- Docente responsabile
- MARIA ANTONIA BROVELLI
- CCS proponenti
- Geoinformatics Engineering - Ingegneria Geoinformatica Ingegneria Informatica
- CFU
- 2
- Ore in presenza
- 16
- Prerequisiti
- Knowledge of Python; Knowledge of Machine Learning fundamentals
- N° max studenti
- 25
- Criteri di selezione
- Priority given to Bachelor's degree students. Average grade.
- Parole chiave:
- Geospatial Foundation Models
- Tag
- Environment and land planning, Computer science, Artificial intelligence, Surveying and monitoring, Natural resources
Descrizione dell'iniziativa
Foundation Models (FMs) represent a transformative shift in AI, enabling reusable and adaptable models trained on massive unlabelled datasets. Geospatial Foundation Models (GFMs) apply this approach to satellite imagery, supporting a wide range of applications such as land use mapping, environmental monitoring, and disaster response.
This Passion in Action course combines theoretical foundations with hands-on experience using two state-of-the-art GFMs: Prithvi-2.0 (NASA-IBM) and TerraMind (ESA-IBM).
By the end of the course, students will have both theoretical understanding and practical experience in applying GFMs to real-world environmental and geospatial challenges.
Detailed Program
- Geographic Information and Earth Observation Introduction (4 hours):Locating points on the surface of the Earth:Reference and Coordinate Systems. Map and projections. Satellites and images. Description of the training dataset used fo GFMs. Overview of GFMs, their relevance in geospatial context and potential applications.
- Deep Learning Introduction (4 hours): Neural Networks: from a single perceptron to a Multi-layer Perceptron. Computer vision: learning from images with Convolutional Neural Networks. Transformers, Vision Transformers, and Foundation Models. Self-supervised learning on images with Masked Autoencoders
- Prithvi-2.0 (4 hours)
- Overview of the architecture, pre-training dataset, capabilities and limitations
- Hands on exercises on real data to fine-tune the model for different geospatial applications, such as land cover mapping, illegal runway detection, flood mapping using SAR data. The exercises will be conducted on Google Colab using the Python programming language.
- arxiv.org/abs/2412.02732
- huggingface.co/ibm-nasa-geospatial/Prithvi-EO-2.0-300M
- TerraMind (4 hours)
- Overview of the architecture, pre-training datasets, capabilities and limitations
- Hands on exercises on different data types, showcasing the range of possibilities of TerraMind. The exercises will focus on fine-tuning the model for different geospatial applications presented in the previous session using Prithvi, plus additional use cases specific to TerraMind's capabilities. The exercises will be conducted on Google Colab using the Python programming language.
- arxiv.org/abs/2504.11171
- huggingface.co/ibm-esa-geospatial/TerraMind-1.0-base
Periodo di svolgimento
dal October 2025 a October 2025
Calendario
Schedule
8 October 2025 16:00 - 20:00
16 October 2025 16:00 - 20:00
22 October 2025 16:00 - 20:00
27 October 2025 16:00 - 20:00