结合对象单元和Transformer网络的城市功能区分类  

Object units and Transformer networks combined with urban functional zone classification method

作　　者：鲁伟鹏 贺清康 李佳铃 李诗逸 陶超[1] LU Weipeng;HE Qingkang;LI Jialing;LI Shiyi;TAO Chao(School of Geosciences and Info-Physics,Central South University,Changsha 410012,China;Department of Land Surveying and Geo-lnformatics,Hong Kong Polytechnic University,Hong Kong 999077,China)

机构地区：[1]中南大学地球科学与信息物理学院,长沙410012 [2]香港理工大学土地测量及地理资讯学系,中国香港999077

出　　处：《遥感学报》2024年第8期1927-1939,共13页NATIONAL REMOTE SENSING BULLETIN

基　　金：湖南省杰出青年基金(编号:2022JJ10072);湘江实验室开放基金一般项目(编码:22XJ03007);国家自然科学基金(编号:42171376,41771458);湖南省自然科学基金(编号:2021JJ30815);中南大学高性能计算平台。

摘　　要：准确识别各类城市功能区并全面掌握其分布情况,对合理规划和科学管理城市至关重要。针对该问题,本文提出一种结合对象单元和Transformer网络的城市功能区分类方法。该方法首先以多尺度分割所获得的过分割对象作为最小分析单元,以避免出现同一分析单元包含多种城市功能区的情况。在此基础上,针对现有方法着重于对分析单元内部特征提取而忽略了分析单元之间的空间关系问题,提出利用Transformer框架和对象地理属性作为位置编码对不同分析单元之间的空间关系进行建模,从而实现兼顾分析单元内部特征和不同分析单元之间空间关系的城市功能区分类。结果表明,使用过分割对象作为最小分析单元能够更加准确地获取城市功能区地边界,从而避免基于规则格网单元所导致的锯齿状边缘及基于路网单元所导致地无法区分路网内不同功能区的问题;与仅考虑分析单元内部特征的传统方法相比,通过对不同分析单元之间的分析单元进行建模可有效提升城市功能区分类精度。Urban Functional Zones(UFZs) refer to specific areas within a city that have distinct functionalities and land uses. These zones are designated based on their primary activities and the roles they play in the urban environment. Accurate extraction of UFZs and a comprehensive understanding of their spatial distribution play an important role in urban planning and management. Traditional Convolutional Neural Networks(CNNs) focus on local features through convolutions, but they often miss the broader spatial relationships.Vision Transformer(ViT), while advanced, still has limitations;its tokenization method and learnable position encoding do not effectively represent geographical entities and their spatial relationships, which is a crucial feature in geospatial analysis.This study proposes a UFZ classification method combining object units and ViT to address this issue. First, this method utilizes oversegmented objects generated from a multi-scale segmentation approach as analysis units to avoid the presence of multiple kinds of UFZs within a single object. Over-segmentation helps in creating smaller, more homogeneous units, thereby increasing the precision of the classification process. Then, considering that current methods often focus on the inherent analysis of objects while ignoring their spatial relationships, ViT is employed for spatial relationship modeling between objects, with the geographic attributes of objects serving as position embeddings. In this way, both the inherent features of a single analysis unit and the inter-spatial features among objects are considered for UFZ classification. Position embeddings using geographic coordinates allow the model to understand spatial proximity and relationships, which are crucial for accurate classification. We chose Beijing as the study area and downloaded imagery of the area within the Sixth Ring Road from Bing Maps. We also collected labels from OpenStreetMap and reclassified them into 10 typical urban functional zones according to the “Code for classificati

关 键 词：城市功能区 遥感 深度学习 空间关系建模 Transformer网络 

分 类 号：P231[天文地球—摄影测量与遥感] P2[天文地球—测绘科学与技术]