基于压缩感知的机载激光扫描数据完好性检验及特征级融合

Research on the topic of LiDAR data processing and its application is now becoming a hotpot in remote sensing community. A lot of significant publications can be found accordingly in many research literatures. However, two main problems still exist in this field, showing as: 1) the lack of an integrated theory and methodology to detect and solve the data integrity problem of point clouds from LiDAR sensors, and 2) the lack of mechanism discussion on feature level fusion of LiDAR data. Since the existing fusion methods always concentrate on the pixel level fusion based on multi-sensor data, the above problems closely correlate with the innate problem of undersamping and heterogenity in LiDAR data. Actually, the undetermined problem is the essence of the above two problems. Fortunately, the compressive sensing (CS) theory from Applied Mathematics has been shown the capacity to explain the underdetermined problem of data acquisition and processing. Therefore, to solve the problem of integrity verification and feature level fusion for LiDAR data, this proposal aims to put forward an effective theory framework and its relevant methodology. As the basic theoretical research in this proposal, the principle of CS for LiDAR data is firstly presented to perform its multi-scale sparse representations and sparse reconstruction,which consists of the design of Measurement Matrix, the selection of optimal Sparse Matrix and the optimal algorithm for sparse reconstruction. Next, by seeking regular patterns of data missing from LiDAR sensors, we focus our study on the detection of data missing and its effective repairment with the help of surrounding point clouds under the constraints from certain features. With the basic theory above includes sparse reconstruction and Measurement Matrix for LiDAR data, the methodology of feature level fusion is established for LiDAR data and the spectral imagery. Accordingly, the evaluation of fusion results is studied. After that, to verify our theorety and methodology presented, one typical case study on the spatial distribution of wetland vegetation community in ChongMing island from Shanghai is carried out. And the quantative estimations on wetland biomass are achieved with the feature information extracted from fusion results via dimension reduction by manifold learning. Finally, the influence on practical applications from data integrity and feature level fusion of LiDAR data is summarized. Through the research work defined in this paper,our work in this proposal will significantly improve the usability of LiDAR data, which will greatly favor its further applications in feature extraction, fine 3D modelling, environment monitoring and so on.

国内外对LiDAR数据的处理及应用研究已形成热点。但现有数据处理研究存在两个问题：缺少相应的理论与方法对点云数据的完好性进行检测与分析；融合方法多侧重于多源遥感信息像素级融合，缺乏高层次融合机理的探讨。解决这两个问题的关键是对LiDAR采集中的欠定问题求解。本项目采用压缩感知理论对LiDAR数据采集中的欠定问题进行数学解释，提出有效解决数据完好性和特征级融合的理论框架和方法。采用压缩感知理论实现点云的多尺度压缩与重构，检测数据中的缺失区域并寻找缺失规律，实现数据缺失补充，提出点云的完好性检验方法；依据压缩感知的变换基不相关的观测矩阵和信号重建理论，建立LiDAR数据与光谱影像的特征级融合准则并实现融合质量的综合评价；利用融合特征信息分析崇明湿地植被群落的空间分布规律及量化生物物理特征，分析数据完好性与融合质量对实际应用的影响规律。本研究将极大提高LiDAR数据在实际应用中的可用性。

本课题研究围绕激光扫描数据的欠定性和高层次融合机理缺失问题，针对激光扫描数据的完好性检验、点云与多光谱影像的特征级融合方法、多任务验证与原型系统构建等多项关键技术展开研究，给出了激光探测的高质量信号重构模型，构建了完整的特征级融合流程，验证了方法的有效性。.首先开展激光全波形数据的检校模型研究，给出了波形分解的新方法，构建了完好性评价体系，实现了高质量目标点云重建；在此基础上，研究激光点云的有效稀疏表达模型和实现方法，构建了高维点云特征空间，建立了多种稀疏化表达模型，评价了不同表达模型在分类方法中的效果；然后，研究了激光点云与多光谱影像特征提取方法，针对两种数据的差异性，给出了稀疏条件下的特征点识别策略，尤其针对多光谱影像提出了特征学习的降维方法，形成特征级融合输入条件；此外，还研究了高维特征空间下的高效率索引构建方法，给出了激光与多光谱异构高维特征的组织数据结构，提出了特征超平面分割和匹配方法，有效降低搜索算法的复杂度，保障了计算可行性；最后，给出了特征级融合的必要预处理和特征筛选方法，利用融合数据模型在典型场景开展验证，能够在湿地植被群落分类上获得较好的结果，结合多层分类器技术，能够在珍惜物种保护应用中获得单株级高精度定位结果，由此证明了数据模型对目标特征表征能力的提升和计算模型的有效性。.本课题研究验证了压缩感知理论在提升激光点云数据质量和构建高层次数据融合模型上的有效性和先进性，实现了从高质量信息获取、信息整理到高层次识别应用的完整链路，形成了激光点云数据生产和特征级融合分类的若干关键理论和作业流程，能够为新型点云处理模型和目标识别理论的构建提供有力支撑，在城市建模、林业勘察、海洋测绘、电力巡检、生态监测等行业应用中提供工程实践指导。研究成果申请发明专利5项，所取得的系列创新成果已在《Remote Sensing》、《Remote Sensing Letter》、《Journal of Applied Remote Sensing》等国际权威期刊上发表。