scieee AI-readable full text Open interactive document viewer

Geofoundation Models: Unlocking the Power of GeoAI

de Jong, Jakko

Publisher
Zenodo
DOI
10.5281/zenodo.17586979
Repository
zenodo.org
Language
en

Abstract

Geospatial foundation models (geo-FMs) are rapidly transforming the way we analyze and understand our planet. By leveraging vast amounts of satellite and geospatial data, these models provide a powerful foundation for tackling complex spatial challenges with unprecedented scale and efficiency. Similar to the role of large language models in natural language processing, geo-FMs serve as versatile starting points that can be fine-tuned for diverse applications, from environmental monitoring to climate adaptation. They significantly reduce the need for labeled data, improve the consistency of analysis over time, and open new opportunities for collaboration between science, government, and industry. Spheer has developed Carto, a web-based Earth Observation engine powered by an in-house geospatial foundation model trained on time series of Sentinel-2 satellite data. With Carto, researchers, governments, land managers, and companies can develop custom monitoring solutions to analyze large areas and gain insights into biodiversity, agriculture, climate adaptation, infrastructure, and other spatial challenges. In several use cases, such as detecting specific grasses on Dutch heathlands indicative of nitrogen deposition, Spheer's geo-FM has demonstrated improved temporal stability and a hundredfold reduction in required labeled data. Since September 2024, their small-data machine learning approach has been validated across dozens of use cases, both in the continental Netherlands and the Dutch Caribbean islands. This talk Jakko de Jong, Director of Spheer, explains how Carto works and share lessons learned while developing their geofoundation model. He also showcases the existing use cases and discuss potential collaboration opportunities with researchers.

Full text

Unlock the power of Earth Observation for everyone Jakko de Jong, PhD. CEO and co-founder UTwente, 05-nov-2025 Jan Buwalda Fotografie. Why Geo-AI? Why geo-AI? GeoML and (Geospatial) Foundation Models Results SpheerFM Carto: tech + demo + usecases What is next? Our datasets Earth Observation is central to solving urgent global problems Food Security Water quality Tourism pressures Biodiversy loss Climate adaptation Physical models versus Machine learning Physical models •Patterns determined by human experts •Requires deep understanding of physics involved •Explainable •Use-case specific. Not easily scalable •Static Machine learning •Patterns determined by statistical learning model •Requires (lots of) high quality examples •Less explainable •Generic approach. Scalable to many usecases. •Active, can learn via user interaction Is there a long tail Earth Observation demand? Long tail economy: selling many unique items with relatively small quantities sold of each. Source: wikipedia EO-derived products and services € available for product Creating satellite monitoring is cumbersome Requires teams of specialized dataand modeling experts and weeks to months of development time Source: EUSPA market report 2024 We believe that GeoAI will: 1. Make the ‘head’ more efficient. 2. Open up the tail for custom satellite monitoring. Spatial advisor on-demand that complements human experts Requires: 1. Artificial intelligence that understands the physical world (like GPT understands the textual world) 2. Intuitive and fast interaction between human and AI (at ~human thinking speed) Our datasets Why geo-AI? GeoML and (Geospatial) Foundation Models Results SpheerFM Carto: tech + demo + usecases What is next? Our datasets Satellite timeseries: worth the effort Sentinel-2 timeseries of Dutch salt marshes Satellite-AI can analyze the earth But development is still slow and expensive …. Dunal dynamics Molinia Invasive species Our Geo foundation models are case-agnostic: we have taught them to understand earth’s surface. Users only train the last step – teaching the model what they specifically want to monitor, with only 1% of label data that conventional AI requires. Our IP: Geospatial Foundation Models that make geo-AI fast and interactive Insights Earth Observation data User-AI interaction Finetuning: •Very simple model •100x fewer observations •100x faster: ~min instead of hours Train GFM: •Deep model •No label data •Unlimited satellite data •FM trained once: lower CO2 footprint Self-supervised learning Large Language Models (LLMs) are trained on reconstruction tasks cats rule the world, they are way better than dogs. Self-supervised learning on images 1. Original image 2. Masked image 4. Reconstructed image AI Model 3. Model 5. Compare and update model Geospatial Foundation Models are hot •Presto (2021) •Clay (2023) •SpheerFM (march 2024) •IBM: TerraMind (2025) •Google: AlphaEarth (2025) •Cambridge: Tessera (2025) SpheerFM was trained on nature areas in NL… …on satellite video’s of a calendar year 2017 2024 2017 2018 2024 Geospatial Foundation Model ‘Embedding’: Full year summary of a pixel, represented abstractly in a vector of 128 numbers. Embeddings 2017 Embeddings 2024 Embeddings summarize the earth surface Aerial photo of Dutch salt marsh Embeddings created with SpheerFM Comparison with Google AlphaEarth Comparison with TESSERA Carto demo Why geo-AI? GeoML and (Geospatial) Foundation Models Results SpheerFM Carto: tech + demo + usecases What is next? Our datasets Carto lets users create custom satellite monitoring in minutes In the browser No AI or remote-sensing expertise needed Train a model in ~1min, iterate to accurate maps and trends in 30 minutes Needs 100x fewer expensive examples to learn March 2024: First succesful GFM Sept 2024: Carto prototype online April 2025: Launch of Carto version 1.0 Carto tech + demo •Platform is hosted on AWS •Front-end is written in React with MapLibre for the viewer •Backend written in Python: fastAPI •User models are simple and robust scikitlearn models •Training and Inferencing on a Dask cluster •Maps served via a TileServer in AWS Lambda If you want more details, I will bring you in contact with Mark (CTO) Carto use-cases Why geo-AI? GeoML and (Geospatial) Foundation Models Results SpheerFM Carto: tech + demo + usecases What is next? Our datasets Structure map of Rhine basin Structure map of dunes Forest monitoring and checking nature subsidees Search for Habitattypes inside and outside N2000 Groenknolorchis is een kenmerkende soort voor: •Vochtige duinvalleien (H2190), •Trilvenen (H7140) en •Alkalische laagveenmoerassen (H230) Quality checks of human-made vegetation maps Mangroves and Coralita on Caribbean islands Mangroves on Bonaire Source: Carto