Generative Adversarial Networks to Generalise Urban Areas in Topographic Maps

Abstract

Abstract. This article presents how a generative adversarial network (GAN) can be employed to produce a generalised map that combines several cartographic themes in the dense context of urban areas. We use as input detailed buildings, roads, and rivers from topographic datasets produced by the French national mapping agency (IGN), and we expect as output of the GAN a legible map of these elements at a target scale of 1:50,000. This level of detail requires to reduce the amount of information while preserving patterns; covering dense inner cities block by a unique polygon is also necessary because these blocks cannot be represented with enlarged individual buildings. The target map has a style similar to the topographic map produced by IGN. This experiment succeeded in producing image tiles that look like legible maps. It also highlights the impact of data and representation choices on the quality of predicted images, and the challenge of learning geographic relationships.

Representing Vector Geographic Information As a Tensor for Deep Learning Based Map Generalisation

“…To evaluate the success of learning, without real formalisation of the evaluation process for maps generated using deep learning (Courtial et al, 2020b), we decided to only evaluate visually the results for each task. Our visual evaluation is guided by the following constraints for a generalised topographic map (Courtial et al, 2021b) This map generation task involves a GAN (Isola et al, 2018) that aims at the generation of a new image that looks like the target images from our training set. This architecture combines a generator that creates the image and a discriminator that evaluates if the generated image looks like the target domain.…”

“…Some other experiments have explored the generation of other kinds of maps or from stylized geographic information: e.g. Google map from OSM stylized data or artistic map from GoogleMap (Kang et al, 2019), topographic map from detailed national map agency data (Courtial et al, 2021b). These examples involve more complex cartographic processing, especially map generalisation, and are therefore more sensitive to the representation issues highlighted in the introduction of this paper.…”

“…Especially when the target map includes some abstraction, or generalisation, the notions of Isola et al, 2018)(2. Courtial et al, 2021b)(3. Kang et al, 2019) importance, patterns and spatial relations are essential, but are badly accessible in the inputs from the literature.…”

Integrating domain knowledge and graph convolutional neural networks to support river network selection

“…The strategy of converting vector data to raster data and using generative deep learning models to learn map generalization operators has become popular. For instance, building generalization was implemented using a deep convolutional neural network (Feng et al, 2019), while polyline generalization (Du et al, 2022), such as mountain road generalization (Courtial et al, 2022) and topographic map generalization in urban areas (Courtial et al, 2021), was performed using a generative adversarial network. Nevertheless, methods such as end-to-end multiple autoencoders for polyline generalization (Yu & Chen, 2022) or graph convolutional networks for road network selection (Zheng et al, 2021) have certain limitations.…”

“…Supported by remote-sensing big data, deep learning has solved many geospatial problems, including ground object interpretation (Zhang et al, 2019), change detection (Li et al, 2021), and terrain detection (Du et al, 2019). Inspired by such research and image-based deep learning applications like generative adversarial networks, map generalization operators, such as mountain road generalization (Courtial et al, 2022), building generalization (Courtial et al, 2021), and polyline simplification (Du et al, 2022) have been implemented using raster-based deep learning models.…”