EXPERT OPINION / PERSPECTIVE — We are currently witnessing a mobilization of technical ambition reminiscent of the Manhattan Project, a realization that data and compute are the new defining elements of national power. I am deeply energized by recent bold moves in Washington, specifically the White House’s launch of the "Genesis Mission" this past November—an initiative designed to federate vast federal scientific datasets for integrated AI training—alongside the real-world deployment of GenAI.mil.
Yet, when I look at the velocity of the commercial sector—from OpenAI launching its dedicated Science division and NVIDIA attempting to simulate the planet with Earth-2, to Google DeepMind aggressively crossing their AI breakthroughs into the geospatial domain—it becomes clear that we are still aiming too low. These projects are not just modeling data; they are attempting to model reality itself. American technical leadership is paramount, but that leadership is meaningless if it is not ruthlessly and immediately applied to our national security framework. We must take these massive, reality-simulating concepts and focus them specifically on the GEOINT mission.
A perfect example of this is that earlier this year, in July 2025, the geospatial world shifted. Google DeepMind released the AlphaEarth Foundations (AEF) model, and through the hard work of the Taylor Geospatial Engine (TGE) and the open-source community, those vector embeddings are now publicly available on Source Cooperative.
From Google
The excitement is justified. AlphaEarth is a leap forward because it offers pixel-level embeddings rather than the standard patch-level approach. It doesn’t just tell you “this 256x256 square contains a city”; it tells you "this specific pixel is part of a building, and it knows its neighbors."
But as I look at this achievement from the perspective of national security, I see something else. I see a proof of concept for a capability that the United States is uniquely positioned to build—and must build—to maintain decision advantage.
Google has the internet’s data. But the intelligence community holds the most diverse, multi-physics, and temporally deep repository of the Earth in human history.
It is time for the United States to propose and execute a National Geospatial-Intelligence Embedding Model (NGEM).
Sign up for the Cyber Initiatives Group Sunday newsletter, delivering expert-level insights on the cyber and tech stories of the day – directly to your inbox. Sign up for the CIG newsletter today.
The Proposal: Beyond RGB
The AlphaEarth model is impressive, but it is limited by its training data—primarily commercial optical imagery. In the national security domain, an optical image is just the tip of the spear. We don't just see with light; we see with physics.
I am proposing that we train a massive, pixel-level foundation model that ingests all of its holdings. We aren't talking about just throwing more Sentinel-2 data at a GPU. We are talking about a model that generates embeddings from a unified ingest of:
- Multi-INT Imagery: Electro-optical (EO), Synthetic Aperture Radar (SAR), Infrared/Thermal, Multispectral, and Hyperspectral.
- Vector Data: The massive stores of Foundation GEOINT (FG)—roads, borders, elevation meshes.
- The Critical Missing Modality: Text. We must embed the millions of intelligence reports, analyst notes, and finished intelligence products ever written.
The Approach: "The Unified Latent Space"
The approach would mirror the AlphaEarth architecture—generating 64-dimensional (or higher) vectors for every coordinate on Earth—but with a massive increase in complexity and utility.
In AlphaEarth, a pixel’s embedding vector encodes "visual similarity." In an NGA NGEM, the embedding would encode phenomenological and semantic truth.
We would train the model to map different modalities into the same "latent space."
- If a SAR image shows a T-72 tank (through radar returns), and an EO image shows a T-72 tank (through visual pixels), and a text report describes a "T-72 tank," they should all map to nearly the same mathematical vector.
- The model becomes the universal translator. It doesn't matter if the input is a paragraph of text or a thermal signature; the output is a standardized mathematical representation of the object.
The Outcomes: What Does This Give Us?
If we achieve this, we move beyond "computer vision" into "machine understanding."
1. The "SAM Site" Dimension In the AlphaEarth analysis, researchers found a "dimension 27" that accidentally specialized in detecting airports. It was a serendipitous discovery of the model's internal logic. If we train NSEM on NGA’s holdings, we won’t just find an airport dimension. We will likely find dimensions that correspond to specific national security targets.
- Dimension 14 might light up only for Surface-to-Air Missile (SAM) sites, regardless of whether they are camouflaged in optical imagery, because the thermal and SAR layers give them away.
- Dimension 42 might track "maritime logistics activity," integrating port vectors with ship signatures.
2. Cross-Modal Search (Text-to-Pixel) Currently, if an analyst wants to find "all airfields with extended runways in the Pacific," they have to rely on tagged metadata or run a specific computer vision classifier. With a multi-modal embedding model, the analyst could simply type a query from a report: "Suspected construction of hardened aircraft shelters near distinct ridge line." Because we embedded the text of millions of past reports alongside the imagery, the model understands the semantic vector of that phrase. It can then scan the entire globe’s pixel embeddings to find the mathematical match—instantly highlighting the location, even if no human has ever tagged it.
3. Vector-Based Change Detection AlphaEarth showed us that subtracting vectors from 2018 and 2024 reveals construction. For the intelligence community, this becomes Automated Indications & Warning (I&W). Because the embeddings are spatially aware and pixel-dense, we can detect subtle shifts in the function of a facility, not just its footprint. A factory that suddenly starts emitting heat (thermal layer) or showing new material stockpiles (hyperspectral layer) will produce a massive shift in its vector embedding, triggering an alert long before a human analyst notices the visual change.
The Cipher Brief brings expert-level context to national and global security stories. It’s never been more important to understand what’s happening in the world. Upgrade your access to exclusive content by becoming a subscriber.
The Intelligence Use Cases
- Automated Order of Battle: Instantly generating dynamic maps of military equipment by querying the embedding space for specific signatures (e.g., "Show me all vectors matching a mobile radar unit").
- Underground Facility Detection: By combining vector terrain data, gravity/magnetic anomaly data, and hyperspectral surface disturbances into a single embedding, the model could "see" what is hidden.
- Pattern of Life Analysis: Since the model is spatiotemporal (like AlphaEarth), it learns the "heartbeat" of a location. Deviations—like a port going silent or a sudden surge in RF activity—become mathematical anomalies that scream for attention.
Conclusion
Google and the open-source community have given us the blueprint with AlphaEarth. They proved that pixel-level, spatiotemporal embeddings are the superior way to model our changing planet.
But the mission requires more than commercial data. It requires the fusion of every sensor and every secret. By building this multi-modal embedding model—fusion at the pixel level—we can stop looking for needles in haystacks and start using a magnet.
This is the future of GEOINT. We have the data. We have the mission. It’s time to build the model.
Follow Mark Munsell on LinkedIn.
The Cipher Brief is committed to publishing a range of perspectives on national security issues submitted by deeply experienced national security professionals.
Opinions expressed are those of the author and do not represent the views or opinions of The Cipher Brief.
Have a perspective to share based on your experience in the national security field? Send it to Editor@thecipherbrief.com for publication consideration.
Read more expert-driven national security insights, perspective and analysis in The Cipher Brief, because national security is everyone’s business.
