Automatic Extraction Of Buildings Based On Fully Convolutional Neural Network

Building is one of the artificial features focused on in the fields of intelligent survey of buildings,urban planning,real estate registration,urban and rural construction and urban security.Building information extraction based on high spatial resolution remote sensing images has important research significance and application value.Although the shape and layout of buildings are obvious in high-resolution remote sensing images,which can be easily distinguished from other kinds of ground features in visual interpretation,due to the diversity of building roof materials,building height and architectural styles,automatic extraction of buildings based on high-resolution remote sensing images has been facing severe challenges.At present,although the method of manually drawing the outline of buildings has high precision,it is inefficient and can't be popularized on a large scale.Although the pixel-by-pixel prediction of buildings based on the traditional machine learning algorithm and convolution neural network(CNN)algorithm based on manual screening features is faster than that of manual labeling,the accuracy is often poor.In view of the shortcomings of the above existing research,based on the comparison and analysis of the existing full convolution neural network(FCNs)algorithm,this paper introduces the attention mechanism which can capture context features to improve UNet++,and proposes an attention nested U-shaped network AUNet++ for building extraction.The main research contents and achievements of this paper are as follows:1)Using INRIA aerial image marker data,a high-resolution remote sensing image building automatic extraction data set based on FCNs is constructed,which covers a wide range and has a large amount of data.2)The automatic extraction effect of its application in buildings is analyzed according to the commonly used full convolution neural networks FCN,U-Net and UNet++,.Through experiments,the superiority of UNet++ in automatic extraction of buildings is verified.Compared with FCN and U-Net,UNet++ is the best model with an intersection ratio of 72.05% and an overall accuracy of 95.88%.And the deep supervision of UNet++ makes the model can be pruned and deployed according to the needs after a training.3)Aiming at the defect that FCNs can only extract local information,and fully considering the spatial context information and the attention mechanism of human vision,an AUNet++ building extraction network is constructed.The expected maximum attention and channel attention are introduced to capture the long-range dependence in remote sensing images.The PR curve and ROC curve are drawn according to the extracted building probability map,which qualitatively shows the superiority of AUNet++.AUNet++ obtains the best accuracy with 0.9810 AUC and 0.856 BEP.4)The extracted binary map of the building is processed by the least vertex compression algorithm,and the regular polygon outline of the building is obtained,which makes the real-time update of the electronic base map of the building more convenient and fast.The experimental results show that:(1)compared with other common full convolution neural networks,UNet++ shows obvious advantages in the task of automatic building extraction,(2)after introducing the expectation maximization attention module of context features and the channel attention module of channel weight,the robustness of the AUNet++ building extraction model proposed in this paper is significantly improved.(3)the post-processing of the building is carried out by the least vertex compression algorithm,which further improves the accuracy of the building contour.