基于知识迁移的高分辨率遥感地表覆盖信息自动变更方法研究

One of the current hotspots in the researches of remote sensing cognition is to develop intelligent technologies for remote sensing information extraction by associating with the historical archived knowledge under the geographic background. In this study, a knowledge transfer-based model for cognizing spatial-spectral remote sensing information will be proposed. In this model, historical knowledge is gradually transfered into each step of land-cover information update through multi-level computation. It aims to construct a automatic procedure of land-cover information update with the technological implementation of “constructing objects by segmentation-extracting change by detection-updating attribute by classification”. Specifically, in order to improve the degree of automation and the level of intelligent of land-cover information update by temporal-spatial expansion and transfer learning of the spatial distribution knowledge of land types, we will carefully investigate several technical difficulties in the interpretation process of high-resolution remotely sensed imageries and focus on the following pivotal aspects: (1) mean-shift-based multi-scale segmentation under the guidance of previous land cover information; (2) object-oriented change detection using coupled analysis of spatial-spectral features; (3) fully automatic object-oriented classification by combining the methods of change detection and knowledge transfer. In this research, typical areas in East China and West China will be selected as demonstration regions. Finally, we will design and develop a high-performance software prototype, and demonstrate how to rapidly update regional land-cover information with high-resolution remotely sensed imageries from Chinese high-resolution satellites. Generally, this research will provide technical support and decision foundation for land resources monitoring under the background of China’s new urbanization construction.

关联历史存档的地理背景知识来发展智能化遥感信息提取技术，是当前遥感认知领域一个非常活跃的研究方向。本项目拟提出一种基于知识迁移的遥感信息图谱认知模型，力图通过多层次分步计算将历史专题知识逐步融入至地表覆盖信息变更的各个技术环节中，以期实现以“基元分割－变化检测－分类更新”为技术体系的地表覆盖信息自动变更流程。针对高分辨率影像解译过程中的难点问题，重点研究“基期土地覆盖信息指导下的均值漂移多尺度分割”、“图谱特征耦合分析的对象级变化检测”以及“协同变化检测与知识迁移的对象级全自动分类”三大关键算法，旨在通过地类空间分布知识的时空拓展与迁移学习提高地表覆盖信息变更的自动化程度与智能化水平。项目将选取我国东西部典型区域为研究区，以国产高分辨率影像为主要数据源，搭建高性能软件原型系统，并开展土地覆盖信息自动变更应用示范，为国家新型城镇化建设背景下的土地资源监测提供技术支持和决策支撑。

高分辨率遥感与传统遥感对于应用而言的优势在于其精准特性，也即其呈现的“精细探知、真实检验、全体遍历、动态可控”的大数据特征。将“智能化”机器学习技术与“精细化、定量化、动态化”地学图谱分析思想进行紧密结合，开展遥感信息图谱认知研究正是本基金项目探索的主题，目标就是要在知识迁移机制下为发展“精准” 而又“高效”的遥感影像地学理解工程技术提供支撑！经过三年的持续深入研究，本项目首先综合了高分辨率遥感影像“图结构”与“谱特征”的双重特征优势，提出了“图-谱”耦合的遥感认知与协同计算理论，取得了以学术专著形式体现的原创性基础理论成果；其次，重点围绕“基元提取－变化检测－分类更新”等技术环节，提出了面向精细化土地信息分析的高分辨率遥感影像地块提取与指标反演的方法，突破了土地利用地块智能生成、多源外部知识融入的样本自动优选与迁移学习、中分时序数据处理与重建、地块内覆盖类型判别与指标计算、地块多模态信息融合与功能属性推测等一系列关键算法，通过多层次分步计算将多源知识逐步融入至地表覆盖信息提取流程中，取得了一批学术论文与技术专利等的创新研究成果；最后，研发了地表覆被信息智能提取与变更的软件原型，并通过典型区域的应用示范，验证了土地覆被因子提取过程中的“图-谱”特征逐步融入与外部知识迁移机制，提高了地表覆盖信息自动变更的智能化水平与精准化程度，为大数据时代高分遥感精准服务模式的创建提供了技术支持。