基于改进U-Net模型的高分辨率遥感影像中城市建筑物的提取  

Extraction of urban buildings from high-resolution remote sensing images based on improved U-Net model

作　　者：秦梦宇 刘勇[1] 张寅丹 张洋[2] 侯建西 QIN Meng-yu;LIU Yong;ZHANG Yin-dan;ZHANG Yang;HOU Jian-xi(College of Earth and Environmental Sciences,Lanzhou University,Lanzhou 730000,China;Supercomputing Center,Lanzhou University,Lanzhou 730000,China;Hebei Changfeng Information Technology Co.,Ltd.,Shijiazhuang 050000,China)

机构地区：[1]兰州大学资源环境学院,兰州730000 [2]兰州大学超算中心,兰州730000 [3]河北长风信息技术有限公司,石家庄050000

出　　处：《兰州大学学报（自然科学版）》2022年第2期254-261,269,共9页Journal of Lanzhou University(Natural Sciences)

基　　金：国家自然科学基金项目(41271360)。

摘　　要：针对高分辨率遥感影像中城市建筑物周围环境复杂多样,易被阴影遮挡,难以精细化提取的问题,提出一种改进的U-Net网络用于图像中的城市建筑物提取.该网络在标准U-Net网络的编码器末端嵌入双重注意力模块,可以通过捕获全局建筑物信息和长通道建筑物信息,实现建筑物特征的增强.在交叉熵损失函数的基础上加入Lovász损失函数,构成的复合损失函数增强了对建筑物提取结果的约束能力,进一步提高了模型的鲁棒性.将该模型在美国马萨诸塞州数据集上进行验证,提取建筑物的F1-score为87.83%.结果表明,本方法对高分辨率遥感影像中周围环境复杂多样、被阴影遮挡的城市建筑物具有较强的提取能力.Urban buildings in high resolution remote sensing images usually have the characteristics of intricate surroundings and shadows,which make fine-scale image information extraction a problematic task.An improved U-Net network was proposed for extraction of urban buildings from high-resolution remote sensing images.The network embedded a dual attention module at the end of the encoder of the standard U-Net network.This module could capture global and long-channel building information to enhance building features.The Lovász loss function was added on the basis of the cross-entropy loss function,and the composite loss function formed enhanced the constraint ability of the building extraction results,and further improved the robustness of the model.Our results showed the F1-score was 87.83% verified by the public Massachusetts building dataset.The findings suggested that the improved U-Net could be a robust solution to extracting high-resolution urban buildings.

关 键 词：高分辨率遥感影像 城市建筑物提取 U-Net 双重注意力模块 复合损失函数 深度学习 

分 类 号：P237[天文地球—摄影测量与遥感]