Learning from vector data: enhancing vector-based shape encoding and shape classification for map generalization purposes

Abstract

Map generalization is a complex task that requires a high level of spatial cognition, and deep learning techniques have shown in numerous research fields that they could match or even outplay human cognition when knowledge is implicitly in the data. First experiments that apply deep learning techniques to map generalization tasks thereby adapt models from image processing, creating input data by rasterizing spatial vector data. Because image-based learning has major shortcomings for map generalization, this article investigates possibilities to learn directly from vector data, utilizing vector-based encoding schemes. First, we enhance preprocessing methods to match essential properties of deep learning models – namely regularity and feature description – and evaluate the performance of Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Graph Convolutional Neural Networks (GCNN) in combination with a feature-based encoding scheme. The results show that feature descriptors improve the accuracy of all three neural networks, and that the overall performances of the models are quite similar for both polygon and polyline shape classification tasks. In a second step, we implement an exemplary building generalization workflow based on shape classification and template matching, and discuss the generalization results based on a case study.