融合高光谱与LiDAR数据的城市地物精细化分类

The mixture of artificial and natural targets on urban surface make the methods based on single remote sensing data source have less classification accuracies, and a number of categories is difficult to distinguish. The deep integration of geometric, spatial and spectal features in multi-source remote sensing data provides the possibility of solving the above problems. In view of the classification chanllenges from the same spectrum from different materials, the same materials with different spectrum, shadow and height, a new model of hyperspectral and LiDAR data fusion is established. First, the spatial and spectral information from multi-source remote sensing data is used to construct a space-spectal joint manifold perception graph to reveal the nonlinear distribution of the data. Based on this joint manifold graph, multi-modal features are extracted for high-precision classification by the semi-supervised learning theory. The project intends to introduce the latest theoretical results in the fields of sparse representation, semi-supervised learning and mathematical morphology, and builds spectral and spatial low-rank semi-supervised learning models with urban remote sensing data, to extract multi-mode semi-supervised features for urban refinement classification. The research results from this project will effectively enhance the computer interpretation ability about urban remote sensing data, and improve its potential application in urban planning and management.

城市地表人工与自然目标混杂，单一来源遥感数据往往分类精度不高，若干类别难以区分。多源遥感数据中几何、空间、光谱等特征的深度融合为上述问题的解决提供了可能。针对城市同谱异物和同物异谱现象、阴影和高度直接影响地物分类等难点，研究建立高光谱与LiDAR数据融合的新模式，先利用多源遥感数据的空间信息和光谱信息构建出空-谱联合流形感知图，以揭示数据分布的非线性结构，基于此感知图提取出多模态特征，在半监督学习理论支持下进行地物精细分类。项目拟引入稀疏表达、半监督学习和数学形态学等领域最新的理论成果，构建城市遥感数据谱域和空域的低秩半监督学习模型，实现城市地物多模态半监督特征提取和精细化分类，研究成果可有效提升城市遥感数据的解译能力和在城市规划管理中的应用潜力。

基于多源数据融合技术的城市复杂地表解译分析是城市遥感领域的一个重要研究方向，本项目以城市为研究对象，以多源数据为基础，研究城市地物精细化分类。根据项目计划，已经完成了下列工作：（1）采集和处理了多个研究示范区的同一范围的多源遥感数据，包括高分卫星遥感数据，机载多光谱数据，机载雷达数据，无人机三维倾斜摄影数据等，并且对多源数据进行了标准化处理；（2）构建出了空-谱联合半监督流形感知图，能够挖掘高维遥感数据的流形结构，并基于流形结构提取出遥感数据易判别的多模态特征，实现城市地物精细化分类；（3）基于遥感数据对城市地物精细化分类的结果，融合城市POI大数据，设计深度神经网络模型，对广州、深圳、杭州、西安、武汉、成都这6个典型城市进行了城市边缘区的提取，将城市分为核心区，边缘区和外缘区的划分，定量分析了不同城市的城市化进程。本项目研究结果有助于推进多源数据融合技术的发展，提升城市遥感数据的解译能力，可为城市规划、城市统计分析和城市生态人居环境政策制定等提供参考依据。