| Remote sensing image provides a new observation method for earth resource exploration and development,urban construction and development,ecological monitoring and even global research.Buildings are one of the most common targets in remote sensing imagery for earth observation,and their height information can provide key basic data for commercial 3D scene modeling,digital city design,and virtual environment simulation,and has broad application prospects.In this regard,in order to extract building height information accurately,efficiently,in batches and automatically,this thesis carried out research on building height restoration methods in remote sensing images according to the characteristics of remote sensing images.The main contents are as follows:(1)Proposing a rotating detection method of building objects based on remote sensing images.Building object rotation detection method can extract and locate the building object,suppress the interference information,and provide the basis for building height restoration.Aiming at the problems of complex and diverse background,different building scales and random pose in remote sensing images,a dense Inception network is proposed to complete building rotation detection in remote sensing images.Its dense Inception structure and rotation detection structure are the key technical points,which can detect multi-scale and rotating building targets adaptive respectively.Experimental results show that the average accuracy of AP is 82.33% and the recall rate is 93.82% in remote sensing image building data set.In DOTA and HRSC2016 open data sets,compared with the current mainstream methods,the accuracy AP increased by more than1.72% and 1.12% respectively..(2)Proposing a method of building height restoration in remote sensing images based on location information.Based on the rotating detection of building object,this thesis proposes a method to restore building height in remote sensing images based on location information.Among them,the location information processing module can locate the building target according to the location information of the building and reduce the diversity of image pixels.Then,a building height restoration network based on improved UNet was constructed.The residual network was used as the feature extractor,and the location and scale correlation of pixels were considered spatially and globally through the multi-skip connection structure,so as to capture the contextual information related to height,such as shadow,texture and inclination plane.Finally,the local MSE loss function is used to guide the network to focus on the regression of building height information.Experimental results show that the proposed method can significantly reduce the error of high reduction,and MAE and RMSE error indexes on US3 D dataset are only1.22 and 2.17,which are lower than other commonly used methods.(3)Designing and implementing a building height restoration system in remote sensing images.According to the needs of application scenarios,this thesis designs a building height restoration system in remote sensing images from four levels of user interaction,business processing,data storage and basic resources,and integrates the algorithm proposed in this thesis into the system,finally realizing accurate,efficient,batch and automatic building height information acquisition.After testing,the logic and structure of each module of the system are complete,and the whole system meets the function and performance requirements. |
