卷积神经网络支持下的无人机低空摄影测量DEM修补

发布时间：2018-03-17 01:06

本文选题：低空摄影测量　切入点：卷积神经网络　出处：《东华理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：DEM(Digital Elevation Model)可为数字城市建设、军事、基础测绘实施和灾后应急救援等方面工作提供重要的数据支持。无人机低空摄影测量DEM生成技术是测绘地理信息数据处理研究热点之一,低空摄影测量通过密集匹配可获取数字地表模型(Digital Surface Model,DSM),相比卫星摄影测量,其获取的地表细节信息更为丰富突出,但也为DEM自动生成带来较大困难,解决此类问题常用方法主要包括人工后处理和DSM滤波方法。但人工后处理耗时、自动化程度低,而已有滤波方法难以针对性过滤建筑物、树林等地物高程信息,对其他区域也会进行平抑,具有一定盲目性。因此,自动识别建筑物、树林等目标区并进行DEM修补,对DEM自动生成具有一定的价值。近年来,以深度学习为代表的人工智能算法在遥感目标识别与分类中表现了优异的性能,本文采用卷积神经网络(Convolution Neural Network,CNN),利用构建的CNN低空遥感分类模型识别建筑物、树林等目标区,通过抗差径向神经网络高程曲面拟合法修补目标区高程,旨在实现由无人机低空遥感数据自动化修补DEM。针对上述内容,本文主要开展以下研究工作:(1)基于卷积神经网络原理,构建CNN低空遥感分类模型,并测试该模型对房屋、植被、道路等非地面要素的分类精度,取得较好效果,验证了该CNN模型的有效性。(2)针对DSM中含有大量非地面要素的点云数据,利用构建的CNN低空遥感分类模型对DSM数据进行判别,提取非地面要素构建DEM修补目标区,剔除修补目标区高程点,并利用目标区邻近高程点拟合其高程。采用高差能量衰减函数迭代搜索修补目标区邻近高程点的选取区间,同时顾及邻近高程点的粗差,通过抗差径向神经网络高程曲面拟合法实现修补目标区的高程曲面拟合。(3)采用DSM滤波以及人工后处理方法与本文研究方法进行对比实验,分别生成DEM、三维地形、等高线,同时选取均匀分布的检核点进行精度比较,结果表明本文方法残差较小且外符合精度与人工后处理方法接近,验证了本文方法的有效性。(4)采用克里金、IDW、RBF、局部多项式四种插值算法与本文研究方法进行对比实验,实验结果表明DEM精度受建筑物、树林等非地面要素干扰较大,本文方法精度高于克里金、IDW、RBF、局部多项式插值算法,适用于低空摄影测量DEM自动修补,修补后的DEM能准确表达地形地貌。
[Abstract]:DEM (Digital Elevation Model) for the construction of digital city, the military, in support of the implementation of basic surveying and mapping and post disaster emergency rescue and other aspects of the work. The UAV low altitude photogrammetry DEM generation technology of Surveying and mapping geographic information data processing is one of research hotspots, low altitude photogrammetry through dense matching can get a digital surface model (Digital Surface Model, DSM), compared with satellite photogrammetry, the surface details of its acquisition more prominent, but also brings great difficulties for the automatic generation of DEM, commonly used methods to solve these problems mainly include manual postprocessing and DSM filtering methods. But the manual postprocessing time-consuming, low degree of automation, it is difficult to filter for filtering buildings, trees and other surface elevation information, will stabilize to other regions, have certain blindness. Therefore, the automatic recognition of buildings, trees and target area And DEM repair, which has a certain value for the automatic generation of DEM. In recent years, with the deep learning artificial intelligence algorithm represented in remote sensing target recognition and classification showed excellent performance, this paper adopts convolutional neural network (Convolution Neural Network, CNN), using CNN low altitude remote sensing classification model to identify the construction of buildings, trees as the target area by RBF neural network robust elevation surface fitting elevation of legal repair targets, to achieve by the UAV low altitude remote sensing data automatic repair DEM. in view of the above content, this paper mainly carry out the work to study: (1) the principle of convolutional neural network based on the construction of low altitude remote sensing classification CNN model, and test the model of vegetation housing, roads, the classification accuracy of non ground elements, to achieve better results, verify the validity of the CNN model. (2) for the DSM contains a large number of non ground point cloud elements The data, to identify the data using the DSM CNN model of low altitude remote sensing classification, extraction of non ground elements to build DEM repair the target area, excluding the repair target elevation point, and use the fitting target area adjacent elevation point elevation. The attenuation function iterative search repair interval selection target area adjacent elevation point by high energy, at the same time the gross error adjacent elevation point, through the robust neural network elevation surface fitting method elevation surface fitting repair target area. (3) using DSM filtering and processing method of artificial and the research methods of this paper were compared respectively to generate DEM, 3D terrain, contour, and select the uniform distribution of check point accuracy comparison results show that the method error is small and the outer precision and artificial postprocessing method to verify the validity of this method. (4) by Riggin IDW, RBF, G, Bureau Part four polynomial interpolation algorithm and the research method of comparative experiments, the experimental results show that the accuracy of DEM by the buildings, trees and other non ground elements of noise, the accuracy of this method is higher than that of IDW, RBF, Kriging interpolation, local polynomial interpolation algorithm, suitable for low altitude photogrammetry DEM Auto repair, repair after DEM can accurately express the topography.

【学位授予单位】：东华理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P208;P231

【参考文献】