UAV Remote Sensing Images Based On Convolutional Neural Network Building Extraction

Buildings are the main components of human life scenes and one of the primary census information in earth observation tasks,so it is of great significance to the mapping and geography industry to obtain accurate building information.In recent years,small UAV tilt photogrammetry technology plays an important role in geographic condition monitoring,smart city construction and disaster emergency response with the advantages of high timeliness,good economy and high accuracy.Compared with satellite remote sensing images,UAV remote sensing technology can obtain building image information with higher resolution and a higher degree of fineness.Therefore,how to automatically,accurately and efficiently extract building information from UAV high-resolution remote sensing images has become a research hotspot in the field of photogrammetry and remote sensing in recent years.In this paper,the problems related to building extraction using deep learning are studied using high-resolution remote sensing images obtained by UAV photogrammetry as data sources,and the specific research contents and results are as follows.(1)The processing process of UAV multi-view tilt images is studied.Firstly,the multi-view tilt images of the study area are collected by using UAV tilt photogrammetry,and then the true projection images of the study area are obtained after data pre-processing,multi-view image joint area network leveling and dense matching steps.Finally,the orthophoto of the study area is used to construct the UAV remote sensing image dataset by manual annotation,which provides the data basis for the method of deep learning in the later paper.(2)To address the problem of poor building extraction by traditional convolutional neural networks,this paper adds an attention mechanism to the jump connection structure of the U-Net network and uses a strategy of supervised training with a hybrid cross-entropy loss function and Lovasz loss function.Both qualitative and quantitative experimental results in the open-source dataset show that the method has fewer errors and omissions in building extraction,and the building structure extraction is more complete,with the F1 score and cross-merge ratio reaching 93.13% and 87.17%,respectively,and several evaluation indexes have been improved compared with the comparison network.The above method can effectively reduce the semantic loss problem caused by feature splicing at the end of jump connections of different layers,and the strategy of mixed loss function can also integrate the advantages of several different mixed loss functions to a certain extent,thus enhancing the robustness of the model.(3)To address the problem of missing spatial and contextual information in the semantic segmentation of remote sensing images by full convolutional networks,this paper studies a method for extracting buildings from UAV remote sensing images based on object contextual information.The method firstly adopts high-resolution network(HRNet)as the backbone network to extract multi-scale high-resolution features with complete spatial information;then divides object regions under the supervision of real labels and calculates the contextual information of pixels and object regions;finally combines the high-resolution features extracted by the backbone network with object contextual information to realize feature enhancement,and then realizes the extraction of buildings in UAV remote sensing images based on the enhanced features.The enhanced features are used to extract the buildings in the remote sensing images.When the above method is applied to the Mengzhou dataset and WHU building dataset,the intersection ratio(IOU)reaches 86.99% and 89.93%,respectively,and the building extraction accuracy is better than the current mainstream deep learning networks.The experimental results show that the method can effectively improve the building extraction accuracy of UAV remote sensing images with comprehensive consideration of multi-scale high-resolution features and contextual information of the object area.