基于U-Net的高分辨率遥感影像建筑物自动解译

Automatic Interpretation of High Resolution Remote Sensing Images of Building Based on U-Net

建筑物基础信息在智慧城市建设、地理国情监测等领域有着重要作用。针对传统方法提取高分辨率卫星影像精度低的问题,提出一种基于U-Net的高分辨率遥感影像建筑物自动解译方法。首先,通过ArcGIS制作遥感图像建筑物数据集;其次,针对建筑物数据样本不足的问题,采用数据增强的方式扩充数据;然后,采用迁移学习的方法在开源ImageNet数据集预训练U-Net模型;最后,通过预训练U-Net模型训练与预测研究区域数据。实验结果表明,该方法能够准确地从影像中识别出建筑物,整体的识别精度达到98%,能够自动的解译出遥感影像建筑物轮廓信息,可为建筑物的提取提供一定的参考价值。

The basic information of buildings has an important significance in the fields of smart city construction and geographic condition monitoring.Aiming at the problem of low accuracy of the traditional method to extract high-resolution satellite images,an automatic interpretation method based on U-Net for high-resolution remote sensing images of buildings is proposed.Firstly,the remote sensing image building dataset is produced by ArcGIS;secondly,the data is expanded by data augmentation for the problem of insufficient building data samples;then,the migration learning method is used to pre-train the U-Net model in the open source ImageNet dataset;and finally,the pre-trained U-Net model is used to train and predict the study area data.The experimental results show that,The method is accurate enough to identify buildings from images,and the overall recognition accuracy reaches 98%.It can automatically decode the building outline information of remote sensing images,which can provide certain reference value for building proposals.