MSc in Geoinformatics Engineering, School of Civil, Environmental and Land Management Engineering, Politecnico di Milano. 5 credits. First taught in 2026/27.
I teach this course together with Vasil Yordanov, who holds it, and Federico Giannini. Vasil brings the geospatial side, I bring the artificial intelligence side, and the course lives in the overlap.
What the course is about
Earth observation has more data than anyone can label, and that is exactly the condition under which foundation models are interesting. The course gives students a critical and practical understanding of recent AI methods for geospatial data and workflows: transformer-based architectures, geospatial foundation models, multimodal data, model adaptation, and reproducible open-source pipelines.
The emphasis is on judgement rather than recipes. Through real geospatial use cases, students learn how to select a method, adapt it, and evaluate it against the data they actually have — taking seriously the model’s limitations and the reproducibility of what they build. A dedicated thread looks at what Large Language Models can and cannot do for geospatial reasoning: natural-language interfaces, code generation, documentation, metadata querying, and the reliability problems that come with all of them.
What students take away
By the end of the course a student knows the principles of transformer architectures in a geospatial context, understands the role of geospatial foundation models and multimodal representations, and can prepare datasets, implement a workflow on pretrained models, fine-tune one, and judge where it holds and where it breaks. The final step is the hardest: designing a complete GeoAI workflow, justifying the methodological choices behind it, and explaining the result to someone else.
Topics
- Deep learning foundations for spatial data — feed-forward networks, gradient-based optimisation, backpropagation, convolutional networks, spatial dependencies in images and cross-channel correlation in multi-band inputs.
- Sequential models and language-model foundations — recurrent networks and their limits, hardware parallelisation, word embeddings, text transformers, self-attention, pretraining.
- Transformers, Vision Transformers and the foundation-model paradigm — images as sequences of visual tokens, and what is gained and lost moving from the local inductive bias of CNNs to global attention.
- Geospatial foundation models for Earth observation — model families pretrained on satellite imagery and multimodal EO data, and general-purpose vision models adapted to geospatial contexts.
- Preparing geospatial data — access, preprocessing, tiling, spatial resolution, coordinate reference systems, masks, labels, dataset organisation.
- Multimodal learning and shared semantic spaces — aligning text and vision in a joint embedding space, contrastive frameworks such as CLIP, lightweight fine-tuning.
- Adapting and fine-tuning foundation models — task formulation, training configurations, data and label requirements, and the choices that decide transferability.
- Open-source ecosystems for GeoAI — notebook environments, libraries for geospatial foundation models, and Python tooling for data handling and experimentation.
- Applications — segmentation, classification, temporal analysis, multimodal retrieval, land-cover analysis, environmental monitoring, flood mapping, wildfire analysis.
- LLMs and generative interfaces for geospatial workflows — prompt design, code generation, human-AI collaboration, tool calling and introductory agent-based workflows for GIS and EO tasks.
- Project work — running through the semester, with periodic reviews.
Project work
The project runs alongside the lectures for the whole semester: students implement, adapt, evaluate and present a reproducible GeoAI workflow built on foundation models and open geospatial datasets. Periodic reviews are there to discuss methodological choices while they can still be changed, not to grade progress.
Prerequisites
None. The course is designed for students arriving from different backgrounds and builds up the necessary knowledge along the way.
The course contributes to UN 2030 Agenda goals SDG 11 (sustainable cities and communities), SDG 13 (climate action) and SDG 15 (life on land).
Official course sheet
The full syllabus, assessment rules and bibliography are on the Politecnico di Milano course sheet for 064255 — Advanced Geospatial Artificial Intelligence.