面向高分辨率遥感影像道路提取的深度特征融合模型与方法研究

As more and more high-resolution remote sensing satellites are launched, we can obtain more and more high-resolution remote sensing images. It is a major issue that how we extract the road information from these massive data accurately and efficiently. The importance of road extraction for multi-scale and multi-resolution fusion images is increasing. The existing methods are not well extracted under occlusion and shadow conditions, and the structural features and multi-context features of roads cannot be effectively modeled. Therefore, this project aims at solving the challenges of the road extraction in high-resolution remote sensing images, which is based on the fusion of deep features. The first one is the research on the road extraction based on Multi-granularity deep neural network model combining local and global features. The second one is the research on road extraction based on hybrid deep neural network, in which road network structure and semantic information are considered. The last one is road extraction based on multi-context joint optimization convolutional neural network, which well integrates the cross-layer context, global image-level context, within-super-pixel context and cross-super-pixel neighborhood context into a unified network. With the researches of above topics, we will propose a series of road extraction models, which make the extraction performance efficient and accurate. The models studied in this project have strong versatility and can be applied to high-resolution remote sensing images of different application scenarios. The implementation of the project can provide a basis for the cross-development of high-resolution remote sensing images and deep learning.

随着越来越多的高分辨率遥感卫星的发射，可以获得越来越多的高分辨率遥感影像。如何从这些海量影像数据中自动、精确和高效地提取道路并对其加以利用，则成为必须解决的重大问题。针对面向多尺度和多分辨率融合影像中道路提取的重要性逐渐增加、现有方法在遮挡和阴影条件下提取效果不佳以及无法有效地对道路的结构特征和多上下文特征进行自适应建模等问题，开展基于深度特征融合的道路提取研究。包括：1）为充分利用多尺度和多分辨率信息，研究融合局部与全局特征的多粒度深度神经网络模型；2）提出面向道路网结构和语义信息的混合深度神经网络模型；3）建立基于多上下文联合优化的道路提取模型，充分利用多种多样的上下文信息和相互语义关系信息，从而在保持道路提取精确度的同时减少后处理步骤。本项目研究的模型和方法具有较强的通用性，可以适用于不同应用场景的高分辨率遥感影像。该项目的实施可为高分辨率遥感影像与深度学习的交叉发展提供依据。

道路网络是交通基础设施的重要组成部分，同时也是地理国情的重要基础数据。完整、及时、准确的道路网络，在日常出行、电子商务、物流配送等都具有广泛的社会需求，同时也在社会经济活动中发挥着重要作用。准确、可靠、快速的道路提取算法对于城市道路网信息更新、突发事件应急、交通基础设施现状评估等具有重要应用价值和社会意义。本项目主要内容包括：针对遥感图像中带标签样本数量有限的问题，重点研究了基于半监督的分类方法，尤其是基于集成投影的图像分类方法，然而它存在一些不足之处，为此，提出基于高层特征选择、MRF与岭横向方法的高分辨率遥感图像道路中心线提取算法，提出基于局部统计频域信息的鲁棒多模态遥感图像匹配算法。此外，还研究了和遥感图像道路提取相关的图像分类、人体行为识别和轻量级网络等内容，重点突破基于深度神经网络进行道路提取的关键技术。本项目对提高高分影像道路提取技术和多尺度空-谱信息融合的集成应用具有重要的理论价值；也对准确、可靠、快速的城市道路网信息更新、突发事件应急、交通基础设施现状评估等具有重要的应用价值和社会意义。