🛰️🌍 GeoFMs in 5 Minutes: From Earth Observations to Embeddings

If you’ve heard about foundation models mainly in the context of text (LLMs), the same idea is now showing up for Earth observation. Instead of learning from words, Geo Foundation Models (GeoFMs) learn from satellite imagery + weather + other geospatial context — and produce embeddings you can reuse across many tasks.

This post is a short, semi-technical walkthrough you can read (or present) in ~5 minutes.

What is a GeoFM? 🛰️

A GeoFM is a large neural network pretrained on massive amounts of geospatial data, such as:

• 🛰️ Sentinel‑2 (optical reflectance)
• 🛰️ Sentinel‑1 (SAR radar backscatter)
• 🌦️ ERA5 (reanalysis weather signals)
• 🌿 NDVI and other vegetation proxies
• ⛰️ Topography (e.g., SRTM elevation)
• 🗺️ Land cover context (e.g., Dynamic World)

Like the way “ImageNet-pretrained” models can serve as general-purpose feature extractors for images, a GeoFM can serve as a feature extractor for Earth observation time series.

What is an embedding? 🧠

An embedding is a compact numeric representation (a vector like ([x_1, x2, \dots, x{128}])) that captures what the model “understands” about an input.

In practice, embeddings act like coordinates in a learned feature space:

• Similar places / patterns end up close to each other
• Unusual behavior ends up far away (useful for anomaly detection)

In many GeoFM pipelines (including the one described here), you don’t fine-tune the model. You run a forward pass, store the embeddings, and build lightweight downstream logic on top.

The simple flow 🔁

At a high level, a GeoFM pipeline often looks like this:

Ingest (EO + weather)  →  GeoFM encoder (frozen)  →  Embeddings  →  Many tasks

What “ingest” typically means

You build a consistent input tensor per location (e.g., a field polygon) and per time window (e.g., 12 monthly “slots”), aligned across modalities:

• optical + radar sequences
• weather sequences
• static context (elevation, land cover, position)

What the GeoFM does

The GeoFM processes the time series and emits a vector embedding. Many GeoFMs use Transformer-style encoders over sequences (just like text models do — only the tokens are sensor/time features rather than words).

What happens after embeddings

Once you have embeddings on disk, everything downstream can often be cheap:

• similarity search (“find places like this”)
• anomaly detection (“this is unusual for this season/region”)
• clustering (“group similar trajectories”)
• small supervised models (train a classifier/regressor on embeddings)

GeoFMs vs LLMs 🤖 (a simple comparison)

What’s similar

• Both are usually built on Transformer-style architectures
• Both learn general-purpose representations via large-scale pretraining
• Both can be reused for many downstream tasks

What’s different

• LLMs are optimized for text generation: tokens in → tokens out
• GeoFMs are commonly used as feature extractors: geo signals in → vectors out

An easy mental model:

LLMs are often “generators.” GeoFMs are often “encoders” you use to turn messy multi-sensor data into clean vectors.

What can you do with GeoFM embeddings? 🧭

Here are common, practical patterns:

• 🔎 Similarity search: find regions/fields that behave like a reference example.
• 🚨 Anomaly detection: detect unusual vegetation stress, flooding, disruption, or change.
• 🧩 Clustering: segment a landscape by phenology or land-use behavior.
• 🎯 Lightweight supervised learning: add a small model (logistic regression, XGBoost, shallow MLP) on embeddings for a specific label.
• ⏱️ Monitoring: compute embeddings once, then score continuously as new observations arrive.

Where this helps (examples) 🌾🏗️

🌾 Agriculture

• early drought/stress signals
• yield risk monitoring
• crop type stability / rotations
• targeted scouting and prioritization

🚆 Mobility & infrastructure

• monitoring roads/rail corridors for flood/landslide impacts
• detecting construction, degradation, or disruption
• prioritizing inspection after extreme events

🧾 Insurance

• rapid triage after events (hail, drought, flood)
• portfolio-level risk monitoring
• parametric triggers with explainable signals

⚡ Energy

• vegetation encroachment risk (transmission corridors)
• solar farm soiling / performance context
• siting support with long-term environmental signals

🌍 Climate & environment

• land-cover change detection
• water stress / drought monitoring
• ecosystem degradation signals

Closing thought

The real win of GeoFMs is not a single “magic prediction.” It’s the ability to turn multi-sensor, messy geospatial time series into a stable embedding space — so you can build many reliable products on top with simpler downstream logic.