【作者】 李波；

【导师】 李汉武； 黄昕；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2017， 硕士

【摘要】 近年来,城市化进程不断加快,城市的发展规划越来越需要地理空间信息的支撑。随着遥感平台和传感器的不断进步,遥感数据朝着多元化、多极化的方向发展,从而展开了遥感对地观测的新纪元。光学遥感影像是城市遥感的主要数据源之一,其能方便、直观地展示空间格局地物分布。近年来,光学遥感影像朝着高光谱分辨率或高空间分辨率的方向发展,能够获取更加丰富的地物光谱信息和空间特征。但是,更高的光谱分辨率和空间分辨率并不一定能带来影像解译结果的精度提升。由于同物异谱和异物同谱等现象的存在,高光谱影像分类精度存在一定的瓶颈。而高空间分辨率虽然提供了详细的地物细节信息,但是过高的空间分辨率也使得类内方差增大类间方差减小,并最终导致光谱可分性降低。因此,采用光学影像单一数据源不足以完成城市区域高精度影像分类任务。机载激光雷达作为一种新型对地观测技术,可以快速、准确地获取被观测区域的高程信息,弥补二维光学影像的信息不足,因此在城市遥感任务中得到了迅速推广。然而,激光雷达的数据组织格式是一种离散三维点云数据,无法直接与二维影像特征进行信息融合。因此,探索多源遥感数据的互补融合方法,是遥感领域的前沿研究和重要应用方向之一。多源数据融合分类包括像素层、特征层和决策层三个不同的融合层次。其中,像素层融合是多特征之间的简单堆叠,然后输入支持向量机或随机森林等分类器;特征层融合考虑对不同特征进行有效组合,以实现地物识别和分类;决策层融合是在像素层和特征层融合的基础上进行投票,从而提高分类结果的可靠性和精度。本论文的研究目标是利用机载激光点云数据和高光谱遥感影像进行多源数据融合分类研究,以提高多源数据分类结果精度。论文首先针对激光点云和高光谱影像的数据特点,分别提取了高程、光谱和空间等有效特征信息,并分析了不同特征信息对不同地物类型的判别效果。此外,论文首次提出了融合点云伪波形特征和高光谱数据进行空间特征提取的方法。伪波特征与高光谱数据进行波段堆叠可以增强光谱可分性,而基于光谱角距离的像元形状指数可以充分挖掘组合特征的光谱-空间特征信息,从而获取更加可靠的像元形状指数特征。在特征提取的基础上,我们利用多源特征库进行波段组合,输入支持向量机或随机森林分类器进行像素级融合分类,并评定分类精度。然后,我们利用高程信息辅助光谱特征数据进行多尺度影像分割,获取影像对象。最后,利用像素级融合分类结果进行面向对象的决策层融合分类。论文采用一组激光点云数据和高光谱遥感影像验证特征提取和融合分类效果。实验结果表明:1)以伪波特征和高光谱影像作为基影像的像元形状指数特征相比原始像元形状指数特征在分类过程中效果更好;2)在高程信息辅助的情况下,基于支持向量机和随机森林的像素级融合分类相比原始光谱分类,其精度分别提高了 10%和12%左右;3)在面向对象的决策层融合分类试验中,采用高程辅助的分割方法相比基于光谱的分割方法取得了更好的分类效果,我们所提出的面向对象决策层融合分类则取得了 94.05%的分类精度,相比支持向量机和随机森林两种像素层分类方法分别有了 6%和4%左右的精度提升。更多还原

【Abstract】 In recent years,with the accelerating process of urbanization,geospatial information is increasingly demanded for urban planning and development.With the continuous advances of platforms and sensors,remote sensing data is developing towards multi-sources and multi-direction,which launched a new era of earth observation.Optical imageries are one of the main data sources of urban observation,which can be convenient and intuitive to show spatial pattern distribution.In recent years,optical imageries have developed towards the direction of hyperspectral and high spatial resolution,which can obtain richer spectral information and spatial features.However,this phenomena does not necessarily improve the accuracy of image interpretation.Due to the presence of spectral heterogeneity,there is a certain bottleneck in the accuracy of hyperspectral image classification.As the high-spatial-resolution,although providing many details of the object,but also makes the variance within the class increases the variance between the class decreases,and eventually lead to the decrease of spectral separability.Therefore,the use of a single data source is not enough to complete the urban classification with high-precision.LiDAR is a new type of earth observation technology.It can quickly and accurately obtain the elevation information which supplements the information of 2D optical image.Therefore,LiDAR is widely promoted in urban remote sensing task.However,the data format of LiDAR is a discrete point cloud,which cannot be directly fused with image features.It is one of the hottest research and most important application that fuse the complementary information of multi-source data.Multi-source data fusion includes three levels:pixel level,feature level and decision level.The pixel level fusion is a simple stack of multiple features,and employ the support vector machine or random forest classifier;feature level considers the effective combination of different features to achieve feature recognition and classification;decision level fusion is a comprehensive combination of features and classifiers,which can significantly improve the reliability and accuracy of classification results.The purpose of this paper is to fuse airborne LiDAR point cloud data and hyperspectral imagery to improve the accuracy of multi-source data classification.In this paper,the information of elevational,spectral and spatial features is extracted from LiDAR data and hyperspectral image.In addition,the method of spatial feature extraction is proposed,which combines the pseudo-waveform of LiDAR and hyperspectral data.The pseudo-waveform can be used to enhance the spectral separability of hyperspectral data,and then the spectral angle distance(SAD)is utilized for the pixel shape index(PSI),which can fully extract the information of fused data,and obtain a more reliable PSI feature.On the basis of feature extraction,we use the multi-source features for pixel-level fusion classification with support vector machine(SVM)and random forest(RF)classifier.Then,the elevation information is employed to assist the spectral and spatial data and carry on the multi-scale segmentation,resulting with image object.Finally,object-oriented decision fusion is carried out with pixel-level classification results.In this experiment,we use a set of LiDAR data and hyperspectral remote sensing image in the same region to verify the framework we proposed.The experimental results show that the SAD-PSI feature with pseudo-waveform and hyperspectral images is better than the original PSI feature in the classification.At the same time,the SVM and RF classification accuracy of pixel level fusion with elevation information is increased about 10%and 12%respectively,compared with the original spectral classification.In the object-oriented decision fusion classification experiment,the elevation-assisted segmentation method achieves better classification effect compared with the spectral-based segmentation method.And object-oriented decision fusion classification achieves an overall accuracy of 94.05%,with 6%and 4%higher than to the pixel level classification with SVM and RF,respectively.更多还原