| With the development of artificial intelligence technology and the continuous improvement of remote sensing image resolution,it brings us a new idea of extracting building data from remote sensing images.However,urban surface buildings are composed of scenes with highly complex features.Traditional feature extraction methods are difficult to accurately distinguish the contours of small-scale features such as buildings.Building target extraction has always been a difficulty in the field of remote sensing image segmentation.With the development of artificial intelligence technology,the deep learning algorithm based on artificial intelligence is continuously optimized.It is widely used in the field of image recognition.The use of deep learning image segmentation method to achieve accurate extraction of buildings has a technical basis.On this basis,a deep learning method based on high-resolution remote sensing images is studied.It has important theoretical and practical application value.In order to achieve accurate building extraction,the specific research contents are as follows :(1)Analyze the complex,diverse,small and dense features of buildings in highresolution remote sensing images.Considering the semantic features of image context,a lowto-high guided enhancement network for remote sensing building extraction(LTHNet)based on high-resolution remote sensing images is proposed.The hierarchical structure from low to high is constructed by using the relationship between different channels.The fusion of lowlevel features generates spatial details while using high-level features to construct rich semantic information to achieve accurate extraction of buildings.(2)A Low-to-High Decoder Module(LTH)is proposed to obtain the key features of buildings with complex and isolated backgrounds and realize the fusion of adjacent features.Background Exploration Mod is proposed.More sufficient context information is introduced into foreground features and background features,and more spatial feature details of buildings are retained to achieve compact feature representation within the class(3)Two building datasets are extracted using six networks : U-Net,Deeplabv3 +,HRNet,PSPNet,DSNet and Build Former.In order to explore the extraction effect of the network model in different regions,this paper selects the remote sensing image data of Gaofen No.2 in Nanchang City,Jiangxi Province as the original data,and establishes the Nanchang building vector data set by preprocessing and manual visual interpretation of the original data.At the same time,the Massachusetts building dataset with complex building coverage commonly used internationally is introduced.The two data are cut into 512 × 512 size images,and divided into training set,test set and verification set as the data source of building extraction.The research sample set is further subdivided into three categories : main urban area,rural area and industrial area for qualitative analysis.The results show that :(1)LTHNet effectively captures enough global information and has a good effect on extracting irregular and multi-scale buildings in the main urban area;the key information of a single building with random and irregular distribution in rural areas can be effectively extracted;effectively eliminate the internal adhesion of large-scale buildings in industrial areas;there are fewer problems of false detection and missed detection,and the accuracy is higher.(2)It can effectively restore the spatial resolution,and the building boundary detected by LTHNet is obvious,which is closer to the real building boundary.(3)The proposed LTHNet achieves the highest accuracy in the callback rate,Io U and F1 score indicators,reaching 82.54 %,72.82 % and 84.27 %,respectively The extraction accuracy results on the Massachusetts dataset and the newly built Nanchang building dataset are better than the six methods of U-Net,Deeplabv3 +,HRNet,PSPNet,DSNet and Build Former with high citation rate and high accuracy in recent years.The innovations of this paper are summarized as follows :(1)The original remote sensing image of Gaofen No.2 in Nanchang City was preprocessed,and the building data set of Gaofen No.2 in Nanchang City was constructed.Comparing and analyzing the foreign open source Massachusetts building data set and the Nanchang building data set,the characteristics of buildings at home and abroad are found.(2)The LTH decoder module and BEM module are proposed to reduce the loss of spatial semantic information when the resolution of the feature map is gradually reduced,which has the comprehensive ability to enhance small target recognition,complex building extraction and improve boundary blurring.(3)According to the existing problems of building extraction,the LTHNet model of building extraction based on high resolution remote sensing image is constructed.By refining the rough outline of the building into a more refined building footprint prediction map,the problem of building extraction is improved.The effect of extracting irregular buildings,smallscale buildings and large-scale buildings is better.It can be used as an effective segmentation method for extracting buildings from remote sensing images. |
