Intro to Earth Foundation Models
EFMs WTF?
The landscape for Earth Foundation Models (EFMs) has been confusing to me… every week, if not every day, I see a post about a newly released EFM that is groundbreaking, game-changing, and disruptive to the earth observation ecosystem.
Um. Maybe! But I haven’t been able to understand any of these in context. Each post is advertising a single model, and the braggadocious superlatives for them are getting tired.
This is the first post in a series that is my attempt to start making sense of this (perhaps: disruptive! ground-breaking!) emerging aspect of the EO/RS industry.
I have a few follow-on topics queued up, but please let me know what comments, questions, and curiosities you have. I’d love to help unmask these masked autoencoders (sorry) and the like for those who are interested.
Two distinct disciplines
Let’s get one thing straight right out of the gate: the term EFM is being used for two distinct domains: weather/climate forecasting and land observation analysis. While these two types of models share some underlying conceptual frameworks, they work differently, require different data, and solve different problems
Climate models: built for prediction
Weather/Climate models (WxC) predict future atmospheric conditions and are used for weather forecasting and climate projections. They are built on decades of dense meteorological observations and the models have an understanding of how conditions change over time
In weather forecasting, a process called data assimilation integrates real world observations with physics-based models of the world in order to encode atmospheric dynamics & temporal relationships into a single data product. This assimilated data is what WxC EFMs get trained on, so the models are provided with a temporal framework from the start.
Land observation models: built for pattern recognition
Land Observation Models analyze what's happening on the earth's surface right now. They’re good for tasks like monitoring crops, detecting deforestation, and identifying the extents of urban areas. They're better at spotting patterns in optical satellite data than they are at predicting what happens next.
As a result, many land observation EFMs were initially designed to have a “single-snapshot” understanding of the world. This means that while they may take in data from different points in time, they don’t actually understand change over time in the same physics-informed way that weather models do. This paradigm is shifting, however, as a few natively multi-temporal models have arrived and others are actively working to catch up on out-of-the-box multi-temporal capabilities.
A shared conceptual framework
While WxC and land observation models have important differences, we can treat them as part of the same family at a high level.
At their core, all of these EFMs are deep learning models trained to contain a mathematical representation of the earth. Or, more accurately, a portion of the earth’s systems or subsystems, such as the land surface, the oceans, and/or the atmosphere. These representations are built from various data inputs, some more sophisticated or diverse than others.
Removing the need to build from scratch
Once trained, EFMs serve as the "backbone" for fine-tuned earth observation modeling. Think of them as brains containing an existing earth model, but spongy enough to absorb new intel. The mathematical foundation eliminates the need to train a bespoke model from scratch, saving practitioners hours of data acquisition, labeling, and model design.
As an example, a user could add a few hundred labeled wildfire burn scar images to a pretrained EFM and achieve accurate burn scar detection. Training a specialized model from scratch, on the other hand, would require hundreds of thousands of labeled images for the same accuracy.
We still have an accessibility problem
Like customizing a GPT on OpenAI's ChatGPT, you can layer analytical “instructions” onto an EFM's pre-trained earth model. Unlike consumer LLMs, however, accessible EFM interfaces don't exist yet. This technology is still gate-kept for the data science and ML elite, typically published as GitHub repos and HuggingFace resources containing model architecture, massive pre-curated datasets, training instructions, and fine-tuning docs.
Great for a certain audience. But, we still need to bridge the gap to consumer and business users.
What makes each model different
As you can start to see, there are a handful of characteristics that differentiate models from one another, even when they share a general discipline of WxC or land observation. Getting a handle on the inputs, outputs, and accessibility of a model helps us gauge what that model might be useful for.
What goes in
Understanding the input training data helps you know what each model might be good at.
Single vs multi-modal: Some models only work with one type of data, like optical satellite imagery. Others can combine different data types - SAR, optical imagery, elevation maps, even text descriptions.
Data specifics: Does this model need clean, harmonized Landsat data, or can it handle messier sources like drone imagery? What spatial, spectral and temporal resolutions can it handle as inputs? Are the suggested pre-training datasets openly licensed, or do I need to make a data purchase to get this to run optimally?
What comes out
The output data format determines what you can actually do with the model.
Granularity: Are you getting annual trends or daily insights? Is information packaged in 10m x 10m chunks, or is this a pixel-level analysis based on my input data resolution?
Format: Some models produce embeddings you can query semantically. Others give you traditional raster outputs, raw tensors or weather variables.
How do I use it
Here we start to ask deeper questions about practical deployment of these models.
Paper/code/platform availability: Is this model documented solely in a research paper? Is there implementable code provided? How about pre-trained weights and deployment documentation?
Licensing: If I invest in learning to use this tool, can I benefit commercially or is this only for research and academic purposes.
Computational requirements: Does this need high-end GPUs, or can it run on a standard cloud instance? What are the memory and processing requirements for inference?
And, of course, what does this actually solve
All of these questions and characteristics are driving at answering the overall question of “what does this actually solve for me?”. That means being able to calculate an ROI based on understanding the input effort, time, and money as well as the output gains. Particularly since these tools are not consumer-friendly, we want to know that our effortful adoption is going to be worth it!
Some Models Worth Knowing About
As a way to help folks find their footing in this growing EFM landscape, the majority of the work I've put in over the last few weeks has resulted in a distilled spreadsheet (sorry, guilty product person). The tables organize the key characteristics of the most popular and best performing earth foundation models across both the climate and land observation domains. I worked to address questions of data inputs, outputs, accessibility, and also gave my best shot at distilling each of these models into a snippet about real world use cases and a short paragraph descriptor with slightly more detail.
There are, of course, a ton more models out there that I didn't include. This initial view focuses heavily on projects that are "developer friendly", meaning there are more resources for the model than just a published academic paper. I also shied away from commercial SaaS platforms like SunAirIo, Jupiter Intelligence, and Zeus AI for now.
Here are the top open models that you or your ML-savvy colleagues can start experimenting with:
Note: This landscape is changing very quickly! The above information is my best understanding of the current characteristics of these models as of the time of publishing this article (August 27, 2025). If you have conflicting information, or are a developer of or expert on any of these models and would like to share your expertise with me, please do reach out!
What's Next
As a follow on, I’d like to dive deeper into the question of current accessibility for these EFMs. What does it mean to "run the model from scratch" versus using pre-computed embeddings? How do embedding-based models differ from predictive ones in practice? What kind of technical resources do you actually need to get value from these tools?
The gap between "here's a cool GitHub repo" and "here's a tool that solves my business problem" is still wide in the EFM ecosystem. Technical maturity will need to fill that gap in the long run. But, in the short run, I’m hoping to help bridge it with further education on the effort it takes to get up and running with these emerging tools. As mentioned at the outset, I have a few things I’d like to work on next, but please please let me know what questions, comments and other feedback you have!