Convolution Neural Network Supported DEM Correction In UAV Low-altitude Photogrammetry

DEM(Digital Elevation Model)provide important data support for the digital city construction,military,basic mapping and post-disaster emergency rescue and other aspects of the work.UAV Low altitude photogrammetry DEM generation technology is one of the hotspots in geo-data processing,Low-level photogrammetry get the digital surface model(DSM)through dense matching.Compared to satellite photogrammetry,its access to surface details is more informative,but bring greater difficulties to DEM automatic generation,manual post-processing and DSM filtering is the common methods to solve such problems.Manual post-processing has the disadvantage of time-consuming and low degree of automation and the already filtering methods had difficulty in filtering buildings,trees and other objects elevation information,the other regions will be moderate,with a certain blindness.Therefore,the automatic identification of buildings,trees and other target areas and DEM repair,the DEM automatically generated with a certain value.In recent years,the artificial intelligence algorithm represented by the depth learning shows excellent performance in remote sensing target recognition and classification,Convolution Neural Network(CNN)is used to identify the target area such as buildings and trees by using the CNN low-altitude remote sensing classification model in this paper.The elevation of the target area is corrected by robust radial neural network elevation surface fitting method,which is designed to automate the repair of DEM by the low-altitude remote sensing data of UAVs.In view of the above,this paper mainly carries out the following work:(1)Based on the principle of convolution neural network,the author construct CNN low-altitude remote sensing classification model,test the classification accuracy of non-ground elements such as house,vegetation and road and verify the validity of the CNN model.(2)Point cloud data for DSM that contains a large number of non-terrestrial features,using the constructed CNN low-altitude remote sensing classification model to identify the DSM data,extract the non-ground elements to build the DEM repair target area,Eliminate the repair target area elevation point,and use the target area adjacent to the elevation point to fit its elevation.Using the height difference energy decay function iterative search to repair the target area adjacent to the elevation point of the selected interval,taking into account the adjacent elevation point of the gross error at the same time and use robust radial neural network elevation surface fitting method to achieve elevation surface fitting of patching target area.(3)DSM filtering and artificial post-processing are compared with the research methods of this paper,they generate DEM,three-dimensional terrain,contour line and select the uniform distribution of the detection points to compare the accuracy respectively.The results show that the method has the advantages of small residuals and the accuracy of the method is close to that of the artificial reprocessing method,and the validity of the method is verified.(4)The four experiments of Kriging,IDW,RBF and local polynomial are compared with the research methods,the experimental results show that the DEM accuracy is greatly disturbed by non-ground elements such as buildings and woods.The method is more accurate than Kriging,IDW,RBF,local polynomial interpolation algorithm,which is suitable for low-altitude photogrammetric DEM automatic repair and repaired DEM can accurate express topography.