图注意力神经网络支持下的建筑物形状识别  

Building shape recognition supported by graph attention networks

作　　者：张自强 刘涛 ZHANG Ziqiang;LIU Tao(Xi'an Railway Vocational&Technical Instute,Xi’an 710600 China;Faculty of Geomatics,Lanzhou Jiaotong University,Lanzhou 730070,China;Gansu Provincial Engineering Laboratory for National Geographic State Monitoring,Lanzhou 730070,China;National-Local Joint Engineering Research Center of Technologie and Applications for National Geographic State Monitoring,Lanzhou 730070,China)

机构地区：[1]西安铁路职业技术学院,西安710600 [2]兰州交通大学测绘与地理信息学院,兰州730070 [3]地理国情监测技术应用国家地方联合工程研究中心,兰州730070 [4]甘肃省地理国情监测工程实验室,兰州730070

出　　处：《测绘科学》2024年第9期125-133,共9页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(42261076,42061060);西安铁路职业技术学院2024年度立项课题项目(XTZY24J24)。

摘　　要：针对传统建筑物形状识别方法无法顾及空间认知、不适用于复杂建筑物形状的问题,引入图注意力神经网络识别建筑物形状,采用开源的建筑物形状数据集,使用坐标法、序列法、图方法和特征法等四种方法编码建筑物矢量数据,并对比不同深度学习模型分类建筑物形状的效果。实验表明,在编码方式上,坐标法和序列法精度较差,且分类效果与深度学习模型有一定关系;图方法能很好地分类建筑物形状,特征法能取得比图方法更好的分类效果。在模型上,卷积神经网络和循环神经网络模型分类效果较为均衡,对建筑物特征依赖性不强,而图卷积神经网络模型和图注意力神经网络模型只有和建筑物特征相结合才能取得较好结果,注意力机制的使用能够提高形状分类精度。In order to solve the problem that traditional building shape recognition method can not take space cognition into account and is not suitable for complex building shape,this paper introduces graph attention network to recognize building shape.The open source building shape data set is used to encode the building vector data using four methods:coordinate method,sequence method,graph method and feature method,the effects of different deep learning models on building shape classification were compared.The results show that coordinate method and sequence method yield lower accuracy,and their classification performance correlates with the chosen deep learning model.Graph method demonstrates superior capability in classifying building shapes,while feature method achieves better classification results than graph method.Convolutional neural network(CNN)and recurrent neural network(RNN)exhibit relatively balanced classification performance and weak dependency on building features.Graph convolutional neural network(GCN)models and graph attention neural network(GAT)model require integration with building features to achieve satisfactory results,with the utilization of attention mechanisms enhancing shape classification accuracy.

关 键 词：建筑物 形状分析 深度学习 注意力机制 图注意力神经网络 

分 类 号：P237[天文地球—摄影测量与遥感]