Ground Radar Systems
The Silent Sabotage of Bad Data
If you’ve spent any time in this space, you’ve seen it. The radar is solid. The algorithms are state of the art. The model is deployed and running at the edge. But it just won’t work the way you need it to; often even failing before it gets to what warfighter.
You dig in, looking for bugs or calibration drift. But the system is performing as designed. The problem is deeper. It was trained to solve the wrong problem.
What we call “truth” in the data, the labels, the taxonomies, the class definitions, that’s where the mismatch is hiding. And it’s subtle. Not a crash. Not a misfire. Just enough misalignment to seed confusion and erode trust.
That’s what’s failing. Not the signal. Not the model.
Airborne
Ground
Space
/ THE PROBLEM /
The Hidden Flaw in Current AI Training
We’ve built radar AI systems capable of processing massive volumes of signal with speed, clarity, and efficiency at the edge. But when missions change, threats adapt, or real-world conditions diverge from the training scenario, those same systems start to fail. The problem isn’t the model. It’s the data we trained it on.
The real failure is upstream i.e. how we define and structure ground truth.
Too often, we’ve relied on legacy labels and static taxonomies focused on platform type or signal classification. These were designed for controlled environments, not operational ones. They reflect what was convenient to annotate, not what actually matters in the field.
So when a system breaks down, it isn’t because the model made a bad decision. It’s because we trained it to solve the wrong problem. It followed the data. The data just didn’t follow the mission.
/ OUR SOLUTIONS /
Training AI to Think Like a Warfighter
We need to stop thinking like lab assistants and start thinking like warfighters, the operators who use these tools to defend themselves against threats launched from miles away. Their world is dynamic, high-stakes, and unforgiving. Ours needs to meet that standard.
Every labeled radar return encodes a set of assumptions—what the object is, what it’s doing, whether it matters, and what action to take. When those assumptions are crafted for analysis instead of operations, the system fails to perform for the end user and only adds burden.
This is not a model problem; it is a failure in how we define the data itself.
Fixing it means changing how we think about ground truth:
- We design taxonomies that follow operator reasoning, not spreadsheet structure. Labels must reflect how decisions get made under pressure, not how assets are categorized in documentation.
- We capture behavior over time, not just frame-by-frame classifications. Models need motion profiles, revisit consistency, and pattern recognition to match how operators assess threat.
- We anchor everything to mission relevance. Identity is secondary; context determines importance. A benign object can become critical in seconds. Labels must reflect that shift.
This isn’t about collecting more data, annotating more returns, or optimizing a pipeline. It’s about designing data that teaches the model how to think like the people who depend on it.
/ TECHNICAL DEEPDIVE /
Beyond the Metrics: What Your Radar AI is Missing
Radar gives more than you’re using
Radar isn’t just a detection system; it’s a behavioral sensor. You get range, velocity, azimuth, cross-section, revisit rate, and in some cases polarization or micro-Doppler signatures. These signals carry intent, and other useful insights that can be leveraged.
But most models ignore the majority of that richness. Why? Because the data wasn’t structured to take advantage of it. If the labels reduce everything to “vehicle,” “object,” or “clutter,” that’s all the model learns to care about.
The output matches the label logic, not the operational need. If your taxonomy doesn’t differentiate between a loitering contact and a pass-through track, the model won’t either. You left the real signal on the cutting room floor.
Your metrics are lying to you
Precision, recall, and F1-score sound scientific. But they reflect performance on held-out validation sets drawn from the same flawed data. They reward repetition, not adaptation.
A model can ace the test set and still fail in the field. Why? Because the test set is clean, labeled under lab conditions, and built for statistical certainty—not for ambiguity, spoofing, or changing ROEs.
What matters is whether the model produces outputs that help an operator decide, act, or respond faster and more accurately. If the data doesn’t reflect operational ambiguity, your metrics are just theater.
Fixed taxonomies don’t survive contact
Most radar datasets are built on rigid class lists. They’re organized by asset type, signal family, or commercial category. That works in PowerPoint. It doesn’t work when adversaries change appearance, spoof known profiles, or blend into civilian patterns.
A label like “military truck” is meaningless if it includes both friendly and hostile signatures. If your model isn’t trained to detect threat behavior—only object identity—it will fail when identity stops correlating with intent.
Taxonomies must evolve with tactics. If your data can’t adapt, neither will your system.
There is no ground truth in the field
We like to talk about “ground truth” as if it’s a fixed reference. In radar, it isn’t. Truth shifts with time, context, and mission phase. A contact that was irrelevant at 0800 might be critical by 0810.
Signals are noisy. Tracks are discontinuous. Correlation is messy. If your labels treat each return as an isolated fact—with no temporal or situational encoding—you’ve already lost the thread.
Truth in this domain is a hypothesis, not a fact. Your data needs to reflect that. If it doesn’t, your model is solving a version of the world that doesn’t exist.
/ CONCLUSION /
What We Do
At Deca, we start with the assumption that the data is the problem. And we work backwards from there.
Before we build or deploy a model, we audit the label set. Not just for accuracy, but for relevance. Does it capture behavior? Does it include context? Does it reflect what operators actually need to know?
We work with SMEs to rebuild taxonomies around mission logic, not just object type. We emphasize track-based annotation and time-windowed labeling. We look for patterns the model might need to learn and make sure those patterns are represented in the truth set.
And after deployment, we treat model outputs as part of the feedback loop. If a system misclassifies a high-priority contact, we don’t just retrain. We trace it back to the labeling decision that shaped the error.
This is not flashy AI. It is the kind of work that keeps models trustworthy when it counts.
If your radar AI systems are underperforming, don’t start with the model. Start with the data. Ask whether your labels reflect the environment you’re operating in. Ask whether your taxonomy matches the way your teams think. Ask whether your model is learning what you want it to learn or just memorizing what was easy to label.
At Deca, we help teams align data with mission. Not in theory, but in the real world, where ambiguity is the norm and decisions can’t wait for perfect inputs.