数据驱动的高光谱遥感影像特征表达、迁移学习及其在城市地理信息提取中的应用

高光谱遥感影像的自动化处理是人们普遍关注的研究方向。高光谱遥感信息处理的关键问题在于：1）高维数据的有效表达，既能紧凑表达特征的同时又能保持光谱-空间的结构关系；2）光谱的准确匹配，克服由于光谱变化导致同类地物不能匹配的难题；3）与数据无关的信息模型，使高光谱数据的解译具有泛化能力。针对这三个问题，本项目从数据-特征-模型-信息的自学习和互反馈的思路出发，围绕一体化特征表达、多层次迁移学习、地物信息的互反馈机制开展深入研究，并进行城市地区典型地物的地理信息提取的实用化验证，形成数据驱动的高光谱遥感信息处理理论方法。本项研究成果不仅可以提升高光谱影像信息处理的解译能力和自动化水平，也对推动我国高光谱遥感技术的广泛应用具有重要的现实意义。

The automatic processing of hyperspectral remote sensing images (HRSI) is world-widely under intensive study. The key issues of the HRSI information processing consist of: 1) the efficient description of data with very high dimensionality preserving both spectral and spatial features of HRSI; 2) the accurate spectrum matching in HRSI taken spectral variability into consideration; and 3) the data-independent information processing model allowing high capability of generalization for HRSI interpretation. Starting from the self-learning and the feedback of the “data-feature-model-information” processing chain, our research will focus on the integrated representation of data features, the multi-level transfer learning and the mutual feedback mechanism of geographic information. Validations on the geographic information extraction of the typical urban areas will be carried out. The outcomes of this project could not only improve the interpretation capability and the automation level of HRSI information processing, but also promote the Chinese hyperspectral remote sensing technology.