Research On Fine DEM Generation Method Of Loess Landform Area Based On UAV

Digital elevation model(DEM)is the digital simulation of terrain surface through limited terrain elevation data.It is the modeling expression and process simulation of the geographic phenomenon with continuous change characteristics in two-dimensional geospatial space.As a part of modern digital map 4D products,it is widely used in geomorphology,hydrology,soil science and other fields.In the loess landform area,the surface shape is easily affected by external forces such as rainstorm erosion and human activities in a short period of time,which makes the rapid and accurate acquisition of high-precision DEM data in the area an important basis for the analysis and application of digital topography in the Loess Plateau.Unmanned aerial vehicle(UAV)photogrammetry can obtain massive point cloud data in a short time,which provides data support for building DEM and digital surface model(DSM)representing elevation,and is an important data source for high-resolution surface modeling.Based on the original terrain surface DSM data obtained by UAV photogrammetry,DEM can be obtained by point cloud filtering.The quality and quantity of ground points determine the refinement degree of DEM.However,in the loess landform area,the vegetation type is complex,natural vegetation and crop vegetation interweave,which affects the conversion process from DSM to DEM to a certain extent.In the past,most of the researches were based on rough division of vegetation area,unified reduction of vegetation height to restore DEM,resulting in a large error of ground point elevation,and the DEM data obtained was seriously distorted,which became an urgent problem in the relevant research.In view of the excellent performance of deep learning in image recognition,target detection and semantic segmentation in recent years,this paper proposes a DEM generation method based on deep learning.Firstly,vegetation segmentation is realized by full convolutional network(FCN)model.The model uses the idea of jump connection and u-net multi-level architecture to stack and splice the feature map and the output map of the upper sampling,expand the dimension and express the multilayer characteristics of vegetation.Moreover,regularization and cross validation are introduced to prevent the model from over fitting.Then,based on the extraction results of the surface vegetation elements,an adaptive elevation processing method considering the height of vegetation is used to modify the elevation of the surface vegetation elements,classify and process the sparse single vegetation and continuous vegetation area in the terrain,and finally produce high-precision DEM.In the network model,by comparing the results of different channel combinations,selecting the combination of image map and DSM map,by comparing the vegetation label map,training the image map and DSM map of relevant areas,modifying the parameters to make them have learning ability.The verification accuracy is 94.97%,and the training loss is 0.12.The results of vegetation segmentation in the new sample area show that the F1 harmonic average is greater than 78%;in the conversion from DSM to DEM,three small sample areas in the Loess geomorphic area are selected,and the results are evaluated by using high-resolution image,mapping basic geographic data and measured ground point data.The overall height difference range is between(-0.42 m,0.16m).Compared with a variety of methods and data,the high-resolution DEM generation method proposed in this paper can reduce the vegetation elevation more accurately,express the terrain details in high fidelity,and realize the automatic generation of fine terrain in the loess landform area.