Research On Urban Building Extraction From Deep Learning Approach

High resolution remote sensing images can reflect the rich details of underlying surface,and provide a reliable data source for the extraction of surface feature information.Due to the high complexity of the city,the extraction of urban building information has always been a hot and difficult issue based on high-resolution remote sensing image information extraction.Accurate extraction of urban buildings is conducive to the in-depth application and expansion of high-resolution remote sensing in urban planning,smart city construction and other fields.Object-oriented urban building information extraction method has some limitations in practical application,such as low accuracy and large workload of post-processing.In recent years,the development of depth learning technology has made it possible to improve the accuracy of urban building information extraction.Depth convolution neural network can automatically learn the high-level abstract features of the object in the image from a large number of datasets composed of images,thus completing the object recognition and segmentation of the image,such as FCN,U-Net and SegNet.There are complex feature information in high resolution image of urban buildings.If the deep convolution neural network can fully learn and fuse the high-level feature information of buildings implied in the image and apply it to image classification,the accuracy of building information extraction based on high-resolution image will be improved.Based on this,this paper analyses the advantages and disadvantages of FCN,U-Net and SegNet.On the basis of U-Net,a multi-level features involved fully convolutional network is constructed by fusing multi-level feature layers.The network is tested with the open Mnih building datasets,and the IoU index can reach more than 80%,which is obviously better than the building extraction results of U-Net and SegNet.In order to make full use of the panchromatic and multispectral data of the GF-2 satellite image,this paper designs two models: single input model and dual-input model,in which the dual-input model can synthetically learn the characteristics of the two resolution data,and can learn and fuse the information of urban buildings.By comparing with the building extraction results of object-oriented GF-2 image,the results shows that the precision,recall and IoU of the building extraction results of single input model and dual-input model have been greatly improved,while the precision and IoU of GF-2 building extraction based on dual-input model is improved by 5.51% and 5.78% respectively compared with single-input model.The above results show that the multi-level features involved fully convolutional network algorithm and dual-input model designed in this paper can effectively improve the accuracy of urban building information extraction,and provide an effective way for large-scale extraction of urban building information.