Is deep learning the new agent for map generalization?

Abstract

The automation of map generalization has been keeping researchers in cartography busy for years. Particularly great progress was made in the late 90's with the use of the multi-agent paradigm. Although the current use of automatic processes in some national mapping agencies is a great achievement, there are still many unsolved issues and research seems to stagnate in the recent years. With the success of deep learning in many fields of science, including geographic information science, this paper poses the controversial question of the title: is deep learning the new agent, i.e. the technique that will make generalization research bridge the gap to fully automated generalization processes? The paper neither responds a clear yes nor a clear no but discusses what issues could be tackled with deep learning and what the promising perspectives. Some preliminary experiments with building generalization or data enrichments are presented to support the discussion.

Exploring the Potential of Deep Learning Segmentation for Mountain Roads Generalisation

“…Beyond more traditional machine learning techniques, the use of deep learning for map generalisation is recent, but the potential is real [8]. As map generalisation is a graphical problem, deep learning techniques should be able to learn the implicit graphic structures, such as key spatial relations, necessary for a good map generalisation.…”

“…There are other possible ways to model road generalisation as a deep learning problem, for instance by using graph convolutions [29]. But we think that CNN segmentation is adapted because all the information necessary to draw a generalised version of a road can be included in the image of the road [8]. The usual constraints defined for mountain road generalisation are the following [16]:…”

“…Machine learning was an early solution to overcome traditional methods and capture the knowledge of human cartographers on how to orchestrate geometrical transformations [4][5][6][7]. The main limitation of all these approaches seem to be the capacity of machine learning to mimic the complexity of human reasoning, particularly with implicit spatial relations (e.g., buildings aligned along a road); however, as in many other fields where machine learning is used, deep learning approaches might be able to overcome these problems [8]. The first attempts to generalise buildings with a U-Net (a convolutional neural network for image segmentation) [9] showed very promising results [10].…”

A Multi-scale Embossed Map Authoring Tool for Indoor Environments

“…Advances in machine learning techniques have prompted researchers to revisit the problem of cartographic generalization. Several neural-network models have been developed for the tasks of recognizing, grouping, and typifying buildings [20,7,24]. However, these models are only able to learn and predict a building's contour and the geographical distribution if groups of buildings, and may not generalize well for the representation of the layout of an indoor space.…”

OSMWatchman: Learning How to Detect Vandalized Contributions in OSM Using a Random Forest Classifier

“…One of the main challenges of supervised machine learning is domain adaptation, also known as transfer learning [10,11]. This issue refers to the ability of the learning algorithm to classify the data regardless, in our case, of the geographical region considered.…”