AI Space Surveillance: Why Maui’s Telescopes Matter

Forty thousand.

That’s roughly how many objects the U.S. tracks in orbit today—satellites, spent rocket bodies, and debris fragments that can turn a routine launch into a high-speed collision problem. Now add a harder layer: adversaries that don’t just operate in space, but actively try to avoid being seen.

That’s the point Gen. Chance Saltzman, the U.S. Space Force’s chief of space operations, made during a visit to the Maui Space Surveillance Complex. China is experimenting with ways to reduce observability—changing apparent brightness, maneuvering in perceived blind spots, and using operational timing to complicate tracking. If you care about defense and national security, this isn’t trivia. It’s the practical difference between knowing what’s happening and getting surprised.

This post is part of our “AI in Defense & National Security” series, and Maui is a perfect case study. Not because it’s just a set of telescopes on a volcano, but because it illustrates the real future of space domain awareness: sensors produce the signal; AI produces the understanding.

Space domain awareness is shifting from “where” to “why”

Space domain awareness (SDA) is no longer a cataloging task. It’s an intelligence problem.

In the early years, the mission was straightforward: prevent collisions, maintain safe operations, and maintain basic custody of objects. Today, SDA also demands you answer questions that sound like classic intelligence analysis:

• Who is operating the object?
• What capability does it imply?
• What behavior is normal for that operator?
• What behavior suggests intent—surveillance, interference, staging, or rehearsal?

Saltzman has described a “theory of success” with components such as avoiding operational surprise and confronting malign activity. That theory depends on SDA. You can’t prevent surprise if you’re only seeing 10% of an orbit at a time.

Here’s the thing I’ve found when talking with teams building mission analytics: the technical bottleneck is rarely “we can’t collect data.” It’s we can’t convert data into decisions fast enough—especially when the other side is trying to make your data messy.

Maui’s advantage: geography buys you time—and data quality

Maui isn’t “special” because it’s remote. It’s special because it’s predictably clear.

Perched on Haleakalā at about 10,000 feet, the complex sits above much of the weather. That matters for optical surveillance, where atmospheric distortion and cloud cover can erase your window. The Space Force officers on-site emphasize that this is among the best locations in the world for telescope performance, including daytime viewing.

Why optical sensors still matter in an AI-first defense stack

A common misconception is that optical telescopes are old-school compared to radar. In reality, they’re complementary, and optical has unique strengths:

• High-resolution characterization: seeing shapes, solar panel configurations, and relative orientation
• Passive collection: optical systems don’t advertise themselves the way active emitters can
• Deep-space utility: observing objects in geosynchronous orbit (GEO) and beyond

In practice, optical sensors provide “identity clues” that feed a modern fusion pipeline. And fusion is where AI earns its keep.

The “orbital hide-and-seek” problem is a data-association problem

If adversaries maneuver, the hard part isn’t noticing a change. The hard part is maintaining custody.

At Maui, the Advanced Electro-Optical System (AEOS) telescope can track low-Earth orbit (LEO) satellites moving fast across the sky—often visible for only five to nine minutes per pass. That’s enough to observe, but not enough to guarantee you know what happens next.

Saltzman noted a key constraint: a sensor might only see a fraction of an object’s orbit. If an object performs an unexpected maneuver after leaving the field of view, your prediction model may be wrong when it reappears.

That challenge has a name in analytics: data association. You’re trying to decide whether the object you see now is:

1. The same object you saw earlier (with a maneuver), or
2. A different object with a similar signature, or
3. A deceptive tactic designed to create ambiguity

Where AI fits (and where it doesn’t)

AI can materially improve orbital surveillance—but only if it’s used for the right jobs.

AI is strong at:

• Multi-sensor fusion: combining optical, radar, telemetry, and commercial tracking data
• Anomaly detection: flagging behavior outside established patterns (operator-by-operator)
• Classification: identifying likely satellite types via signatures and imagery
• Trajectory prediction under uncertainty: learning maneuver patterns and operator “habits”

AI is not a substitute for:

• High-quality sensor calibration
• Physics-based orbit determination
• Credible analytic tradecraft (especially for intent assessments)

The winning approach is hybrid: physics + AI + analyst judgment, with auditability built in.

Upgrades on Maui show the real race: algorithms and processing

Hardware matters, but the faster race is in software-defined performance.

Maui hosts multiple telescopes, including Ground-Based Electro-Optical Deep Space Surveillance System (GEODSS) instruments. These are receiving Ground-Based Electro-Optical Sensor System upgrades, modernizing sensors, optics, algorithms, and post-processing to detect “smaller, dimmer things further” out.

This is exactly where AI in defense becomes concrete. Post-processing is where you:

• Denoise and enhance imagery
• Correct atmospheric distortion
• Identify features consistently across lighting changes
• Generate a usable track from imperfect observations

A small improvement in detection threshold can have outsized strategic impact. If your system can reliably characterize dim objects and maintain custody through maneuvers, an adversary has fewer safe windows to test capabilities without attribution.

If you can’t track it, you can’t deter it.

That’s not a slogan. It’s a systems reality.

“Intent” is the hardest deliverable—and the most valuable

Knowing what an object is doing is good. Knowing why it’s doing it is priceless.

Col. Barry Croker described the modern requirement clearly: track not only what is where, but who is where—and what their intent is. This is the analytic jump from surveillance to intelligence.

Consider a capability that can rendezvous with another satellite and move it to a different orbit. That could be:

• A legitimate debris-removal experiment
• A servicing demonstration
• A rehearsed co-orbital anti-satellite technique

Intent assessment depends on context:

• Where and when it’s tested
• The proximity operations profile
• Prior behaviors by that operator
• Correlation with geopolitical events

A practical AI pattern for SDA intent assessment

If you’re building or buying AI for space domain awareness, don’t ask for “AI that predicts intent.” Ask for systems that deliver these three outputs reliably:

1. Behavior baselines: normal maneuver patterns per operator and per satellite class
2. Change detection: what changed, by how much, and how often
3. Explanation artifacts: the features that drove the alert (for analyst trust)

When those three exist, analysts can do the final step—judgment—quickly and defensibly.

Space surveillance is also a governance problem (and Maui proves it)

A sensor advantage isn’t sustainable if the community rejects it.

The Maui complex sits on a site that many Native Hawaiians consider sacred. Proposals for additional telescopes have faced protests, and a 2023 fuel leak intensified public anger. The Space Force has described a multi-year cleanup process designed to avoid removing soil from the site.

From a national security perspective, it’s tempting to treat this as “non-technical noise.” That’s a mistake. The access, permitting, and legitimacy of key sensor sites are part of the architecture.

A mature SDA strategy has to include:

• Transparent environmental stewardship (measurable milestones, not vague promises)
• Community partnership (not just public affairs)
• Resilience through diversification (don’t depend on a single mountaintop)

If you want resilient architectures, you need resilient relationships.

What defense leaders should do next (actionable, not theoretical)

AI-enabled surveillance and intelligence analysis only create an advantage if the program is engineered for operations.

Here are five practical moves that separate “tech demos” from mission capability:

1. Treat custody as a performance metric. Track how often you lose and regain objects, the latency to re-establish custody, and how that changes after upgrades.
2. Fund the processing pipeline like it’s a weapon system. Sensors without scalable post-processing are just expensive cameras.
3. Require explainability for alerts. If an anomaly model can’t show why it fired, analysts will ignore it—or overreact.
4. Build operator-specific models. “One model for all satellites” is convenient and usually wrong. Adversaries win in the edge cases.
5. Plan for contested data. Assume deception: brightness changes, spoofed signatures, timing games, and orbital crowding. Design fusion logic accordingly.

This is where leads are actually created in defense tech: not by promising a miracle, but by offering a credible path to faster warning, better attribution, and fewer surprises.

The future is sensor networks plus AI—built for attribution

Maui’s telescopes matter because they highlight a blunt truth: space is no longer a quiet utility layer beneath modern warfare. It’s a place where rivals probe, posture, and practice.

Optical surveillance from high-performance sites helps the U.S. maintain an edge, but the edge won’t come from the telescope alone. It comes from connecting telescopes, radar, commercial data, and operator intelligence into an AI-enabled space domain awareness system that can keep custody through deception and produce analyst-ready conclusions.

If your organization is modernizing ISR, building AI-driven intelligence analysis capabilities, or trying to operationalize anomaly detection for space surveillance, now is the time to pressure-test your architecture. Are you building for pretty pictures—or for attribution and response?

Where do you think the next advantage will come from: better sensors, better AI models, or better integration between operators and analysts?