<scp>PointConv</scp> ‐ <scp>AE</scp> : A Multi‐Stage Model for Building Simplification in Vector Maps

ABSTRACT The simplification of building footprints is essential for cartographic generalization, particularly in tasks that involve scale transformation from large to medium scales. However, traditional simplification algorithms struggle to achieve intelligent and efficient performance because of their dependence on manual parameter tuning and expert knowledge formalization. Moreover, existing graph convolutional deep learning methods for simplification suffer from diminished accuracy because of redundancy in input information. This study introduces a multi‐stage Point Convolutional Autoencoder (PointConv‐AE) model that first performs point removal classification and subsequently conducts point movement prediction to achieve building simplification. The PointConv‐AE model directly utilizes vector data as input, performing convolutions at the point level for relevant points without relying on graph structures or extensive feature extraction. The point AE module processes only those points classified with a movement label and predicts point movement on the basis of the preservation constraints inherent in building simplification. The model was evaluated on the Stuttgart building dataset, revealing that the proposed method outperforms other deep learning approaches in evaluation metrics. Furthermore, sensitivity analyses of the model's components were conducted to identify the optimal model structure.