The AI4EO Spring School 2026 provided participants with an intensive
introduction to modern Artificial Intelligence methods for Earth
Observation (EO). The event combined scientific lectures, practical
coding sessions, and interdisciplinary discussions concerning
geospatial foundation models, remote sensing workflows, deep
learning, and ethical AI applications. This report outlines the main
scientific topics covered during the Spring School, evaluates the
educational value of the sessions, and reflects on how the event
contributed to current developments in GeoAI research.
Day 1 / Foundation Models and Emerging AI Technologies
One of the central themes throughout the Spring School concerned the
increasing importance of foundation models within Earth Observation.
Several lectures demonstrated how large pretrained AI systems are
beginning to transform EO workflows through transferable geospatial
embeddings and multimodal learning approaches.
Particularly interesting was the ESA-oriented presentation on
emerging innovation technologies, which discussed:
- augmented intelligence,
- generative AI,
- explainable AI,
- onboard AI systems,
- neuromorphic computing,
-
and advanced computational paradigms for future EO
infrastructures.
Deep Learning for 3D Point Clouds
The lecture delivered by Iris de Gélis focused on
deep-learning-based change detection in raw 3D point clouds. The
session introduced advanced geometric deep-learning methods designed
for irregular spatial data and explained how modern convolutional
approaches can operate directly on point clouds rather than
rasterized representations.
Particular attention was given to Kernel Point Convolution (KPConv),
which extends convolution operations to unordered 3D point
distributions. The presentation discussed several Siamese KPConv
architectures developed for semantic change detection in urban
environments. A notable contribution involved the Encoder Fusion
SiamKPConv architecture, where temporal change information is
introduced directly into encoder layers to improve segmentation
performance.
The lecture clearly demonstrated the challenges associated with
sparse point-cloud structures and geometric neural networks. Unlike
classical image-based convolutional networks, point-cloud approaches
must learn directly from irregular spatial neighbourhoods and
geometric relations.
The practical importance of these methods became especially visible
when benchmark results on the Urb3DCD dataset were presented. The
experiments highlighted the influence of encoder-fusion strategies
and backbone architectures on semantic change-detection performance.
Practical Sessions and Reproducible Research
An important strength of the Spring School was the strong emphasis
on practical implementation and reproducibility. Several sessions
involved hands-on notebook-based experimentation using modern Python
frameworks and EO datasets.
The practical session conducted by Pierre Adorni focused on flood
mapping using:
- Sentinel-1 SAR imagery,
- Sentinel-2 optical imagery,
- and AI-driven segmentation workflows.
The tutorial demonstrated how multisensor Earth Observation
pipelines can support disaster monitoring applications through
multimodal data fusion.
Day 2 / TorchGeo
The presentation by Adam Stewart introduced TorchGeo, an open-source
geospatial machine-learning framework built on top of the PyTorch
ecosystem. The library is designed to simplify the development of
Earth Observation workflows by providing efficient tools for EO
dataset management, geospatial preprocessing, semantic segmentation,
and reproducible machine-learning experimentation. In addition,
participants were introduced to extensive tutorials and practical
resources available through the official TorchGeo documentation (
https://github.com/torchgeo/torchgeo
). A particularly valuable aspect of the session involved the
discussion of reproducibility standards within machine-learning
research.
Ethical AI and Societal Implications
A particularly engaging lecture during the AI4EO Spring School 2026
was presented by Pedram Ghamisi and Weikang Yu and focused on
Responsible AI in Earth Observation. The session explored how
Artificial Intelligence is increasingly influencing the analysis and
interpretation of geospatial data and why responsible development
practices are becoming essential within modern EO research.
The lecture covered topics such as bias mitigation, AI security,
geo-privacy, and the broader societal implications of AI-driven
satellite analysis. What I found particularly interesting was how
the speakers connected advanced deep-learning research with
practical challenges related to transparency, fairness, and secure
deployment of AI systems. The presentation clearly demonstrated that
Earth Observation technologies are no longer limited to purely
scientific applications, but are also becoming increasingly relevant
for environmental governance, urban planning, and large-scale
monitoring systems.
I found the depth of both Pedram Ghamisi’s and Weikang Yu’s research
especially impressive. The questions they raised concerning
responsible AI development, trustworthy geospatial models, and
secure AI infrastructures are highly relevant for the future of the
Earth Observation community and provided an important perspective
alongside the more technically focused lectures of the Spring
School.
Day 3 / Conditional Flow Matching and Generative AI
The lecture and practical session delivered by Nicolas Audebert
focused on Conditional Flow Matching and modern generative AI
methods. Although the underlying mathematics was highly complex, I
found the lecture particularly interesting because it demonstrated
how modern AI algorithms can be used to improve the resolution and
quality of satellite imagery. It was especially fascinating to see
how generative and probabilistic models are capable of
reconstructing missing spatial details and enhancing Earth
Observation data through learned feature representations.
Another aspect that I found very impressive was the discussion about
transforming SAR imagery into RGB-like visual representations. Since
SAR data is fundamentally different from optical imagery, the
possibility of converting radar information into visually
interpretable RGB images using AI-based approaches demonstrated the
enormous potential of modern generative models within Earth
Observation research.
The lecture successfully combined rigorous mathematical explanations
with intuitive visualizations and practical examples, making
advanced concepts easier to understand. Overall, the presentation
showed how generative AI is becoming increasingly important for
future EO systems, particularly in applications such as synthetic
data generation, uncertainty estimation, image enhancement, and
multimodal representation learning.
ESA Φ-lab and the GeoFM Challenge
A particularly valuable session during the AI4EO Spring School 2026
was the presentation of the “Reaching New Heights with GeoFM”
challenge by ESA Φ-lab in collaboration with the International
Telecommunication Union (ITU) and AI for Good. More information
about the challenge can be found on the official website:
ESA GeoFM Challenge Website
The challenge focuses on the use of Geospatial Foundation Models
(GFMs) for large-scale Earth Observation applications. The main
objective is to investigate how pretrained geospatial embeddings
generated by modern foundation models such as AlphaEarth, Tessera,
TerraMind, and THOR can be combined in order to perform semantic
segmentation and height estimation tasks. More specifically,
participants are asked to develop AI models capable of detecting
buildings and vegetation while simultaneously estimating their
heights using openly accessible satellite imagery.
What made the challenge especially relevant was its strong
connection to real-world Earth Observation problems. Traditionally,
highly accurate Digital Surface Models (DSM) and Digital Terrain
Models (DTM) rely on expensive airborne LiDAR campaigns, which are
not available worldwide. The GeoFM challenge therefore explores
whether scalable AI-driven alternatives can be created using only
publicly available EO data.
The presentation clearly demonstrated how modern Earth Observation
research is increasingly shifting toward multimodal AI systems and
transferable geospatial embeddings instead of traditional
handcrafted workflows. It also highlighted how foundation models are
becoming universal feature extractors for remote sensing
applications. Overall, the challenge provided a very good overview
of current research directions in GeoAI and illustrated how strongly
Earth Observation is evolving toward large-scale AI-driven
infrastructures for environmental monitoring, urban analysis, and
climate-related applications.
Learning Outcomes
Overall, the AI4EO Spring School 2026 provided an excellent overview
of current developments in Artificial Intelligence for Earth
Observation. One of the most valuable aspects of the event was the
successful combination of theoretical lectures, practical notebook
sessions, reproducible workflows, and interdisciplinary discussions.
Several major scientific trends became especially visible throughout
the Spring School. These included the rapid emergence of foundation
models, the increasing importance of multimodal EO embeddings, the
growing role of generative AI, and the necessity of ethical and
reproducible AI development.
The practical coding sessions significantly strengthened the
educational value of the event because participants were able to
directly interact with state-of-the-art machine-learning methods and
geospatial AI infrastructures. In addition, the Spring School
created an excellent environment for scientific exchange between
researchers, students, and practitioners from diverse backgrounds.
Conclusion and Recommendations
In conclusion, the AI4EO Spring School 2026 was exceptionally well
organised and scientifically highly valuable. The event successfully
combined theoretical depth, practical implementation, and
interdisciplinary discussion within a rapidly evolving research
field.
Particularly successful aspects included the balance between
advanced scientific lectures and practical implementation, the
strong emphasis on reproducibility and open science, and the
integration of modern AI methods such as foundation models and flow
matching into Earth Observation workflows.
The Spring School clearly demonstrated how rapidly GeoAI is evolving
and how important interdisciplinary collaboration has become within
modern Earth Observation research. Similar Spring Schools should
definitely continue to be organised in the future because they
provide an outstanding platform for scientific exchange, practical
learning, and discussion of emerging challenges in Artificial
Intelligence for Earth Observation.
