基于深度卷积网络的高分辨率遥感影像道路提取方法研究

Road extraction from very high resolution remote sensing images is a key issue in vehicle navigation, digital line graphic making and spatial geodatabase updating, and it is also a hot topic in academic research at the same. The shadows and occlusions of ground objects on remote sensing images, the interference of similar linear objects, and the uneven illumination cause problems such as road discontinuity, road misclassification, and low accuracy of centerline extraction. These are the main factors that constrain the robust extraction of roads. In view of this, this project firstly designed a generative adversarial network with the aggregation of localized context and texture for the road discontinuity problem, which is used to repair the broken road from data sources. On the basis of continuous roads, the loss function is constructed to increase the difference between road and similar linear objects and overcome the confusion problem. Finally, the multi-task learning method of road segmentation and centerline extraction is proposed to get rid of the dependence of the cascade prediction of centerline on the road segmentation and improve the accuracy of road segmentation and road centerline extraction. This project aims to achieve robust road extraction from very high resolution remote sensing images under complex background conditions. The research results will help to improve the automation of road extraction and have important research and economic value.

高分辨率遥感影像道路提取是车辆导航、数字线划图制作、空间地理数据库更新等应用的关键问题，也是学术界研究的热点前沿。遥感影像上地物的阴影和遮挡、相似线状地物的干扰以及光照不均等现象，导致道路提取出现断裂、错分误分、中心线提取精度低等问题，是制约道路鲁棒提取的主要因素。鉴于此，本项目首先针对道路断裂问题，设计聚合局部上下文和纹理信息的生成式对抗网络方法，从数据源上修复中断道路；在道路连续的基础上，构建区分道路与相似线状地物的损失函数，增大两者的差异，克服道路与相似地物易混淆的问题；最后提出道路中心线与道路分割多任务学习的方法，摆脱中心线级联预测对分割结果的依赖，提高道路分割与道路中心线提取的精度。本项目力图实现复杂背景条件下高分辨率影像中道路的鲁棒提取，研究成果有助于提高道路提取的自动化程度，具有重要的研究和经济价值。

从高分辨率遥感影像中提取道路是研究的热点与难点问题，随着计算机和人工智能技术的迅速发展，卷积神经网络在图像识别、目标检测等领域崭露头角，本课题开展基于卷积神经网络的高分辨率遥感影像道路提取。针对受阴影、遮挡及周围地物影响提取道路断裂问题，构建多尺度金字塔空洞卷积模块，增加感受野的同时强化道路区域的上下文关系，考虑空间语义信息和连通性信息；为改善道路提取精度，减小相似地物的影响，考虑道路本身的结构特性，设计带状注意力卷积，提升模型特征提取能力，突出道路的细节信息，提高道路提取精度；研究道路分割和中心线协同提取模型，同步开展道路分割和中心线提取，构建双任务并行网络模型，两个任务之间进行信息交流和特征传递，优化道路分割和中心线提取结果。本项目的研究有助于节省人力物力，为地理信息道路数据的更新提供支持。