ASU Workshop on Geospatial AI Foundation Models

Welcome to the hands-on materials for the ASU Workshop on Geospatial AI Foundation Models.

This workshop introduces geospatial foundation models for Earth observation and walks through a practical example using the NASA-IBM Prithvi model for landslide mapping from satellite imagery.

Workshop Exercise

The main hands-on notebook is:

• workshop_materials/prithvi_workshop_final.ipynb

In this notebook, you will fine-tune a Prithvi geospatial foundation model for binary semantic segmentation:

• Input: Sentinel-2 optical bands, including Blue, Green, Red, NIR, SWIR-1, and SWIR-2
• Output: a landslide mask
• Model: Prithvi encoder with a UNet decoder
• Metrics: precision, recall, F1-score, and IoU

The notebook covers:

1. Exploring HDF5 satellite data
2. Building a custom PyTorch dataset
3. Computing channel-wise normalization statistics
4. Applying data augmentation and resizing for Prithvi input
5. Building a TerraTorch Prithvi + UNet segmentation model
6. Training, validation, checkpointing, and test evaluation
7. Visualizing predictions and error maps

Files

• workshop_materials/prithvi_workshop_final.ipynb: main workshop notebook
• workshop_materials/config_prithvi300_UNet.yaml: model configuration for Prithvi-EO-2.0 300M with a UNet decoder
• workshop_materials/readme.md: short note for the workshop materials folder

Data And Checkpoints

Participants should download the required landslide reference dataset and Prithvi checkpoint before running the model-building and training sections.

• Landslide reference dataset: https://zenodo.org/records/17007637
• Prithvi-EO-2.0 300M model files: https://huggingface.co/ibm-nasa-geospatial/Prithvi-EO-2.0-300M/tree/main
• Direct checkpoint download: https://huggingface.co/ibm-nasa-geospatial/Prithvi-EO-2.0-300M/resolve/main/Prithvi_EO_V2_300M.pt?download=true

When Jupyter is launched from workshop_materials, the notebook path setup cell defaults to:

• Data: ./data/reference_data
• Prithvi checkpoint: ./checkpoints/Prithvi_EO_V2_300M.pt
• Fine-tuned training output: ./prithvi_unet_landslide_best.pt

These correspond to the repo paths workshop_materials/data/reference_data, workshop_materials/checkpoints/Prithvi_EO_V2_300M.pt, and workshop_materials/prithvi_unet_landslide_best.pt.

If you store the files somewhere else, update DATA_DIR and BACKBONE_CKPT_PATH in the notebook path setup cell. You do not need to edit workshop_materials/config_prithvi300_UNet.yaml for local paths.

Python Environment

The notebook uses Python 3 with common geospatial and deep learning packages, including:

• torch
• torchvision
• h5py
• numpy
• pandas
• matplotlib
• tqdm
• pyyaml
• terratorch
• segmentation_models_pytorch

A CUDA-enabled GPU is recommended for training and evaluation.

Suggested Workflow

1. Follow workshop_materials/readme.md to create the environment, download files, and launch Jupyter from workshop_materials.
2. Open prithvi_workshop_final.ipynb.
3. Run the notebook cells in order.
4. Confirm that DATA_DIR and BACKBONE_CKPT_PATH point to your downloaded files.
5. Follow the section notes in the notebook to inspect data, train the model, and visualize predictions.

References

• TerraTorch: https://github.com/terrastackai/terratorch
• Prithvi-EO-2.0: https://github.com/NASA-IMPACT/Prithvi-EO-2.0
• Prithvi-EO-2.0 300M model card: https://huggingface.co/ibm-nasa-geospatial/Prithvi-EO-2.0-300M