基于对象的高分辨率遥感地表覆盖信息精细提取及其尺度效应分析

It is of theoretical significance and practical value to use remote sensing technology to accurately and efficiently extract the land cover information, which is essential for qualitatively and quantificationally analyze the geographic background and even for enhancing national capability of natural resource management and global change responsing. Using Chinese high spatial resolution satellite images to accomplish this task is specially meaningful for promoting the development of National Geographic Condition monitoring and perfecting the national satellite information application chain. However, current research is suffering from limitation of accuracy due to the lack of mechanism cognition on human visual perception and scale effect. This project will use Chinese high spatial resolution satellite images and follow the rules of human visual cognition to concentrate on the following work. First, based on local single scale optimizations and multi scale information synthesis, this project will tentatively extract the accurate land cover parcels as what we see. Second, a adaptive high performance object-base classifier with adaptive feature selection will be further explored. Finally, this project will apply the methodology to practice by township level case study. Through this project, it is desired to deeply expand the research on remote sensing scale issues and modifiable area unit problem. Also, the precision and accuracy of extracted land cover information is expected to be methodologically enhanced, so that the application chain of national high spatial resolution satellite data, capturing-processing-analyzing-application and service, will be further improved.

基于国产高分卫星数据对地表覆盖信息进行高效率精细化提取，定性定量地描述地理本底分布，对提高我国资源管理和应对全球变化的能力，推进我国地理国情监测业务进展，完善国产卫星测绘应用链条具有重要的意义。然而由于对人类视觉认知机理和尺度效应机理的理解和利用不足，目前的研究手段难以实现与人类视觉相一致的地表覆盖斑块高精度提取。本研究将采用国产高分卫星数据，遵循人类视觉认知本质规律，突破基于多尺度局部优化与尺度综合的“所见即所得”的地表覆盖图斑精细化提取问题；构造具有自适应特征优选功能的高精度高自动化的影像对象分类器；在应用上，以乡镇级地表覆盖高分遥感信息提取为示范任务开展应用研究。通过课题研究，力争实现从理论上拓展遥感尺度和可塑性面积单元问题研究的深度；从方法上提高基于对象影像分析的精细度和精确度；从应用上服务于我国地理国情监测，推动我国国产高分卫星数据“获取-处理-分析-应用服务”链条的不断完善。

针对地表覆盖信息高精度自动提取问题，本项目进行了如下研究：在理论上，建立了“先全局粗略分割、再局部精细提取、后全局优化综合”的信息提取模型，拓展了遥感尺度和可塑性面积单元问题认知研究的深度；在方法上，探索了面向对象影像分析及深度学习卷积核模板的尺度自适应问题，解析了面向对象影像分类尺度效应机理，并突破了面向对象卷积神经网络分类的卷积核位置自动生成技术难题，提高了面向对象影像分析的精细度和精确度；在应用上，将以上理论与方法应用于地表覆盖调查、专题信息提取及城市功能区识别等工作中，示范研究表明本项目提出的地表覆盖信息自动提取理论与方法一定程度上提高了面向对象信息提取的效率和精度，在自然资源调查及监测方面有广阔的应用前景。基于以上研究，本项目已在国内外著名学术刊物发表学术论文21篇，取得授权发明专利及软件著作权各1项。