DEM Production Technology Based On Low Altitude Remote Sensing And Deep Learning

The Digital Elevation Model(DEM)is an important national geographic information data that uses digital elevation data to simulate fluctuations in terrain surface.The production of DEM through digital photography technology will generate a huge amount of picture data.But the development of drone technology and deep learning technology provide a new way for the acquisition and processing of massive pictures.This paper aims to improve the production efficiency of DEM by combining UAV photography technology,deep learning semantic segmentation technology and DEM repair technology.The main research work is as follows:(1)Design and implement a drone aerial test in the survey area.The collection and pretreatment of remote sensing data is the first step in DEM production.This paper first uses the Dapeng UAV to conduct aerial photography experiments to obtain remote sensing images of the target area in the Beichuan County survey area.Then,the aerial image is preprocessed,and the digital surface model(DSM)of the survey area is generated by image matching technology combined with coordinate information.DSM provide basic data for further production of DEM.(2)A lightweight semantic segmentation network that is more suitable for remote sensing tasks is designed.Identifying non-ground targets in remote sensing images is an important step in making DSM into DEM.This paper analyzes the basic structure of deep learning semantic segmentation network in detail.Considering the massiveness and redundancy of remote sensing data,it is designed based on DeepLabV3.Lightweight semantic segmentation network.The network uses the inverse residual module for feature extraction,uses cascading hole convolution and spaced pyramid sampling to obtain context information,and finally performs classification and upsampling to achieve semantic segmentation of remote sensing images.Under the premise of similar segmentation accuracy,the network operation volume of this paper is reduced by 76%compared with DeepLabV3.(3)The comparison experiment between the network and DeepLabV3 is designed to verify the effectiveness of the network.This paper trains neural networks using ISPR datasets and survey datasets.The survey dataset was created to complement the survey targets not included in the ISPR dataset,such as crops and lakes.In order to complete the data annotation work,this paper uses the watershed algorithm to write the semantic segmentation dataset annotation tool,and completes the annotation of the 2838 remote sensing datasets in Beichuan.The survey data set combines 7776 ISPR split datasets to form the final data set for network training.Experiments from the comparison of pixel accuracy,cross-section ratio,loss function,training time and other indicators,the experiment shows that under the premise of similar segmentation accuracy,the error between the two network is controlled at 5%,but the network of this paper is smaller than the DeepLabV3,more suitable for data massive,high redundancy.Remote sensing data.(4)The deep learning semantic segmentation technology is applied to remote sensing target recognition,and the non-ground objects in the DSM generated by the remote sensing image can be determined.According to the elevation properties of different non-ground objects,this paper has developed three kinds of DEM repair strategies:elevation culling,smoothing filtering and subtracting fixed threshold.Through comparative experiments,the DEM was repaired using the methods of this paper,manual processing and general filtering.Compared with the filtering method,the external conformity accuracy of this method is improved by 63.6%;compared with the manual method,the degree of automation is higher.The DEM repair strategy based on deep learning proposed in this paper can improve the production efficiency of DEM.