基于多级特征融合网络的武汉城市局部气候区分类  

Local Climate Zones Classification in Wuhan urban area based on multi-feature fusion network

作　　者：潘岩松 李英冰[1] 伍智超 李美权 刘波 PAN Yansong;LI Yingbing;WU Zhichao;LI Meiquan;LIU Bo(School of Geodesy and Geomatics,Wuhan University,Wuhan 430079,China)

机构地区：[1]武汉大学测绘学院,武汉430079

出　　处：《测绘工程》2025年第3期1-7,共7页Engineering of Surveying and Mapping

基　　金：国家重点研发计划“国家公共安全应急平台”项目(2018YFC0807000)。

摘　　要：人类活动引起大量土地覆盖变化,加剧了城市内部的环境变化。这种变化不仅表现在自然地表和建筑物之间的转移,也表现在不同建筑类型的转变中,因此需要更精细的方法来研究城市内部的空间结构。针对一般土地分类方案忽略城市结构的复杂性和空间异质性,难以详细描述城市内部的形态的问题,文中基于局部气候区(LCZ)方案提出一种用残差连接改进的多级特征融合网络(Sen2LCZ-MF),在专门为LCZ分类任务设计的So2Sat LCZ42数据集上进行训练,取得了66.6%的分类精度。针对长尾数据集样本数量不平衡的问题,引入类平衡损失作为训练的损失函数,使分类精度提升了2.5%。应用训练好的模型绘制了武汉城区的LCZ地图,结果显示武汉城区不透水面比例达到31.5%,17种LCZ类型中占比最高的是低矮植物。通过混淆矩阵和真实地面图像的对比发现,自然地表和部分建筑类型的分类准确率在75%以上,较为准确地识别了主要的LCZ类型。Human activities have induced significant land cover changes,exacerbating the environmental changes within cities.Such changes are not only manifested in the transfer between natural land surface and buildings,but also in the transformation of different building types,thus requiring a finer-grained approach to study the spatial structure within cities.To address the problem that general land classification schemes ignore the complexity and spatial heterogeneity of urban structures,which makes it difficult to describe the morphology of the urban interior in detail,this paper proposes an multi-feature fusion network(Sen2LCZ-MF)improved with residual block based on the Local Climate Zones(LCZ)scheme,which was trained on the So2Sat LCZ42 dataset designed specifically for the LCZ classification task and achieved a overall accuracy of 66.6%.For the problem of imbalance in the number of samples in long-tailed dataset,the class balanced loss is introduced as the loss function during training,which improves the classification accuracy by 2.4%.The trained model was applied to draw the LCZ map of Wuhan urban area,and the results showed that the proportion of impervious surface in Wuhan urban area reached 31.5%,and the highest proportion of the 17 LCZs was low plants.The comparison between the confusion matrix and the real ground image reveals that the classification accuracy of the natural surface and some building types is above 75%,which accurately identify the main LCZ types.

关 键 词：局部气候区 土地覆被分类 深度学习 特征融合网络 类平衡损失 

分 类 号：P237[天文地球—摄影测量与遥感]