Sociotechnical Foundations of GeoAI and Spatial Data Science

Hard Facts

Date and Time
October 25, 2024 - arrival day
October 26-27, 2024 - meeting
October 28, 2024 - departure day

Springer Schlössl, Vienna, Austria


We are excited to announce a specialist meeting on “Sociotechnical Foundations of GeoAI and
Spatial Data Science” which will take place at the Springer Schlössl in Vienna, Austria on
October 26-27, 2024. We will offer accommodation and travel support for around 30 participants
across career stages, geographic regions, and academic/industry backgrounds. The meeting will
provide an opportunity to discuss the sociotechnical and ethical foundations of GeoAI and
Spatial Data Science in the context of recent advances in generative AI and foundation models.
With recent breakthroughs in foundation models, such as large language models and text-to-
image models in AI and GeoAI, there is an urgent need to develop a community-driven roadmap.
This roadmap will help us to positively and actively shape the next five years by providing our
(geo)spatial perspective to the broader AI community. Otherwise, we risk being passively shaped
by those next five years.

We are already seeing first research on developing geo-foundation models. However, the costs
and computational resources required for training, tuning, and deploying these models may
exceed what most individual labs (or even universities) can handle. For instance, training large
language models or text-to-image models is known to cost hundreds of thousands of dollars.
While the science community is currently trying to catch up by closing the performance gap
between super-large and smaller models, we believe that future GeoAI and Spatial Data Science
research may benefit from a similar approach as implemented by the Physics community. Such
an approach involves forming a joint community-wide consensus to inform funding agencies and
donors about the future long-term research roadmap, e.g., for missions such as the Kepler space
telescope, that benefits the entire community.

The specialist meeting (together with related activities such as online seminars and the
“GeoMachina” workshop) aims at identifying and discussing the sociotechnical foundation of
GeoAI, Spatial Data Science, and geo-foundation models from an ethical perspective in order to
prepare and positively shape the AI-based disruptions ahead of us. Reflecting on our profession’s
ethical implications will assist us in conducting this potentially disruptive research more
responsibly. It will assist us in identifying pitfalls in designing, training, and deploying GeoAI-
based systems, and developing a shared understanding of the benefits but also potential dangers
of artificial intelligence and machine learning research across academic fields, all while sharing
our unique (geo)spatial perspective with others.

To give just one example why such bi-directional exchange is important, it is worth mentioning
that currently AI teams from Google Brain, Sony, and others are trying to understand potential
coverage and representational biases in their training, validation, testing datasets, and models.
They do so by studying their ‘geo-diversity’ and using terms, e.g., the Modifiable Areal Unit
Problem, and technologies that originated in geography and GIScience. Put differently, our skills
and methods benefit the broader AI community.

To provide structure to our discussions, we aim at covering the following key topics:
AI Sustainability: Training of an AI system can cause carbon emissions equivalent to hundreds
of flights across the US. This does not take into account the cost of adjusting and deploying these
systems nor the cradle-to-grave emissions generated through manufacturing, transporting, and
recycling the required hardware (which are substantially higher yet). Given that geo-foundation
models may need substantially more frequent retraining, our community should progress using
GreenAI methods instead of RedAI, where progress is essentially bought through research
consumption thereby excluding most competition. Additionally, the metrics used to measure how
environmentally friendly current machine learning systems are, rely on (over)simplistic models
of space and geography, e.g., by ignoring the population affected by negative environmental
impacts in relation to the benefiting population. Put differently, it needs geospatial analysis
methods to properly quantify how sustainable current progress is as well as a better
understanding of ownership and governance structures. Interestingly, there are first signs that
foundation models don’t necessarily have to be very large (and, thus, resource intensive) to
provide good results if the underlying architectures are improved.

Bias and Debiasing: Are training datasets, pre-processing, neural architectures, evaluation
criteria, prompt engineers, and users (feeding back into the system) biased? What types of bias
are specific to (geo)spatial data and models? How do researchers and practitioners in the broader
AI community think about geo-diversity, and can we contribute new perspectives? What types of
biases affect geographic data, e.g., VGI, and how can we mitigate them? Given that debiasing
will be done algorithmically, how biased will debiasing be? These are just some of the questions
that the GeoAI (and broader AI) community is currently facing, and that cannot be resolved just
by the technical community alone. For instance, debiasing on the data level may lead to models
that more accurately reflect social aspirations at the cost of masking realities expressed by the
original data, which do not (yet) reflect these social and political aspirations. If training data
sources for text-to-image models contain only 3.1% of images from China and India combined,
is this reflected in the way foundation models represent geographic space? Can we as
geographers and spatial data scientists contribute measures of geo-diversity back to the global AI

Schema and Data Diversity: Foundation models rest on the assumption that pre-trained models
of sufficient size can be used across domains and downstream tasks. However, this may neglect
regional variability and lead to less accurate results overall. It is important that models are
trained on a diverse set of datasets across several data types (modes) and that diversity also
includes variability in the schema knowledge underlying these datasets. So far, data diversity is
purely approached from a perspective of representativeness, e.g., of a given data collection.
Local/regional differences in schema knowledge are not broadly taken into account despite their
importance, e.g., due to varying laws, being widely recognized. Given that most AI chatbots are
now utilizing retrieval-augmented generation (RAG) to connect to knowledge graphs and other
data sources to retrieve data instead of dreaming them up, how are these data sources prioritized?
Where does their schema knowledge come from? How would we provide data for a GeoRAG?
GeoAI Neutrality: Given that geo-foundation models will impact how we learn about the world,
and, in a second step, also how we act in the world, it is crucial to understand whether GeoAI
methods are neutral, and if they are not, at which stages, e,.g., data curation, unbalanced results
are introduced. The current lack of consensus within our own community about what algorithmic
neutrality means and implies are posing substantial challenges to our ability to positively shape
the disruptions of (Geo)AI that society will likely face over the next five years. In most cases,
lack of neutrality arises from data curation, during prompt engineering, by selecting certain data
and not others, by inadequately matching the training task to future downstream applications,
but also due to issues of ownership and governance, and so on. Can we develop clear definitions
of GeoAI neutrality and guidelines to achieve it?

Disruption: Ahead of us lie disruptions that make the invention of the Internet pale in
comparison. We must shape these next years actively and positively instead of being shaped by
them. Our community has a lot to offer to the broader AI community; however, the costs of
contributing to the current state of the art, e.g., geo-foundation models, are very high, and the
required hardware, storage, and deployment costs cannot easily be handled by single research
groups and often not even by universities alone. Hence, it is important that we as a community
jointly form a research agenda similar to how this has been done in (Astro)physics for decades.
Agreeing on such a community goals—driven research agenda and approaching funding
agencies with such proposals requires a clear understanding of the benefits and risks of
developing and deploying the geo-foundation models of the near future. If we develop a joint and
informed consensus, the benefits of current AI developments will far outweigh the drawbacks
and potential risks. Communicating this optimism while informing about risks ahead is also key
to educating the future Spatial Data Science workforce.
Format of the Specialist Meeting

The workshop will be held over two full days October 26-27, 2024 with arrival and departure
days before and after. For the arrival day, we will also provide opportunities to jointly explore
Vienna. We will keep the tradition of offering a morning hike alive.

We invite colleagues from all disciplinary backgrounds, career stages, geographic regions,
genders, and ethnicities to apply. We kindly request all potential applicants to fill out the linked
, including a brief biography (up to 200 words),
and a one-pager (400-600 words) detailing their motivation to participate in the meeting. The
one-pagers and biographies will be published on the meeting’s webpage and inform the
discussion. The deadline for applications is July 17, 2024. We aim to provide accommodation
and travel support for around 25-30 external attendees and, therefore, about 35 participants
overall. While the meeting will focus on discussions, each participant will also have the
opportunity to present a lightning talk during the opening session. Other roles will include
panelists, keynote speakers, and so forth. All participants will be co-authors on the meeting
report. Please also use the opportunity to participate in our additional activities that we will offer
before and after the in-person meeting such as our open-access book on “Geography According
to ChatGPT”, our “GeoMachina” autonomous GIS analyst workshop at SIGSPATIAL 2024, the
Spatial Data Science Symposium SDSS 2024, and other webinars.

Please feel free to reach out to Krzysztof Janowicz ( for further
questions and /or Daniela Woelfle ( for administrative requests,
e.g., with respect to the application form.

We gratefully acknowledge support from Esri, AAG, and the University of Vienna.

Krzysztof Janowicz, University of Vienna
John Wertman, Esri
Mike Goodchild, University of California, Santa Barbara
Gary Langham, AAG
Coline Dony, AAG