基于改进UNet网络的遥感影像建筑物提取  

Building Extraction from Remote Sensing Image Based on Improved UNet

作　　者：吴锋振 杨德宏[1] 李俊[1] 何万才 邓云龙 WU Fengzhen;YANG Dehong;LI Jun;HE Wancai;DENG Yunlong(Faculty of Land Resource Engineering,Kunming University of Science and Technology,Kunming 650093,China;Key Laboratory of Geospatial Information Integration and Innovation for Intelligent Mines,Kunming 650093,China;Science and Technology Innovation Team for Integration and Application of Natural Resources Spatial Information in Yunnan Universities,Kunming 650211,China)

机构地区：[1]昆明理工大学国土资源工程学院,云南昆明650093 [2]智慧矿山地理空间信息集成创新重点实验室,云南昆明650093 [3]云南省高校自然资源空间信息集成与应用科技创新团队,云南昆明650211

出　　处：《城市勘测》2023年第5期99-105,共7页Urban Geotechnical Investigation & Surveying

摘　　要：针对卷积神经网络在高分辨率遥感影像建筑物提取中存在边缘模糊和建筑物空洞等问题,提出一种结合似空间注意力模块UNet网络用于建筑物提取。为验证网络的有效性和适用性,分别在WHU和AIRS数据集上进行实验,并将其与FCN-8S、SegNet和UNet进行对比。实验结果表明,在WHU建筑物数据集上,精确率达94.12%,交并比达91.74%,与其他三个网络相比,各项评价指标有一定程度提升,在模型参数量和每轮次运行时间增加的可接受范围内,表现出良好分割性能,验证提出方法的有效性;在AIRS数据集上,交并比、精确率分别为90.40%、90.22%,除F1分数比FCN-8S略低外,其他指标也存在优势,验证方法的适用性。Aiming at the problems of blurred edges and holes in buildings in high-resolution remote sensing image building extraction by convolutional neural networks,this paper proposes a UNet network combined with a spatial-like attention module.To verify the effectiveness and applicability of our network,experiments are conducted on the WHU and AIRS dataset,respectively,and compared with typical building extraction networks FCN-8S,SegNet and UNet.The experimental results show that on the WHU building dataset,precision reaches 94.12%,and IoU reaches 91.74%.Compared with the other three networks,the evaluation indicators have been improved to a certain extent.Within the acceptable range of the increase of the number of parameters and the running time,the method in this paper shows good segmentation performance,which verifies the effectiveness of the method in this paper;on the AIRS dataset,IoU and precision are 90.40% and 90.22%,respectively.In addition to the slightly lower FCN-8S,other indicators also have advantages,which verifies the applicability of the method in this paper.

关 键 词：卷积神经网络 建筑物提取 UNet 似注意力模块 特征校正 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]