Abstract

Identifying the spatial configurations of buildings and grouping them reasonably is an important task in cartography. This study developed a grouping approach using graph deep learning by integrating multiple cognitive features and manual cartographic experiences. Taking building center points as nodes, adjacent buildings were organized as a graph in which cognitive variables including size, orientation, and shape were defined for each node. Then, a learning model combining the graph convolution and neural network was designed to analyse the adjacent buildings modelled by the graph. The center points of groups were used as labels to train the positions of graph nodes and finally, a k-means algorithm was employed to obtain the grouping results based on the predicted node positions. Experiments confirmed that our approach can extract the inherent features describing the grouping relationship between buildings and performed better than two existing approaches referring to the ARI index (from 0.647 to 0.749).

摘要

识别建筑物的空间配置并将其合理分组是制图学中的重要任务。这项研究通过整合多个认知特征和手动制图经验，使用图深度学习开发了一种分组方法。以建筑物中心点为节点，将相邻建筑物组织为图形，其中为每个节点定义了包括大小，方向和形状在内的认知变量。然后，设计了一种将图卷积和神经网络相结合的学习模型，以分析通过图建模的相邻建筑物。使用组的中心点作为标签来训练图节点的位置，最后，使用k-means算法基于预测的节点位置获得分组结果。