机载LIDAR数据特征选择与精确分类技术研究

发布时间：2018-06-07 21:23

本文选题：LIDAR + 地物分类　；参考：《中北大学》2015年硕士论文

【摘要】：激光扫描与测距技术（Light Detection and Ranging,LIDAR)可以快速、主动、自动获取大范围地物密集采样点的三维信息，弥补了传统摄影测量技术获取地物信息单一的缺陷，自上世纪九十年代投入商业使用以来，已被广泛地应用于构造数字地形模型（DigitalTerrain Model,DTM）与数字城市模型、突发自然灾害评估、道路与电力线勘探、生物量估计等领域。如何利用LIDAR系统提供的光谱、纹理、高程、强度等多源信息精确快速地获取地物分布信息已成为当前亟需解决的问题。本文针对机载LIDAR数据地物分类领域内的特征提取、特征重要性分析、分类器构建、分类结果类别混淆出现原因及分类精度优化方法进行了深入的研究和探讨。论文主要研究内容包括： 1.针对以往在对LIDAR数据进行地物分类时，特征选取缺乏依据，主要依赖个人经验与偏好，并由此导致分类精度未能达到最优的问题，首先，，在进行特征提取时，对LIDAR点云与影像数据所提供的特征进行较为完备的提取；其次，在随机森林分类算法框架下，利用袋外样本的特征置换重要性测度评估特征对分类精度的影响程度；最后，选择对分类精度影响较大特征代替原有的高维特征进行分类。 2.从算法原理上，对支持向量机、随机森林、马尔科夫随机场、D-S证据理论等LIDAR数据地物分类领域内常用的分类算法的优缺点进行分析，通过实验验证构建最适合LIDAR数据的分类方案。 3.针对分类器分类结果存在分类精度低、不符合真实地物特性、不符合人们观测习惯等缺陷，研究分析分类结果中易出现混淆的类型、位置；利用目标边缘测度分析混淆出现的原因；根据各类混淆出现的原因，利用地物类间空间限制构建具有针对性的混淆目标类别修正算法，改善分类结果。
[Abstract]:The laser scanning and ranging technology, Light Detection and ranging list (LIDARL), can acquire 3D information of dense sampling points in a large area quickly, actively and automatically, which makes up for the single defect of traditional photogrammetry technology in obtaining ground object information. Since it was put into commercial use in 1990s, it has been widely used in the construction of digital terrain models (DTM) and digital city models, sudden natural disaster assessment, road and power line exploration, biomass estimation and so on. How to use the spectrum, texture, elevation, intensity and other multi-source information provided by the LIDAR system to accurately and quickly obtain the distribution information of ground objects has become a problem that needs to be solved. This paper aims at feature extraction in the field of airborne LIDAR data in the field of ground object classification. The analysis of feature importance, the construction of classifier, the cause of classification confusion and the optimization method of classification accuracy are discussed. The main contents of this paper are as follows: 1. In order to solve the problem that the feature selection of LIDAR data is lack of basis, it mainly depends on personal experience and preference, and thus leads to the classification accuracy can not reach the optimal. Firstly, in feature extraction, The features provided by LIDAR point cloud and image data are extracted completely. Secondly, under the framework of stochastic forest classification algorithm, the importance measure of feature replacement of out-of-bag samples is used to evaluate the degree of influence of features on classification accuracy. Select the feature which has a great influence on the classification accuracy instead of the original high dimensional feature to classify. 2. Based on the principle of the algorithm, the advantages and disadvantages of the common classification algorithms in the field of LIDAR data object classification, such as support vector machine, random forest, Markov random field D-S evidence theory, are analyzed. Through the experimental verification to construct the most suitable classification scheme for LIDAR data. 3. The classification results of the classifier have some defects, such as low classification accuracy, not accord with the real features of ground objects, and do not accord with the observation habits of people, so the types and positions that are easily confused in the classification results are studied and analyzed. The reason of confusion is analyzed by using target edge measure, and according to the cause of all kinds of confusion, a modified algorithm is constructed to improve the classification result by using space restriction between ground objects.
【学位授予单位】：中北大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN249

【参考文献】