基于改进U-Net模型的复杂建筑物轮廓自动提取研究  

Research on Automatic Extraction of Complex Building Contours Based on Improved U-NetModel

作　　者：王瑞亭 喜文飞 白世晗 钱堂慧 丁子天 郭峻杞 WANG Ruiting;XI Wenfei;BAI Shihan;QIAN Tanghui;DING Zitian;GUO Junqi(Faculty of Geography Yunnan Normal University,Kunming 650500,China;Yunnan Survey and Design Institute of Water Conservancy and Hydropower,Kunming 650500,China)

机构地区：[1]云南师范大学地理学部,云南昆明650500 [2]云南省水利水电勘测设计研究院,云南昆明650500

出　　处：《城市勘测》2025年第1期8-13,共6页Urban Geotechnical Investigation & Surveying

基　　金：云南省重大科技专项(202202AD080010);滇中引水工程(滇中高原区)高分综合应用示范(89-Y50-G31-9001-22/23-05)。

摘　　要：复杂建筑物轮廓信息对城市变化监测、城市规划及三维建模有着重要意义。针对现有深度学习模型在复杂建筑物轮廓提取中存在的边缘信息提取不完整的问题,本文通过改进U-Net网络,提出一种基于注意力机制的ATT-Unet模型。该模型通过注意力机制动态调整特征权重,细化边缘特征,增强模型特征提取能力。实验结果表明:引入注意力机制提升了模型对复杂建筑物边缘的识别精度,相较于传统U-Net模型,ATT-Unet精确率提高15.2%,召回率提高20.1%,F1分数提高17.7%,交并比提高26.6%。该模型能够准确提取复杂建筑物的完整轮廓,可以用于监测城市建设动态变化,辅助制定科学的城市规划方案。Complex building information plays a vital role in urban change monitoring,urban planning,and 3D modeling.To address the issue of incomplete edge extraction in existing deep learning models for complex building extraction.This study introduces an improved U-Net model,ATT-Unet,which introduces an attention mechanism.The model is applied to extract complex buildings from high-resolution remote sensing images.Results show that the introduction of the attention mechanism significantly enhances the model's ability to recognize building edges.Compared to the traditional U-Net model,ATT-Unet achieves a 15.2%improvement in precision,a 20.1%increase in recall,a 17.7%boost in F1 score,and a 26.6%improvement in Intersection over Union(IoU).This model accurately extracts the complete contours of complex buildings and can be utilized to monitor urban development dynamics,assist in formulating scientific urban planning strategies,and optimize urban land resource allocation.Additionally,the high-precision building extraction results can serve as foundational data for 3D urban modeling,further supporting the refinement and intelligence of urban management.It provides essential insights for enhancing comprehensive urban governance.

关 键 词：U-Net模型 注意力机制 建筑物提取 深度学习 高分辨率遥感影像 

分 类 号：TP751.2[自动化与计算机技术—检测技术与自动化装置] TP181[自动化与计算机技术—控制科学与工程]