【作者】 董保根；

【导师】 秦志远；

【作者基本信息】 解放军信息工程大学 ， 摄影测量与遥感， 2013， 博士

【摘要】 激光雷达测量（LiDAR）作为一种快速获取空间数据的新手段，在地球空间信息科学领域得到了广泛应用。借助LiDAR技术获取的地物三维点云数据与传统意义上的遥感技术获取的影像数据相比，仍存在语义信息匮乏、特征不连续等缺陷，使得地物智能化分类和识别等处理难度加大。而高分辨率影像虽具有光谱特征显著、纹理信息丰富等优点，但同物异谱、同谱异物等现象导致利用影像进行目标自动解译可靠性降低。因此，针对不同传感器获取数据的优点和局限性，将点云与影像数据融合进行地物智能分类和识别以弥补单个数据源的缺陷具有重要的研究价值。本文基于多源数据融合理论，以支持向量机（SVM）、决策树为分类器，对机载LiDAR点云与高分辨率多光谱遥感影像融合的地物分类理论和技术进行深入研究，通过实验及特征空间分析揭示两种数据融合分类的深刻内涵。研究工作及创新点主要包括：1．从技术层面对机载LiDAR点云与遥感影像融合分类涉及的配准和滤波问题进行了研究。尤其是利用统计矩原理实现了基于偏度平衡的点云滤波算法，该算法最大的优势在于无需阈值或参数支持，且独立于LiDAR数据格式和分辨率，采用非监督分类法将地面点与非地面点自动分离。与传统算法相比，此算法在Ⅱ类误差的控制上略有不足，但高效、智能化处理的特点使其仍具有较大的实用价值。2．针对以往点云与影像融合分类技术中存在的“特征少，精度低”的缺点，重点加强点云特征优化提取策略研究。将点云特征依据不同的提取原则分为直接特征和间接特征，提出了在特征向量中加入原始点云局部几何属性特征以提高分类精度及稳定性的思路，并实现了五类法向量特征提取方法。对特征影像的分析表明，法向量特征能够有效区分道路、不同形状的建筑物、植被，并对车辆和植被识别有较强的敏感性。3．将SVM引入点云与影像融合分类，提出了影像辅助点云及点云辅助影像两种多特征分类模式。分类结果表明，本文方法对不同数据均具有较强的适应性，并且两种分类模式都获得了较高的分类精度，特别是点云总体分类精度和Kappa值达到了90%以上，且成功实现车辆这一特殊地物类型的提取。对比实验表明，无论在分类精度上还是在地物类别数量上，本文方法均明显优于传统的分类方法。4．采用正反向结合方法对影像辅助点云及点云辅助影像两种分类模式下的特征空间进行了深入分析，得出如下结论：①在影像辅助点云模式下，点云归一化高度（NH）特征对分类精度影响最大；②加入点云局部几何属性特征后，大大降低了建筑物与树木类的错分概率，解决了建筑物边缘点提取难题；③两种分类模式的对比分析表明，点云对影像分类的贡献量远大于影像对点云分类的贡献量，点云辅助影像分类更具实际意义；④以机载LiDAR点云为辅助数据的地物融合分类技术将在未来占有重要的战略地位。5．改进了非线性Mode滤波器以改善点云和影像数据的分类结果质量。针对点云和影像两种数据的不同特点，分别利用k近邻型和窗口型Mode滤波器去除斑点噪声和椒盐噪声。对比实验结果证明，改进的Mode滤波器能有效地提高点云和影像两种数据分类结果的精度。6．提出并实现了点云高程数据支持下影像上地物精细分类的方法。为保证高精度地同类别地物再划分，综合考虑辅助数据源、Mode滤波器、点云密度及影像空间分辨率、首次回波四种因素，利用决策树显著地提高了影像上建筑物、植被的分类数量，使点云与影像的融合分类优势得到进一步体现，达到了分类精度与地物类别数量相统一的预期目的。7．提出并实现了多光谱影像辅助下点云中树种分类的方法。为避免特征提取单一而导致分类精度下降问题，除归一化高度、强度等直接特征外，重点提取点云中能够描述树种冠层表面信息的间接特征，包括冠层回波特征、冠层几何属性特征、冠层垂直结构特征。实验结果表明，联合机载LiDAR点云与多光谱影像数据并且利用SVM分类器获得了较高的树种分类精度和更精确的类别数量。更多还原

【Abstract】 Airborne Light Detection and Ranging(LiDAR) is widely applied in the domain ofGeo-spatial information sciences as an emerging means which is a quick access to spatial data.Compared with remote sensing imagery data in tradition, the3D point clouds data acquired byLiDAR technology have a lot of flaws, such as the scarcity of semantic information,discontinuous characteristics and so on. Thus, it is more difficult in intelligent classification andrecognition of land cover types. Meanwhile, though very high resolution imagery can provide alarge amount of spectral features and texture information, the phenomenon of spectralheterogeneity even within the same class and different objects becoming more spectrally similarmake the reliability of automatic target interpretation in imagery decrease a lot. Consequently,according to the specific advantages and disadvantages of different sensors, integrating pointclouds and imagery to complement single data resource to perform intelligent classification andidentification of land cover types has important research values.Based on the theories of multi-source data fusion and classifier of Support VectorMachine(SVM) and decision tree, this thesis aims to investigate an intensive and deep study onthe theory and technology of classification of land cover by fusing airborne LiDAR point cloudsand high spatial resolution multi-spectral remote sensing imagery. Subsequently, profoundsignificance is revealed through experiment and feature space analysis. The main contents andinnovations in this thesis can be summarized as follows:1. The subjects of registration and filtering involved in classification of fusing airborneLiDAR and remote sensing imagery are researched from a technical level. Especially a filteringalgorithm based on skewness balancing for point clouds is achieved by exploiting statisticalmoments. The main advantages of the algorithm are threshold-freedom and independence fromLiDAR data format and resolution, and it separates ground points and non-ground pointsautomatically by unsupervised classification method. Compared with conventional algorithms, inspite of its drawback of the typeⅡerror control, it has great values of practical application due toits high efficiency and intelligence.2. In connection with the shortcomings of "few features, low accuracy" which is existed inthe classification of fusing point clouds and imagery, the emphasis is placed on the optimizedextraction strategy for point clouds. According to the different principles of extraction, thefeatures for point clouds are divided into direct and indirect features. The way of adding localgeometric properties of raw point clouds to feature vector to improve the classification accuracyand stability is proposed, and five kinds of features belonging to normal vectors are extracted.The analysis of feature images indicates that the normal vector can not only distinguish roads, buildings with different shapes and vegetation, but also have a strong sensitivity in recognition ofvehicles and vegetation.3. SVM is introduced to the classification of fusing point clouds and imagery, two types ofmulti-features classification modes which are point clouds complemented by imagery andimagery complemented by point clouds are developed. The classification results show that ourmethods exhibit great adaptability to various types of data, and both of the modes achieve highaccuracy. In particular, the overall classification accuracy and Kappa of point clouds reach above90%, and the vehicles are successfully identified as a special kind of object. The contrastexperiments prove that the methods developed in this thesis evidently outperforms the traditionalclassification methods, no matter on the accuracy or on the quantity of land cover types.4. Combined methods of pros and cons are conducted to analyze deeply the feature spacesassociated with the two classification modes of the point clouds complemented by imagery andthe imagery complemented by point clouds, the conclusion are considered as follows: Firstly, inthe mode of point clouds complemented by imagery, normalized height(NH) of point clouds hasthe greatest impact on the classification accuracy; Secondly, after adding the local geometricproperties of point clouds, the misclassification probability between buildings and trees reducesconsiderably, and the problems in extracting building edge points are tackled; Thirdly,comparative analysis of the two classification modes indicate that point clouds contribute a lot tothe imagery classification, which is far superior to the mode of point clouds complemented byimagery, so the classification mode of imagery complemented by point clouds presents furtherpractical significance; Fourthly, combined classification technologies of land cover, whosecomplementary data are airborne LiDAR point clouds, will have important strategic status in thefuture.5. A nonlinear Mode filter is designed and improved to enhance the classification resultquality of point clouds and imagery. According to the different characteristics of point clouds andimagery, k-nearest neighbor and window-based Mode filters are respectively developed toremove speckle and salt and pepper noises. Contrast experimental results demonstrate that theimproved Mode filters can boost the classification accuracy of point clouds and imagery dataeffectively.6. The method of refined classification in image with the support of point clouds elevationdata is proposed and carried out. In order to create high accuracy of subdividing the same kind ofland cover type, four factors are taken into consideration, which includes supplementary datasource, Mode filter, point clouds density and image spatial resolution and point clouds firstechoes. Decision tree is developed to improve remarkably the classification quantity of buildingsand vegetation, which represents further superiority of classification of fusing point clouds and imagery, and achieves the desired goal of the unity of classification accuracy and quantity.7. The method of classification of tree species in point clouds data complemented bymulti-spectral imagery is proposed and carried out. In order to avoid the low accuracy caused byinadequate features, besides the direct features, such as normalized height, intensity, and so on,this thesis focuses on extracting the indirect features from point clouds which can depict thecanopy information of tree species, including the echoes of canopy, geometric properties ofcanopy, vertical structure of canopy. The experimental results show that, by synergistic use ofairborne LiDAR point clouds and multi-spectral imagery, the method based on SVM classifierobtains higher classification accuracy and more accurate categories of tree species.更多还原