DeepMapScaler: a workflow of deep neural networks for the generation of generalised maps
Résumé
The automation of map generalisation has been an important research subject for decades but is not fully solved yet. Deep learning techniques are designed for various image generation tasks, so one may think that it would be possible to apply these techniques to cartography and train a holistic model for end-to-end map generalisation. On the contrary, we assume that map generalisation is a task too complex to be learnt with a unique model. Thus, in this article, we propose to resort to past research on map generalisation and to separate map generalisation into simpler sub-tasks, each of which can be more easily resolved by a deep neural network. Our main contribution is a workflow of deep models, called DeepMapScaler, which achieves a step-by-step topographic map generalisation from detailed topographic data. First, we implement this workflow to generalise topographic maps containing roads, buildings, and rivers at a medium scale (1:50k) from a detailed dataset. The results of each step are quantitatively and visually evaluated. Then the generalised images are compared with the generalisation performed using a holistic model for an end-to-end map generalisation and a traditional semi-automatic map generalisation process. The experiment shows that the workflow approach is more promising than the holistic model, as each sub-task is specialised and fine-tuned accordingly. However, the results still do not reach the quality level of the semi-automatic traditional map generalisation process, as some sub-tasks are more complex to handle with neural networks.
Origine | Fichiers produits par l'(les) auteur(s) |
---|