Code and notebooks for the ISPRS Congress 2026 tutorial "Towards Geospatial Embeddings: Investigating Accurate and Accessible Deep Geospatial Feature Representations" (tutorial info).

The tutorial pairs lectures with two hands-on coding sessions. Both notebooks are designed to run start-to-finish on a free Google Colab CPU runtime — no GPU, and no Earth Engine account required.

📑 Lecture slides: slides/ISPRS_Tutorial_EE.pdf.

	Notebook	What you do	Open in Colab
Demo 1	demo1_using_earth_embeddings.ipynb	Use pre-made embeddings for prediction: coarse global SatCLIP location embeddings → ecoregion/biome; fine-grained AlphaEarth pixel embeddings → Canadian crop-type mapping. Includes a satellite-imagery baseline and geographic transfer — hold out whole continents (a configurable leave-one-continent-out sweep) and map where out-of-domain predictions fail.
Demo 2	demo2_producing_earth_embeddings.ipynb	Produce your own embeddings with the training-free MOSAIKS random convolutional features, see how spectral bands and image size change downstream accuracy, then compare against a pretrained SSL4EO-S12 foundation model loaded from the Hub. Ends with a similarity ("find places like this") map.

To keep the live sessions fast and dependency-free, each notebook downloads small, pre-packaged datasets from a public Hugging Face dataset repo (kklmmr/isprs26-earth-embeddings) — no login, no Earth Engine. The scripts that build those files from the original open sources live in scripts/ — see scripts/README.md to regenerate them.

notebooks/   the two demo notebooks (the things participants run)
slides/      lecture slides (PDF)
scripts/     offline data-prep (organizers only; some steps need a GEE project)
tutorial/    small reference library (RCF featurizer) used by tests/prep
tests/       unit tests for the reusable code
local/       staged data before upload to Hugging Face (git-ignored)

• SatCLIP location embeddings — Klemmer et al., Microsoft (repo).
• AlphaEarth Foundations "Satellite Embedding" — produced by Google and Google DeepMind; CC-BY 4.0 (Earth Engine catalog).
• RESOLVE Ecoregions 2017 — CC-BY 4.0 (ecoregions.world).
• AAFC Annual Crop Inventory — Agriculture and Agri-Food Canada, Open Government Licence – Canada.
• Sentinel-2 — Copernicus / ESA.
• EuroSAT — Helber et al., 2019 (Sentinel-2 land-cover patches).
• SSL4EO-S12 pretrained ResNet-18 — Wang et al., 2022; weights pulled from the public HF repo torchgeo/resnet18_sentinel2_all_moco (via TorchGeo).

Academics can get premium access to LGND's Embeddings API for free — learn more and apply via the research tier.

• Klemmer, Konstantin, et al. "Earth Embeddings: Towards AI-centric Representations of our Planet." IEEE GRSM (2026). [EarthArXiv]
• Fang, Heng, et al. "Earth Embeddings as Products: Taxonomy, Ecosystem, and Standardized Access." arXiv (2026). [arXiv]
• Rolf, Esther, et al. "Mission Critical — Satellite Data is a Distinct Modality in Machine Learning." ICML (2024). [arXiv]
• Corley, Isaac, et al. "No One Knows the State of the Art in Geospatial Foundation Models." arXiv (2026). [arXiv]
• Betti, Livia, et al. "What's in an Earth Embedding? An Explainability Analysis of Location Encoders." arXiv (2026). [arXiv]
• Kaur, Amandeep, et al. "Pretrain Where? Investigating How Pretraining Data Diversity Impacts Geospatial Foundation Model Performance." CVPR (2026). [arXiv]
• van der Plas, Thijs L., et al. "Better Together: Evaluating the Complementarity of Earth Embedding Models." arXiv (2026). [arXiv]
• Gilch, Luis, et al. "How to Embed Matters: Evaluation of EO Embedding Design Choices." CVPR (2026). [arXiv]
• Vinge, Rikard, et al. "NeuCo-Bench: A Novel Benchmark Framework for Neural Embeddings in Earth Observation." CVPR (2026). [arXiv]
• Corley, Isaac, et al. "From Pixels to Patches: Pooling Strategies for Earth Embeddings." arXiv (2026). [arXiv]

• Bad Tables: Why You Shouldn't Trust Results Tables in RS Foundation Model Papers — Anthony Fuller
• 2025 LIDS Seminar — Konstantin Klemmer (Microsoft Research) — MIT LIDS
• Earth Embeddings: Learning Mental Maps in Neural Nets — AI + Environment Summit 2025

• TorchGeo — join the community on Slack.

Code released under the MIT License (see LICENSE). Data subject to the licenses above.