A deep learning approach for polyline and building simplification based on graph autoencoder with flexible constraints

• Published in: Cartography and geographic information science Vol. 51; no. 1; pp. 79 - 96
• Main Authors: Yan, Xiongfeng, Yang, Min
• Format: Journal Article
• Language: English
• Published: Taylor & Francis 02.01.2024
• Subjects: building simplification; deep learning; graph autoencoder; map generalization; Polyline simplification
• ISSN: 1523-0406; 1545-0465
• DOI: 10.1080/15230406.2023.2218106

Polyline and building simplification remain challenging in cartography. Most proposed algorithms are geometric-based and rely on specific rules. In this study, we propose a deep learning approach to simplify polylines and buildings based on a graph autoencoder (GAE). The model receives the coordinates of line vertices as inputs and obtains a simplified representation by reconstructing the original inputs with fewer vertices through pooling, in which the graph convolution based on graph Fourier transform is used for the layer-by-layer feature computation. By adjusting the loss functions, constraints such as area and shape preservation and angle-characteristic enhancement are flexibly configured under a unified learning framework. Our results confirmed the applicability of the GAE approach to the multi-scale simplification of land-cover boundaries and contours by adjusting the number of output nodes. Compared with existing Douglas‒Peukcer, Fourier transform, and Delaunay triangulation approaches, the GAE approach was superior in achieving morphological abstraction while producing reasonably low position, area, and shape changes. Furthermore, we applied it to simplify buildings and demonstrated the potential for preserving the diversified characteristics of different types of lines.