Radar target classifiers get more accurate when you tell them the angle they're viewing a target from, not just what its echo looks like.
Researchers tested a technique called High-Resolution Range Profile classification, which identifies objects by analyzing the shape of their radar echo, across three datasets and a wide range of model architectures. Most prior work assumed the aspect angle - the angle between the radar and the target - was unknown or only partly known during training. This study instead fed the angle directly into the classifier for every training example, and the result was a consistent accuracy gain of about 7 percent on average, and as much as 10 percent depending on the model and dataset, for both single-snapshot and sequential classifiers. Since real radar systems do not simply hand a classifier the true aspect angle, the researchers also tested a causal Kalman filter that estimates the angle in real time from the same data stream, and it did so with a median error of about 5 degrees.
That estimation step is the real finding here. A 7 percent accuracy bump on a lab benchmark is nice, but it only matters if a fielded system can get that angle information without extra sensors. The Kalman filter estimates preserved most of the accuracy gains, which means the approach does not depend on some idealized, perfectly known angle that only exists in a training set.
None of this has been tested against a target actively trying to confuse the radar, so treat the 10 percent figure as a lab result, not a battlefield guarantee.