基于时空一致性及组稀疏多核学习约束的遥感图像迁移方法研究

With the development of the ability of obtaining high temporal resolution remote sensing image, how to perform automatic classification based on both the historical images and the existing classification results has become an urgent issue. The solving of this issue will be key to the operational process of the high temporal resolution remote sensing satellite and be of great significance to the specific target monitoring. This project aims to utilize the transfer learning methods, which are widely studied in the machine learning community. To make the learning more effective, the characteristics of the high temporal resolution remote sensing images are considered: the temporal-spatial consistency, and various different image features. The feature extraction procedure will be studied based on group sparsity multiple kernel learning. Besides, the transfer learning method will be studied based on temporal-spatial consistency. The technology will be tested in the land monitoring application. Automatic classification of key elements of land monitoring will be performed based on existing historical data base and the newly-obtained image. In summary, the technological breakthrough of this project is expected to provide new solutions to the remote sensing monitoring applications, e.g., national land automatic monitoring, and automatic update of the land cover classification results.

随着髙时间分辨率遥感图像获取能力的提高，如何结合历史图像及分类成果信息对新获取的图像进行自动的分类成为一个迫切需要解决的技术问题。该技术问题的解决为髙时间分辨率遥感卫星的业务化运行，从而进行特定目标的监测具有重要意义。本项目拟采用机器学习领域广泛研究的迁移学习的思路，考虑髙时间分辨率遥感图像特有的时空一致性特点，并充分利用该类图像具有多种区分性特征。研究基于组稀疏约束的多核学习技术进行多特征融合研究；在其基础上，重点研究时空一致性约束下的迁移学习算法模型。并针对国土监测应用，结合已有的基期历史数据以及新获取的图像，进行国土重点监测要素的自动提取应用示范。该技术的突破有望为国土自动监测、地表覆盖分类成果的自动更新等遥感监测应用提供新的思路。

针对研究目标，本项目分别在分类样本的优化、特征的提取以及分类结果的迁移策略等三个方面对课题进行了探索。在样本的优化方面，继续探索了主动学习算法在样本定义的问题：提出了一种基于稀疏表示的主动学习方法；提出了一种基于最大子图优化的主动学习算法模型；并研发了基于主动学习算法的样本采集网站。在特征提取方面，主要探索视觉词包特征在高空间分辨率遥感图像的表达能力以及迁移能力：提出了一种基于特征重要性的多词包模型的场景分类方法；并发布一份高分辨率场景分类数据集。在迁移策略方面，首先研究了基于深度特征的高空间分辨率遥感图像迁移学习策略；并进一步探索了基于地理标记图像的遥感图像迁移学习策略；同时课题组整理了两组地理标记图像数据集，以及相应的两组Landsat数据，用于研究基于数码照片获取的信息迁移到中分辨率遥感图像的地表覆盖分类问题。