Research On Terrain Data Cleaning Method Based On Spatial Interpolation And Machine Learning

In the field of civil engineering,digital terrain models are very important decision-making aids.These models can be used for regional planning,road alignment,resource management,and risk analysis.The original terrain data is the key factor in the construction of the digital terrain model.The original terrain data mainly comes from ground surveying and mapping,the digitization of the original terrain map,Li DAR,and aerial photography.Due to the constraints of technology and environmental factors,especially in the mountainous area with complex terrain,it is difficult to obtain precise terrain data.In this case,Li DAR and aerial photos are generally used to obtain accurate terrain data,but the cost of Li DAR and other methods is too high to be used on a large scale.Therefore,how to use the existing open-source databases such as Google Earth,Geospatial Data Cloud,and USGS to obtain free terrain data,and clean these terrain data to meet the requirements of building a certain precision terrain model is a research direction with practical value in civil engineering and digital terrain model research.In response to the above problems,this paper obtains the original terrain data with a sample accuracy of 3.3m from Google Earth,uses spatial interpolation and machine learning algorithms to clean the acquired terrain data,realizes defect value recovery and local spatial encryption,and compares the performance of spatial interpolation and machine learning algorithms in terrain data cleaning.The main works are summarized as follows.(1)Proposing the ARBF interpolation algorithm.The shape parameters of ARBF are determined by the local density of the space points,and then the radial basis function is adaptive.The adaptive process includes two key steps,one is determining the local data set at the points to be interpolated,the other is establishing the relationship between local point density and shape parameters.In addition,the computational performance of the ARBF interpolation algorithm is evaluated from two aspects of computational accuracy and efficiency.(2)Constructing the Deep Neural Network(DNN)and XGBoost machine learning models for terrain data cleaning,and implementing the algorithm models by Python.The main steps include parameter selection and cross-validation to determine the optimal model parameters,and GPU parallel acceleration of the DNN algorithm.In addition,evaluating the computational performance and applicability of DNN and XGBoost algorithms in the aspect of terrain data defect estimation.(3)Comparing the terrain data cleaning performance of the proposed ARBF interpolation algorithm with other commonly used spatial interpolation algorithms and machine learning algorithms.The terrain datasets are obtained through Google Earth,and the data points of the terrain datasets contain three characteristic values of Longitude,Dimension,and Elevation.Taking the terrain data of a certain area in Fuping County of Shanxi Province as the research object,analyzing the advantages and disadvantages of different algorithms in terrain data cleaning.