Among cartographic generalisation problems, the generalisation of sinuous bends in mountain roads has always been a popular one due to its difficulty. Recent research showed the potential of deep learning techniques to overcome some remaining research problems regarding the automation of cartographic generalisation. This paper explores this potential on the popular mountain road generalisation problem, which requires smoothing the road, enlarging the bend summits, and schematising the bend series by removing some of the bends. We modelled the mountain road generalisation as a deep learning problem by generating an image from input vector road data, and tried to generate it as an output of the model a new image of the generalised roads. Similarly to previous studies on building generalisation, we used a U-Net architecture to generate the generalised image from the ungeneralised image. The deep learning model was trained and evaluated on a dataset composed of roads in the Alps extracted from IGN (the French national mapping agency) maps at 1:250,000 (output) and 1:25,000 (input) scale. The results are encouraging as the output image looks like a generalised version of the roads and the accuracy of pixel segmentation is around 65%. The model learns how to smooth the output roads, and that it needs to displace and enlarge symbols but does not always correctly achieve these operations. This article shows the ability of deep learning to understand and manage the geographic information for generalisation, but also highlights challenges to come.

The "Segmentation", i.e. the three dimensional point clouds partition in different morphological zones, is a necessary operation while approaching the reverse engineering cycle, because it helps the operator in generating the surface model. This operation is usually developed after the acquisition and the pre-processing phases, and it tries to define a boundary grid which the following surface fitting operation will employ for the surface model definition. Many approaches apply the segmentation methods far from the 3D scanner device. On the contrary, this research work proposes an iterative strategy which starts from a first raw point acquisition, then it partitions the object surface and identifies the boundary of those zones showing significant morphological features (shapechanges). As a consequence, they will be re-digitised with deeper scansions, in order to reach a more precise morphological information. This partitioning operation is driven by a morphology descriptor, the Guassian Curvature, giving and estimation of the local surface morphological complexity. Moreover the proposed algorithm employs the 3D scanner measuring uncertainty to define a "curvature variation threshold", in order to identify those zones showing significant morphological shape-changes.

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.

Map generalisation is a process that transforms geographic information for a cartographic at a specific scale. The goal is to produce legible and informative maps even at small scales from a detailed dataset. The potential of deep learning to help in this task is still unknown. This article examines the use case of mountain road generalisation, to explore the potential of a specific deep learning approach: generative adversarial networks (GAN). Our goal is to generate images that depict road maps generalised at the 1:250k scale, from images that depict road maps of the same area using un-generalised 1:25k data. This paper not only shows the potential of deep learning to generate generalised mountain roads, but also analyses how the process of deep learning generalisation works, compares supervised and unsupervised learning, and explores possible improvements. With this experiment we have exhibited an unsupervised model that is able to generate generalised maps evaluated as good as the reference and reviewed some possible improvements for deep learning-based generalisation, including training set management, and the definition of a new road connectivity loss. All our results are evaluated visually using a four questions process and validated by a user test conducted on 113 individuals.

Deep learning techniques have recently been experimented for map generalization. Although promising, these experiments raise new problems regarding the evaluation of the output images. Traditional map generalization evaluation cannot directly be applied to the results in a raster format. Additionally, the internal evaluation used by deep learning models is mostly based on the realism of images and the accuracy of pixels, and none of these criteria is sufficient to evaluate a generalization process. Finally, deep learning processes tend to hide the causal mechanisms and do not always guarantee a result that follows cartographic principles. In this article, we propose a method to adapt constraint-based evaluation to the images generated by deep learning models. We focus on the use case of mountain road generalization, and detail seven raster-based constraints, namely, clutter, coalescence reduction, smoothness, position preservation, road connectivity preservation, noise absence, and color realism constraints. These constraints can contribute to current studies on deep learning-based map generalization, as they can help guide the learning process, compare different models, validate these models, and identify remaining problems in the output images. They can also be used to assess the quality of training examples.