Geo AI: Expert Interview on Artificial Intelligence in Geoanalytics

In your opinion, how big is the potential for artificial intelligence in geoanalytics (Geo AI)?

Geo AI plays a role in every area of geoanalytics. It is a very large field of research, but we do not yet know exactly where it will bring real advantages over classical spatial analyses. It is still too early to say whether Geo AI has the potential to revolutionize Geomarketing tools.

Where is Geo AI helpful today, in 2025?

You can already do a lot with Geo AI models today, especially in the area of evaluating geodata and identifying trends and relationships in data, provided you have well-trained models. However, we very often see in projects that there is a lack of sufficiently high-quality training and verification data.

Can you share some examples from your research on Geo AI?

In the recently completed project "ABM4EnergyTransition", we aimed to find out which criteria homeowners use to decide on their heating system or the thermal insulation of their private homes. For this purpose, a large survey of 30,000 households was conducted, but we also incorporated classic sociodemographic data provided to us by WIGeoGIS into the AI model. Artificial intelligence was a great help in this project in finding out how, among other things, the financial situation, educational status, and public funding influence homeowners' decisions.

In another research project, we want to use Geo AI together with TU Graz to find out how factors such as climate change, socio-demographic change, and mobility behavior will influence the stability of power grids until 2030.

How exactly do you work with Geo AI?

In the example mentioned, we want to use highly precise smart meter data, i.e., electricity consumption data and electricity generation data, which are aggregated, i.e. summarized, so that data protection is guaranteed. This means that the data does not reveal which household turns on the television and starts cooking at exactly what time. We also incorporate sociodemographic data and data from a mobile phone provider into the AI model, which tells us something about commuter behavior, for example. We try to process and overlay all of this data in such a way that the AI model tells us something about the future utilization of the power grid. A research question is at which level we can make predictions, for example, at the neighborhood level or the building block level, although there will always be uncertainties.

What is the difference between ChatGPT and Geo AI models?

Geo AI always deals with questions with a geographical reference. Moreover, large language models, such as ChatGPT 4.0, are aimed at end consumers and are trained using a dataset that is available at a specific point in time. Users interact with the model via prompts. We, on the other hand, train our own AI models. This means we first have to prepare the data we receive, and that is not as easy as it sounds.

Where do you see the biggest challenges in training Geo AI models?

One of the biggest challenges is data availability, where we need not only a certain quantity of training data, but also the highest possible quality. AI models learn similarly to children: The more often they come into contact with something, the better and faster they can recognize patterns: "Ah, that scratches, I know that, that must be a cat. Oh, it is barking, so it must be a dog." The more dogs and cats a child sees, the more likely they are to memorize the patterns and distinguish them better. Only once the algorithm has seen enough, can it also recognize patterns and assign categories using statistical probabilities.

In geomarketing, we often try to predict human behavior. How well are we already able to do this today?

This is a very big challenge that we have been dealing with for several years. For example, in one project we modeled tourism flows in Salzburg. It is important to have a variety of different data available in order to recognize people's movement patterns. This means that we incorporate as many decision-making levels as possible. For such tasks, we like to work with hybrid models, specifically a combination of agent-based modeling, which involves simulation, and the detection of neighborhood effects, and artificial intelligence. AI can uncover patterns within the data that would be difficult or even impossible for a human to detect.

To what extent can the results of Geo AI be trusted, especially considering that the same question does not always produce the same result?

That is true, although even an agent-based AI model or ChatGPT does not necessarily deliver the same result for real-time queries. The same is true for Geo AI: The overall trends will be similar, but the details may vary. For example, when searching for a location, artificial intelligence might consistently display a location within a certain region or grid cell, but the specific location could vary each time. To ensure explainability, it is important to clearly communicate to users and customers the level of detail at which the AI can deliver insights, for example at the building level, neighborhood level, etc. Sometimes the source data is not that granular and you need to choose a higher aggregation level.

How well can AI recognize patterns in unstructured mass data?

Artificial intelligence is already very good at extracting patterns, features and trends from unstructured data and structuring them, thus making knowledge explicit. AI models, like deep learning, provide us with excellent opportunities to turn unstructured data into partially structured or fully structured data and, if necessary, to georeference it, i.e., to assign a location to the data. For example, in a project with the Austrian Armed Forces, we analyzed unstructured media data with Geo AI and were able to identify and geographically locate disaster areas based on the content of media reports. This solution worked quite well.

In your opinion, in what stage of development is Geo AI currently?

We need to make distinctions here. The field of Earth Observation within geoinformatics, for example, is already highly advanced. There are already many standardized image analysis methods that work extremely well. For instance, a colleague at TU Graz successfully developed an AI system that automatically determines whether roofs are equipped with thermal solar panels or photovoltaic modules. In contrast, when it comes to human behavior, for example, when creating movement analyses and forecasts, further development is still needed. While good progress is currently being made in this area, there are still no standardized methods. Especially in Europe, compared to the USA, China, and Korea, we still need significantly more research initiatives.