared with other models,its accuracy,recall,and Iou have reached 98.74%,95.45%,and 96.40%,respectively.The building extraction accuracy of the improved model is 93.44%,which is better than other comparison models in this paper.(2)Aiming at the problem that only relying on the addition of image features and location information cannot effectively improve the accuracy of image segmentation,this paper constructs an improved PSP-Net model that fuses multiple features to carry out building extraction research.The model first adds a Dilated Convolution(DC)module to the PSP-Net network structure to fuse the deep and shallow features of the image to obtain contextual information as well as global and local feature information;and then adds Spatial Pyramid Pooling(SPP),to enhance the generalization ability of network feature extraction.In this paper,the Inria aerial image data set is used to carry out a comparative analysis of building extraction experiments for the improved model.The experimental results show that compared with the U-Net model,the Seg-Net model,the PSP-Net model and the improved U-Net model,the improved model is more conducive to improve the accuracy of building extraction,and its accuracy,recall and Iou reach 95.90% respectively,85.76%,82.13% building extraction accuracy can reach 85.93%.(3)Based on multi-source remote sensing image data,this paper conducts a case study of building extraction in multiple scenarios.Based on the preprocessing of the UAV image rural building dataset and the Inria aerial image dataset,two images of rural buildings and three images of urban buildings in different scenarios were selected to carry out experimental research to verify that in the case of interference such as roads,vegetation occlusion and bare soil,the feasibility and effectiveness of the model constructed in this paper for extracting building information.The experimental results show that the improved PSP-Net model integrating multiple features can effectively extract stadium buildings with a large footprint;for buildings with a small footprint,the improved model has a good extraction effect and can avoid vegetation,roads and bare soil.interference;for rural dense buildings,the extraction accuracy of the improved model is 89.57%.The final results show that the improved model constructed in this paper can effectively extract building information in multiple scenes.