“图-谱”耦合迭代的地表覆被因子自适应遥感提取方法研究

One of the "hotspots" in the remote sensing cognitive researches is to develope intelligentialized data mining technique for remote sensing information extraction, which can be done by fully exploring the Spatial-and-Spectral features possessed uniquely by remote sensing imagery.In this study,a self-adaptive remote sensing model for extraction of land cover factors was proposed based on spatial-spectral integrated iterative approach. Synergic cognization of multi-layered knowledge,such as spectral feature, spatial feature and spatial relationship etc., can be achieved through adopting sub-step with layers an iterative computing method. It aims to improve the ability of automatic information extraction and,then furtherly consummate the remote sensing cognitive theory represented as a hierarchical structure,i.e.,"Spectral single pixel - Feature element - Spatial-spectral combined structure - Landscape pattern - system function evolution". Specifically, we will focus on the following pivotal aspects: the comprehensive representation of "spatial-spectral" feature, automatic selection of the optimal sample points and the precise identification of feature-lacking objects and fuzzy self-adjusted optimization mechanism.After adequately addressing these problems, the accumulated errors in calculating link are under well control and the problems of deficient feature and computing unsteadiness commonly existed in automatic extraction models are also overcome ,thus consequently improving the accuracy of information extraction. In this case, Xinjiang Kaidu River basin is chosen as a demonstration region in which automated classification scheme of land cover factors specifically for arid zone will be established. The land cover factors in this basin,i.g., glacier, snow, lake，river，oasis and etc., then will be automatically extracted based on remote sense imagey. As a result, this will definitely provide technique and decision supports for ecological environment monitoring in Chinese arid region.

充分挖掘遥感影像"图"、"谱"特征来发展智能化遥感信息挖掘技术，是当前遥感认知研究领域所关注的热点问题。本项目拟提出一种基于"图-谱"耦合迭代的地表覆被因子自适应提取模型，力图通过多层次分步迭代计算实现遥感影像波谱特征、空间特征以及空间关系等多个层面知识的协同认知，以提升自动化信息提取的能力，进而完善"波谱像元－特征基元－图谱结构－空间格局－功能演化"为体系的遥感认知理论；重点研究"图-谱"特征综合表达、特征样本自动优选、特征残缺地物精确识别以及模糊自适应优化调整等关键方法, 有效控制计算过程中的误差积累问题，克服当前遥感信息提取中的特征不完备和计算难稳健的问题，以提高信息提取的精准程度。项目将以新疆开都河流域为应用示范区，建立干旱区地表覆被因子的自动化分类流程，开展流域内冰川、积雪、河流、湖泊、湿地、绿洲等地表覆被因子的遥感自动提取，为干旱区生态环境监测提供技术支持和决策支撑。

对于应用而言，高分辨率遥感较传统遥感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“遥感认知与计算”研究正是本基金项目探索的主题，目标就是要为发展“精准” 而又“高效”的遥感影像地学理解工程技术提供一套支撑理论！本项目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，按照空间、时间与属性的维度构建了“由谱聚图-图谱协同-认图知谱”三段论组合的新型认知体系，通过“粒化-重组-推测”的计算逻辑对精准地物“形态-类型-指标-功能-演化”进行了逐级深入的智能挖掘与定量分析，取得了以学术专著形式体现的原创性基础理论成果；提出了面向精细化土地信息分析应用的高分辨率遥感影像地块提取与指标反演的技术方法体系，并按照多层感知、时空协同与多粒度决策上下协同的模型设计，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，解决了遥感提取过程中的特征不完备及计算不稳健的问题，取得了一批学术论文与技术专利等创新研究成果；研发了地表覆被信息智能提取软件原型（PLAC），并通过西部干旱区湖泊冰川制图和东部农业区种植规划制图的应用示范，发展了多层次迭代的自适应计算技术，实现了“图-谱”特征逐步融入与外部知识迁移机制，有效控制了信息传递过程中误差积累问题，提高了覆被因子提取的智能化水平以及地物识别的精准化程度，探索了针对不同应用目标进行专题制图的定制化技术，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。