Resources

Here we list a bunch of useful resources for getting started with geospatial machine learning. By no means is this an exhaustive list, but just a few examples that ML4GEO members have found useful.

For a more comprehensive list the one provided by satellite-image-deep-learning is pretty awesome! 😎

Intro to GeoML

Open Access Courses:

Other introductory resources:

  • Pytorch Tutorials: Popular machine learning library for python.
  • Pytorch Lightning: High-level interface for PyTorch
  • Torchgeo Tutorials: The package is an extension to PyTorch and Pytorch Lightning to include popular datasets, model architectures, and common image transformations for geospatial data.

Geospatial foundation models

Below we’ve collated a list of helpful links and tutorials to get you started on Geospatial foundation models. By no means is this an exhaustive

Foundation Models / Tutorials

  • TerraTorch: flexible fine-tuning framework for Geospatial Foundation Models (GFMs) based on TorchGeo and Lightning, supports models from the Prithvi, TerraMind, and Granite series as well as models from TorchGeo and timm.

    • https://github.com/IBM/terratorch/tree/main/examples/notebooks

  • TerraMind: The latest GFM from ESA, can be applied for classical deep learning tasks, as well as generative tasks (e.g. S1 from S2 data)

    • https://github.com/IBM/ML4EO-workshop-2025
    • Examples of disaster response with S1 and S2; multi-temporal data crop segmentation

  • Google Satellite v1 Embeddings:

    • We have scripts to help download rasters of embeddings with python API

  • Earth Index:

    • https://www.earthgenome.org/earth-index
    • No coding tool with great interface to find similar satellite images tiles in an active learning environment, currently early access only

  • IBM-NASA Prithvi Models: Three foundation models have been released to date: Prithvi-EO-1.0 and Prithvi-EO-2.0 which uses earth observation data from NASA’s Harmonized Landsat and Sentinel-2 (HLS), and Prithvi-WxC-1.0 which uses weather and climate data from NASA’s MERRA-2.

    • https://huggingface.co/ibm-nasa-geospatial

  • SatCLIP: Predict location coordinates given satellite imagery

    • https://github.com/microsoft/satclip/tree/main/notebooks

  • Clay: EO foudnation model trained on Landsat, S2, S1, NAIP, LINZ, MODIS

    • https://clay-foundation.github.io/model/index.html
    • Has a nice tutorial visualising embeddings

  • DOFA: a unified multimodal foundation model for different data modalities in remote sensing and Earth observation.

  • Aurora: A Foundation Model for the Earth System.

Embeddings - Things to Consider when using the Google Ones

  • https://www.linkedin.com/posts/mbforr_ai-isnt-magic-its-math-and-sometimes-activity-7353415998424711169-CKsX?utm_source=share&utm_medium=member_desktop&rcm=ACoAADV71oUBf4-Yt6U7QfjkZLmIRivB7oVFbHA

Technical Understanding of Key Concepts

  • https://developers.google.com/machine-learning/crash-course/embeddings/embedding-space

Challenges

ML-ready datasets

Some useful dataset created with machine learning in mind.

Large ML datahubs

Here are some datahubs which host many ML-ready training sets:

  • Source Cooperative
  • Torchgeo DataModules - torchgeo has a number of datasets readily available to import prepped in a format read for ML. If you are using the TerraTorch, this is particularly useful as it is built on top of TorchGeo so can be used off-the-shelf.

Unlabelled

Labelled

  • PhilEO - 400GB Sentinel-2 dataset of the PhilEO Bench containing labels for the three downstream tasks of building density estimation, road segmentation, and land cover classification.