Forschungsprojekt
GeoWiKI – Integrating Geographical Knowledge into Artificial Intelligence
AI models are increasingly being used successfully in geographical applications, including the image-based mapping of geographical objects such as roads, buildings or land use areas. However, the training of these models is very energy-intensive and the results sometimes include obvious misclassifications, as the models are purely data-driven and do not include any knowledge about the typical or logical arrangement of objects in space. Figure 1 shows an example for this, where roads detected by a satellite image-based AI model are shown in an agricultural area.
Figure 1: Incorrectly detected roads (in red) by a satellite image-based AI model developed and published by Microsoft. (Image: own representation, data: Microsoft/RoadDetections, 2021/2024)
Humans, even without having domain expertise, can easily detect that this is a misclassification, as they include the spatial context along with a plethora of additional intuitive observations when examining geographical objects (e.g. roads should be connected to a road network and have reasonable curve characteristics among others). This knowledge of spatial relations and arrangements, which is a mixture of intuition based on common sense, logic, and learned concepts, know-how based on scientific observations and analyses, as well as rules based on laws and physical restrictions, can be summarized as geographical knowledge.
The unique problem in the geographical realm is that this knowledge cannot be represented as vectorized data, i.e. values defined by specific differential equations with an absolute validity in all cases, the only type of data AI models can understand. This is the main difference to other research fields like physics, where all things can be described by deterministic equations. However, we believe that existing geographical objects in the real world already indirectly incorporate geographical knowledge, as humans tend to act based on such knowledge.
As AI models are usually data-driven without general knowledge about the system they are applied in, the learning and training process is elongated drastically and made much more complex as they have to learn the rules of the system from scratch, opening the possibility of learning misleading unrealistic misconceptions. In this project, we want to investigate whether the training time and thus the energy consumption of AI models can be reduced and their performance increased by integrating geographical knowledge into them. This includes tackling three main research questions:
- What is geographical knowledge and what are methods of representing it as vectorized data?
- What are effective methods to include such knowledge into AI models?
- How do such implementations impact the models’ efficiency and predictive performance?
As it is the world’s largest community-based Volunteered Geographic Information mapping project, OpenStreetMap can provide a suitable database to extract this knowledge and enrich it in a way that AI models can understand and utilize it. Through its open design, it enables the generalization of the concept and its application to a plethora of different tasks.
The project is funded by the Vector Foundation and worked on by the GIScience Research Group in cooperation with the HeiGIT gGmbH from October 2024 until December 2025.
About GIScience Research Group
The GIScience Research Group at Heidelberg University focuses on innovative basic and applied research as well as on the latest technology at the interface between geography and computational sciences. Due to the very close cooperation with the HeiGIT, it is possible to realize state of the art research, finally enabling the development and production of suitable tools.
About HeiGIT gGmbH
The aim of the Heidelberg Institute for Geoinformation Technology (HeiGIT) is to improve the transfer of knowledge and technology from basic geoinformatics research into practice by using innovative geoinformation technologies. It is scientifically directed by Prof. Dr. Alexander Zipf and funded by the Klaus Tschira Stiftung gGmbH. As an affiliated institute of Heidelberg University, there is close cooperation with the staff of the Department of Geoinformatics at the Institute of Geography. Hence, innovative solutions can be realised at the cutting edge of research and technology.
About Vector Foundation
The Vector Foundation was established as a company-affiliated foundation. It is an expression of gratitude for the success achieved and ensures the long-term existence of the company. The most important driving force behind the foundation's work is the desire to tackle social challenges effectively.
Microsoft/RoadDetections. (2024). [Computer software]. Microsoft.
https://github.com/microsoft/RoadDetections (Original work published 2021).
