基于深度学习的高分辨率光学卫星遥感影像建筑物变化检测方法

A change detection method for high resolution optical satellite remote sensing image based on deep learning

随着遥感技术的发展与遥感影像更新频率的提升,自动化建筑物识别与变化检测,在城市规划与管理、城市数据更新等领域都起着极为重要的作用。本文针对建筑物变化特征复杂的特性,提出了一种基于深度学习的建筑物变化检测方法。首先,针对上海市卫星影像特征建立了建筑物变化样本数据库,并利用翻转、改变亮度等多种方式进行样本的扩充;然后,结合深度学习Resnet、Unet等网络结构进行算法的优化与改善,得到优化后的变化检测网络结构;接着,用样本库进行网络参数的训练;最终得到适合上海区域的变化检测网络模型。通过在上海市0.5m高分辨率卫星数据上进行多种算法的对比实验,表明本文方法能够有效进行城市建筑物变化检测,并能得到较优的效果。

With the development of remote sensing technology and the increase of satellite imagery update frequency,automatic building identification and change detection play an extremely important role in urban planning and management,urban data update and other fields.According to the complexity of building change,a change detection method based on deep learning is put for-ward.Firstly,the sample database is established using the satellite images of buildings in Shanghai,and the data sets are expanded by rotating images,adjusting brightness,etc.Then the detection algorithm and network are optimized and improved by using the architecture of neural network,such as Resnet and Unet.Finally,after training the network parameters with the sample data sets,the building change detection network model suitable for Shanghai is obtained.