基于图卷积神经网络的建筑物合并方法

Building Aggregation Method Based on Graph Convolutional Neural Network

建筑物合并是地图综合中的复杂决策问题,基于规则的合并方法试图通过一系列认知参数和约束条件来解决“哪些合并”“如何合并”等问题,但是由于区域环境的复杂性、算子选择的多样性以及参量调整的多变性,基于规则的合并结果总是不尽人意。在人工智能技术支持下,将深度学习方法引入,则是从数据驱动视角通过样本训练以期突破规则方法的局限。本研究即是基于图卷积神经网络构建的一种建筑物分组及合并的方法。利用约束Delaunay三角网和最小生成树构建图结构,从图结构中提取建筑物节点特征和距离特征。通过图卷积层和点积运算得到的节点余弦相似度与距离特征拼接作为连接边特征,最后通过全连接层得到二分类结果。通过对比实验分析,本方法的ARI指数高于现有的方法 (从0.674到0.800)。

Building aggregation is a complex decision-making problem in map generalization,where rule-based aggregation methods attempt to address the questions of“which to aggregate”and“how to aggregate”through a set of cognitive parameters and constraints.However,due to the complexity of regional environments,the diversity of operator selections and the variability of parameter adjustments,the results of rule-based aggregation are often unsatisfactory.With the support of artificial intelligence technology,introducing deep learning methods offers a way to overcome the limitations of rule-based approaches from a data-driven perspective.We presented a method for building grouping and ag-gregation based on graph convolutional neural network.We utilized constrained Delaunay triangulation and the minimum spanning tree to con-struct graph structure,extracting building node features and distance features from the graph.Then,we took the node cosine similarity obtained through graph convolution layers and dot product operations,concatenated with distance features as edge characteristics.Finally,we achieved a binary classification result through a fully connected layer.Through comparative experimental analysis,the ARI index of this method is higher than that of existing methods(from 0.674 to 0.800).