高分光学遥感图像道路提取对国家安全、交通网建设等具有重要意义，在智能导航等民用领域也有重大应用价值，但面临高人工标记代价、成像缺失鲁棒性不足的问题。主要原因是缺乏高效训练集获取研究，有监督与无监督结合不足以及深层语义与浅层表观特征融合不充分。为此，本课题围绕高效训练集获取，深层语义与浅层表观特征深入融合，有监督与无监督结合框架三个方面展开研究。创新点包括：（1）拟提出结合历史GIS数据和基于运动车辆标记的道路自动标记方法，高效获取高质量道路提取训练集；（2）拟提出超像素分割和深度特征融合的无监督道路提取，提升无监督模型的效率和鲁棒性；（3）拟提出深入融合浅层表观特征与深层语义特征的道路同步提取与修复的深度网络，克服由成像缺失导致的提取断裂；（4）综上，拟提出多模型多层级的有监督与无监督组合框架，实现高效鲁棒的道路提取。研究将为基于高分遥感道路提取的理论和应用实践提供重要的补充和推动作用。

Automatic road extraction from high resolution optical remote sensing images not only is significant for national security, road network construction etc., but also has great application values for civilian areas such as intelligent navigation etc.. However, current methods suffer from hard burden of manual labeling work for training models and lacking of robustness to missing in imaging. The major reasons for the above two problems are: (1) lacking of research on high efficiency training dataset acquisition; (2) insufficient mixture of supervised and unsupervised models; (3) insufficient fuse of deep semantic features and shallow appearance features. To solve the above problems, this research project focuses on three aspects: (1) the high efficiency of training dataset acquisition; (2) the deep fusion of deep semantic feature and shallow appearance features; (3) the combination framework of supervised and unsupervised models. This research project contains four innovation points. First, this project will propose an automatic road label extraction method based on moving vehicles’ labels and history GIS data to obtain high quality road extraction training dataset with high efficiency. Second, this project will propose an unsupervised method which embeds superpixel segmentation and deep learning features for road extraction, improving the computational efficiency and robustness of unsupervised methods. Third, this project will propose new deep networks which can extract road areas and repair extraction breakages at the same time, conquering the extraction breakages resulted by missing in imaging. In the proposed networks, the deep semantic features and shallow appearance features are deep fused. Fourth, this project will propose hierarchical and multiple models mixture framework of supervised and unsupervised models for promoting the robustness and efficiency of road extraction framework based on the above researches. This study will supplement and promote the theory and application practice of road extraction based on high resolution optical remote sensing images.