深度迁移学习支持下的高分辨率遥感影像变化检测研究

The dynamic analysis of earth surface changes is crucial to understanding the relationship between human development and natural phenomena. At present, remote sensing technology can achieve continuous observation with high spatial resolution, wide range and high frequency, which provides a solid data base for detecting, identifying and analyzing changes. In fact, supervised change detection is widely used as it can show the shift of change types in detail, which often relies on adequate and reliable change samples. However, the cost of selecting a large number of reliable samples is usually high, which seriously restricts the accuracy and the level of automation of change detection. Therefore, this project aims to improve the accuracy and intelligence level of high-resolution remote sensing change detection by introducing the theories and methods of transfer learning and deep learning. It concentrates on three aspects: First of all, the project will develop a deep transfer learning network which can effectively solve the problem of scene transfer and automated sample selection. Secondly, the determination mechanism of reliable samples is then proposed to improve the reliability of the change detection samples by the similarity measure of the features of the multi-phase high resolution imagery. Finally, a multi-task framework using multi source knowledge fusion is established, which can mine the related deep features of multi temporal remote sensing images and improving the accuracy of change detection. The research results will provide theoretical and technical support for the applications of ecosystem monitoring, urban planning, disaster assessment and many other aspects.

地表变化的动态检测分析对于理解人类发展与自然现象之间的关系至关重要。目前，遥感技术可以实现高空间分辨率、大范围、高频次的持续观测，为探测、识别和分析变化提供了坚实的数据基础。方法上，监督变化检测可以详尽的展示变化类型的转移并因此得到了广泛应用，但往往依赖于充足且可靠的变化样本。然而，大量可靠样本的获取成本通常较高，严重制约了变化检测的精度及自动化水平。因此，本项目拟借助迁移学习对知识迁移的能力及深度学习高层次特征的自学习能力，提升高分辨率遥感变化检测的精度与智能化水平。研究内容包括：构建有标注历史场景向无标注变化场景的迁移模型，有效解决变化样本少的问题；通过对多时相高分辨率像对特征的相似测度，进行变化样本筛选以提高可靠性；建立多源知识联合的多任务变化检测框架，挖掘多时相遥感影像的相关特征，提升变化检测的精度。研究成果将为生态系统监测、城市规划、灾害评估等诸多方面的应用中提供理论和技术支撑。

本项目针对多时相高分辨率遥感影像变化检测方法的瓶颈，提出了深度网络层次化特征学习、知识迁移以及多任务深度融合等相关模型，来提升高分辨率遥感变化检测的精度与智能化水平。该解决方案可为城市规划、生态系统监测、灾害评估等诸多方面的应用提供支撑。主要成果如下：（1）针对传统遥感影像分类算法特征学习能力不足的问题，发展了三种用于典型地物识别（建筑物、树木、土地覆盖）的层次化卷积特征提取与分类模型，实现了面向多时相影像的高精度分类；（2）针对变化样本缺失的问题，利用目标域样本迁移与可靠性测度方法，构建了多时相遥感影像变化检测样本集；（3）提出了一种多源知识联合的多任务高分辨率遥感影像变化检测框架，有效提升了变化检测的精度及其自动化水平。本项目基本按原计划执行，针对不同领域，包括建筑物、植被等的变化实例进行了应用研究。围绕本项目的研究内容发表了系列原创性的论文，在International Journal of Applied Earth Observation and Geoinformation等期刊上发表论文8篇；申请国家发明专利2项；项目成果还获得了省部级科技进步奖2项；项目执行期，5人次参加了国际交流；本项目协助培养了博士4名，硕士1名。总体上来讲，本项目的研究成果已基本达到预期目标，但也有部分内容需要继续改善：项目研究中建立的样本集范围仅限于亚热带城市，全球更大范围内不同城市的样本需进一步囊括进来。随着海量高分辨率影像的不断积累，利用该方法可以实现更多领域、更密集时间尺度下内的变化检测，为其他相关研究提供相关的技术和数据支撑。