Integrating Artificial Intelligence in Cartography: Using Deep Learning for Map Style Transfer and Map Generalization

Abstract

Maps have always been considered as a combination of science and art. Following a set of stylistic design criteria that integrates human creativity, perception, and experience, cartographers are able to produce unique map aesthetics that deliver geospatial information. Recently, the advancement of the Artificial Intelligence (AI) technologies makes the line between human and machine increasingly blurred. Machines are able to “see”, “listen” to the world, understand human feelings, and even produce “true” artworks such as visual arts and stylistic design, which bring new opportunities for cartography. In this thesis, I propose a systematic framework that integrates AI in cartography and illustrate two specific cartographic topics: map style transfer and map generalization. I first illustrate the workflow for producing large-scaled tiled maps from GIS vector data with open source software. Then, by training convolutional neural networks such as generative adversarial network (GAN) models and deep neural network (e.g., U-Net), the cartographic knowledge, namely stylistic elements and generalization rules can be learned from existing maps and transferred to target maps across multiple map scales. The architecture and design of deep learning approaches, and how and what cartographic knowledge is encoded into the model, are illustrated in detail. Additionally, I used two approaches to evaluate the experiment results and judge if deep learning approaches perform well, namely machine-based metric and human-centered evaluation to assess the outputs comprehensively. These two approaches can measure the performances of machine learning-based methods from different aspects. Though many challenges remain requiring future research, the thesis show great potential of using deep learning methods for solving cartographic tasks. Integrating AI in artistic part of maps may even provide a potential new paradigm for the next decades of cartography which worth exploring.