The DoD's UAS group classification is a size-and-speed taxonomy: Group 1 is small (under 20 lbs) and slow, Group 2 is medium (21-55 lbs) with slightly higher altitude and speed envelopes, Group 3 is larger and faster still. It's a useful classification for procurement. It's also, quietly, a boundary that most CUAS training datasets fail to respect.
The failure looks like this: a program trains a detector on a mixed bag of drone footage and reports strong aggregate performance. The detector goes to the field. It classifies Group 1 quadcopters at ranges where they're 8 pixels wide, and it classifies Group 2 fixed-wing platforms at ranges where they're 30 pixels wide. Different resolutions, different aspect ratios, different motion signatures, different backgrounds because the operating envelopes differ. The detector was learning three different problems and reporting the average.
What Actually Changes Between Classes
Aspect ratio changes. A Group 1 quadcopter is roughly square from most angles. A Group 2 fixed-wing is a horizontal streak, sometimes with visible wing planform, sometimes just a fuselage silhouette depending on aspect. A detector trained on one class's aspect distribution generalizes poorly to the other.
Motion changes. Quadcopters hover, translate, and can pivot in place. Fixed-wing platforms cannot hover; they orbit or transit. The temporal signature — velocity vectors, acceleration patterns, trajectory smoothness — differs enough that tracking algorithms tuned to one class will drop the other.
Pixels-on-target changes for the same detection range. A Group 2 platform at 800 meters presents more pixels than a Group 1 at the same range because it's physically larger. The detection threshold for one is not the detection threshold for the other, and a training set that treats them uniformly will optimize for the wrong tradeoff.
The Implication For Collection
Every serious CUAS program should specify target class explicitly in its data plan and its collection contract. "Give us drone footage" is not a specification. "Give us Group 1 quadcopter footage at 300-800 meters slant range across five environment types" is a specification.
The vendor should ship data tagged by target class and hold the class balance the program specified. Cross-class validation should happen before delivery, not after the model underperforms in field trials.
Aggregate detection performance across mixed target classes is a vanity metric. The metric that matters is per-class performance at operational range.
The fix isn't exotic. It's discipline about what's being collected and honesty about what a detector trained on Class A can be expected to do against Class B. Which is: not much, and program managers who assume otherwise are the ones getting paged when the field results come back.