高分辨率遥感影像多特征稀疏表达与智能理解方法研究

High-resolution remotely sensed data can provide a large amount of ground information in the multiple temporal-spatial-spectral dimensions and present accurate, huge, and complicated geographical patterns. Therefore, it has become one of the basic information sources for landscape analysis, environmental evaluation, rapid response, and resources surveying. With the spatial resolution increased, the spatial correlation between neighboring pixels is substantially enhanced and the semantic-visual information becomes more significant. In this context, the traditional low-level feature extraction and classification techniques are inadequate for obtaining high-accuracy geospatial information. Consequently, the primitive-semantic feature representation and automatic image understanding can be regarded as the core scientific problem of the huge-volume remote sensing image interpretation. Accordingly, in this project, we propose a series of innovative models to solve the aforementioned scientific topic, including: multilevel primitive feature extraction, feature-semantic reasoning, multifeature sparse analysis and fusion, and semantic scene recognition and understanding. This research is potential for improving the intelligent processing level of high resolution remote sensing data and is expected to exceed the current technological bottlenecks. It has important academic value and practical significance in high resolution remote sensing.

高分辨率遥感数据给我们提供了丰富的地表多时-空-谱信息，展现了精细、海量、复杂的地理模式，已成为景观分析、环境评价、快速响应、资源调查的基础数据源之一。随着分辨率的提高，影像中像素空间关系显著增强，语义视觉信息更加明显，传统的低层特征提取与影像分类技术无法获取高精度的地理空间信息。因此，高分辨率遥感影像的基元-语义特征表达与智能化理解，是解决海量影像自动高效解译的关键科学问题，也是目前遥感科学技术的学术前沿。本项研究围绕这一科学问题展开研究，原创性的提出了多层基元特征提取、特征-语义推导模型、多元特征稀疏流行分析与融合、语义场景认知与主动学习等原创性的方法。研究成果能显著提升当前高分辨率遥感影像的智能化处理能力，有望突破当前的技术瓶颈，具有重要的学术价值和现实意义。