• Post category:Articles
  • Post published:27.3.2024

How can AI help us work smarter? The state of GeoAI in Finland

In all workplaces, people have been discussing the possibilities and benefits of artificial intelligence (AI) in their fields. Naturally, this discourse has also been strong in the field of geospatial data, which utilises a huge amount of data.

Author

You are currently viewing How can AI help us work smarter? The state of GeoAI in Finland
Geospatial Artificial Intelligence, or GeoAI, is a hot topic in the geospatial industry. This is AI's vision of what the synergy between artificial intelligence and geospatial data could look like.

Researchers and companies have been actively developing AI methods for tasks such as satellite image analysis and traffic planning optimisation. The spectrum of solutions is vast, given that geospatial data is ubiquitous and diverse.

For example, researchers at the Finnish Geospatial Research Institute have developed deep learning methods for identifying changes in aerial images and point clouds, detecting damage to forests caused by bark beetles, and identifying anomalies in GNSS signals from large data sets. On the private sector side, ongoing developments include AI methods facilitating the digitisation and indexing of zoning regulations, optimisation methods for land-use planning, and various data processing automations. Much more is continuously being developed.

Looking Towards the Future

Although many AI solutions developed in Finland are excellent and address key challenges in the geospatial field, the GeoAI scene still lags behind other parts of the world, especially the United States and China. Currently, most solutions focus on relatively narrow problems one at a time. Often, the bottleneck has been either a lack of expertise or an insufficient amount of high-quality training data.

The next clear step in the GeoAI world could be the broader utilisation of spatial autocorrelation. In this approach, a wide range of geospatial data could be used as training data, much like ChatGPT uses a large portion of the internet regardless of its specific format. Leveraging spatial autocorrelation would allow GeoAI training datasets to be aligned based on location, enabling models to understand not only the content (or modality) of the data but also its spatial relationships.

For instance, multimodal geofoundation models (GeoFMs) could analyse geospatial structures and shapes from satellite images, extract information from text reports or events in the area, and combine them with sensor data. Models could then take into account an increasing number of factors and act as generative general models for specific geographical areas, also accounting for problems that have not been considered in the training phase.

Many might associate the description above with other geospatial buzzwords such as the metaverse, digital twins, or data spaces. It is essential to consider the role of AI in digital twins or data spaces and whether entire systems should be built on the foundation of AI. How much unnecessary work do humans currently perform that could be delegated to AI?

General GeoAIs Are Still Missing

Currently, research into GeoFM is still in its infancy. Perhaps the best, but misleading, example of such operations is given by the well-known ChatGPT. Users can ask questions like “How many churches are there in Helsinki?” and ChatGPT will respond to the best of its ability.

Screenshot from ChatGPT: The user has asked How many churches there are in Helsinki? The answer is: The number of churches in Helsinki varies, but there are approximately 80 churches in Helsinki.
ChatGPT’s response to a difficult question.

For that question, ChatGPT provides an approximate answer of 80, which is not a bad estimate. The actual number in the year 2024 is around 66 churches or religious spaces according to the Helsinki service map. Additionally, the number of churches has decreased in recent years, making the estimate at least better than that of many individuals.

Users can continue the inquiry by requesting a GeoJSON file for the churches, and ChatGPT can generate the syntax accurately. Therefore, the results can be directly visualised in QGIS.

The GeoJSON representation of the coordinates for the 10 churches in Helsinki, given by ChatGPT.
ChatGPT’s response in GeoJSON format.
Screenshot of a map showing Uspensi Cathedral's location in Helsinki.
The blue dot indicates the actual Uspenski Cathedral location, and the red dot indicates the coordinates provided by ChatGPT.

Of course, as a general language model, ChatGPT does not understand complex spatial relationships or data, so the deviation between real churches and resulting coordinates is a few hundred metres for each church. However, the approximate area is surprisingly well estimated. The information provided could be useful for some purposes, if the data accuracy is not so relevant. For more general use of GeoFMs, there is still a lot to do before reaching truly general GeoAI or GeoFM solutions.

Interested in GeoAI solutions more broadly? Contact Location Innovation Hub and tap into our extensive network of companies and researchers. It includes a vast pool of expertise in various AI methods, training data sets and referense solutions. Almost all our services are free.