基于多视图学习的高分遥感影像和LiDAR数据城市不透水面提取研究

Impervious surface mapping with high spatial resolution remote sensing image and airborne LiDAR data can take advantage of observational features of different mechanisms and obtain the distribution of urban fine-scale impervious surface. Most traditional methods utilize a single model to analyze the stacked data of high spatial resolution remote sensing image and airborne LiDAR data, without considering their specific physical mechanisms and statistical distributions. At the same time, the effect of landscape changes caused by multi-source and multi-temporal high spatial resolution remote sensing data were also ignored in practical problems. Therefore, the performance of impervious surface mapping is limited. In order to solve the aforementioned problems, this project will develop an accurate impervious surface mapping method based on multi-view learning with high spatial resolution remote sensing image and airborne LiDAR data. Firstly, this project proposes a multi-source change detection method based on multi-view learning, which can generate the change mask accurately, and provide unchanged training samples for multi-view feature extraction. Then, a multi-view learning model based on deep canonical correlation analysis and deep autoencoder is proposed to extract multi-view discriminant features with the same statistical distribution. Finally, multi-view discriminant features are combined by D-S evidence theory to map the impervious surface accurately. The research findings of this project will provide a new and effective way for urban fine-scale impervious surface mapping.

融合高分辨率遥感影像和机载LiDAR数据的不透水面提取能够利用不同机理的观测特征，提取城市精细尺度不透水面分布信息，具有重要的研究意义。现有方法大多采用单一模型将高分辨率遥感影像和机载LiDAR数据叠加分析，缺乏对多源数据独特物理意义和统计分布的深入挖掘，忽略了实际问题中数据获取时间不同所带来的地物变化影响，限制了不透水面提取精度的提升。本项目将围绕多视图学习理论，发展基于多视图学习的高分辨率遥感影像和机载LiDAR数据城市不透水面高精度提取方法。本项目首先研究基于多视图学习的多源遥感数据变化检测方法，精确掩膜变化地物，为多视图特征提取提供训练样本；然后，构建基于深度典型相关分析和深度自动编码机的多视图学习模型，提取具有统计一致性的多视图判别特征；最后，采用D-S证据理论融合多视图判别特征，实现不透水面的精确提取。本项目的研究成果可为精细尺度下的城市不透水面提取提供一条新的有效途径。

本项目融合高分辨率遥感影像和机载LiDAR数据的多源信息，开展了多源遥感数据变化检测、特征提取、信息融合与不透水面提取等方向的研究，项目进展顺利，达到预期目标，共发表SCI期刊论文4篇、EI会议论文1篇、已接收SCI期刊论文1篇，包括中科院一区SCI期刊论文1篇、二区SCI期刊论文2篇、三区SCI期刊论文1篇，其中3篇为第一标注资助项目。在人才培养方面，项目负责人获得副教授职称，项目组成员1名硕士顺利毕业、4名硕士顺利开展学习研究工作。本项目具体取得的成果包括：（1）针对多时相多源遥感影像数据结构的不一致性，本项目提出了基于典型相关分析的多源变化检测算法，并实现不同时相高分辨率光学和LiDAR影像的信息融合，精确提取不透水面分布；（2）针对遥感影像中地物特征的复杂性问题，本项目将迁移学习和测度学习结合，提取更加具有判别性和通用性的影像特征；（3）针对多视图信息无法融合的问题，本项目提出了基于D-S证据理论的信息融合策略，在高分辨率影像变化检测任务上取得了较好的效果；（4）针对城市不透水面信息受阴影影响、难以精确提取的问题，本项目提出了结合线性和非线性光谱混合模型的不透水面提取方法以及基于时间序列影像的不透水面提取方法，增强了城市不透水面分布提取能力。综上所述，本项目严格按照计划顺利实施，成功达到项目研究目标，取得预期成果。本项目的研究成果能够提高城市高分辨率光学遥感影像和LiDAR数据的融合处理能力，为我国城市可持续发展与科学规划提供技术支持。