Research Collaboration on Artificial Intelligence for Topographic Mapping

• Forschungsstelle: SWISSTOPO
• Projektnummer: ECEO 2021-01
• Projekttitel Englisch: Research Collaboration on Artificial Intelligence for Topographic Mapping

Schlüsselwörter (Englisch)

topographic mapping, deep learning, machne learning, aerial imagery, image segmentation, map update, siamese CNN,

Kurzbeschreibung (Englisch)

1. Nowadays, we can map the territory with very high accuracy. Given high quality image data, historical records and a team of trained photointerpreters, topographic maps can be updated at regular cycles and continue representing the reality of the field. Recently, machine learning and artificial intelligence have been massively put to the purpose of land cover and land use mapping with excellent results. This technology is now explored at the level of the Swiss Federal Office for Topography, notably via collaborations with the ECEO group of EPFL. Following these encouraging and successful first collaborations, ECEO and swisstopo would like to push collaborations further in a joint research project focusing on the efficiency of updating the topographic mapping layers.

2. The update process of topographic mapping layers at swisstopo is successful, but also very labor and time consuming. The work of the team of (photo-) interpreters could be made more efficient and exciting by a system that would focus their attention on critical features, land cover/use classes and situations where an update is likely to be necessary.

3. Following these requirements, the contractor has to explore solutions based on interactive updates: He has to design update strategies that first select the most interesting examples and then present them to trained human operators. By this interaction between human and machine, also known as active learning, the updates are kept efficient and the effort of the operators is put only where it is necessary. Opposed to traditional mapping, this strategy focuses on acting only where it is truly needed and has the potential of drastically reducing the amount of human effort (and the costs involved) for the update of the maps. At the same time, the two approaches can co-exist, where manual verification and reviewing is still part of the mapping process.

Projektziele (Englisch)

The objective of this research is to design image recognition models that assist (photo-) interpreters to map accurate topographic feature layers and update them efficiently. To do so, advanced deep learning algorithms and smartly designed data collection strategies with human operators in the loop will be investigated. The focus is on critical classes for swisstopo, in particular alpine land cover and classes of instability.