一种高分辨率SAR图像变化检测数据集制作方法  

A High-Resolution SAR Images Change Detection Dataset Production Method

作　　者：李泽宇 赵江洪 

机构地区：[1]北京建筑大学测绘与城市空间信息学院,北京 [2]建筑遗产精细重构与健康监测北京市重点实验室,北京 [3]自然资源部城市空间信息重点实验室,北京 [4]地理信息工程国家重点实验室,陕西 西安 [5]武汉大学测绘遥感信息工程国家重点实验室,湖北 武汉

出　　处：《测绘科学技术》2023年第2期115-121,共7页Geomatics Science and Technology

摘　　要：针对可用于变化检测模型训练的高分辨率SAR图像数据不足、SAR图像处理与样本标注复杂的问题,本文开展SAR图像变化检测数据集制作方法研究,简化SAR图像处理与样本标注流程,对多幅高分三号SAR影像依次进行预处理、精配准、样本标注与数据增强等操作,得到一套高分辨率SAR图像变化检测数据集,为深度学习变化检测网络的训练与测试提供数据基础。通过多种网络模型训练的实验测试与结果分析,验证了该数据集能够有效支持多场景变化检测任务,对于不同变化检测模型训练精度与效率的提升具有较好的参考价值。Aiming at the problems of insufficient high-resolution SAR image data that can be used for change detection model training and the complexity of SAR image processing and sample labeling, this pa-per carried out research on the production method of SAR image change detection data set to sim-plify the SAR image processing and sample labeling process. A set of high-resolution SAR image change detection data set is obtained by preprocessing, fine registration, sample labeling and data enhancement of several high-resolution third-level SAR images successively, which provides data basis for the training and testing of deep learning change detection network. Through the experi-mental test and analysis of STANet, DSAMNet and SNUNet network model training, the validity of the SAR image change detection data set obtained by the data set making method proposed in this paper is verified in the multi-scene change detection task.

关 键 词：变化检测 SAR图像处理 模型训练 数据集 SAR影像 深度学习 数据增强 高分三号 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]