高分辨率遥感影像房屋特征迁移学习与变化检测研究

Building change detection is an important technique for monitoring the development in urban areas. However, it is facing problems like the multitemporal building "projection difference" caused by different imaging angles, the low separability between buildings and other objects because of the regional variability of objects’ distribution, and the increase the errors arouse from repeating building extractions in the multitemporal images. In this condition, a model of building change detection in high (spatial) resolution remote sensing images is proposed in this project. After building extraction in one tempore, the building neighboring feature, which is not affected from the imaging angle, is extracted and transfer learnt to obtain the unchanged buildings in the multitemporal images. Consequently, the side-looking feature of unchanged buildings, unaffected from the regional objects’ distribution, is extracted and transfer learnt in different regions to generate the increased buildings. By extraction of building neighboring and side-looking features, the proposed model could reveal the building imaging rules in the multitemporal images or different image regions, and avoid the effect from building “projection difference” and regional objects’ distribution. Meanwhile, by respectively multitemporal and multiregional transfer learning of building neighboring and side-looking feature, this model could avoid the repeating building extraction, and establish the relationship among the building neighboring feature, side-looking feature and change building areas, which has both promotable scientific significance and practical value.

房屋变化检测是监测城市化进程的重要方法之一。房屋变化检测面临的主要问题有：多时相影像成像角度差异导致的房屋“投影差”，区域性地物分布差异造成的房屋与其他地物可分性较低，以及各时相房屋重复提取造成的误差累积等。针对以上问题，本课题拟提出一种高(空间)分辨率遥感影像房屋变化检测模型，首先提取某时相影像中的房屋区域及其不受成像角度影响的邻域特征，对该特征进行多时相迁移学习，获取非新增房屋区域；然后提取非新增房屋不受区域性地物分布影响的侧视特征，并对其进行多区域迁移学习，获取新增房屋区域。该模型通过房屋邻域与侧视特征的提取，揭示了房屋在多时相影像及同一影像不同区域中的成像规律，避免了房屋“投影差”与区域性地物分布对房屋变化检测的影响。同时，利用上述房屋特征的多时相与多区域迁移学习，可避免房屋重复提取，建立多时相房屋邻域特征、多区域房屋侧视特征与变化房屋区域的内在联系，具有较强的科学意义与应用价值。

房屋变化检测是监测城市化进程的重要方法之一。本课题针对房屋变化检测面临的主要问题，分别定义了房屋空间统计与邻域特征，实现了房屋空间特征在多时相影像之间的迁移学习，该过程可用于同源或多源高分辨率遥感影像变化检测；通过双层卷积神经网络（two-stage convotional neural networks, two-stage CNNs）的构造，实现了样本房屋整体（屋顶及侧墙）特征学习及其在非样本区域的提取，即同一影像不同区域的特征迁移学习，获得了高分辨率遥感影像房屋提取结果。该课题结果对高分辨率影像信息提取、同源/多源影像变化检测以及地理国情监测等具有重要的理论意义和实用价值。