Look at almost any CUAS training dataset and you'll find the same distributional bias: daylight, clear weather, high-visibility conditions. The reason is obvious — that's when it's easiest to collect. The consequence is also obvious, once you think about it. Half the operating envelope has no training data, and the adversary knows.

Drone incursions do not stop at sunset. In some theaters they concentrate at night, specifically because defensive detection posture degrades when the light does.

What Night And Thermal Actually Look Like To A Detector

An EO camera at night doesn't see a small drone at 500 meters. There is nothing to see. The drone is unlit, the background is dark, the contrast is gone. A detector trained on daylight EO data has no useful signal to work with.

A thermal camera at night sees the drone's heat signature — motor bells, battery pack, ESC assemblies. But that signature looks nothing like the visual signature the detector was trained on. Different silhouette shape, different aspect, different background statistics. A detector that classifies daylight EO can be trained on thermal, but it needs to have seen thermal in training. Zero-shot transfer from EO to thermal is not a thing.

Low-light EO — dusk, dawn, urban ambient light — is its own regime. The scene has just enough illumination to produce noisy, low-contrast frames. Detection thresholds have to be recalibrated. Backgrounds are cluttered with light sources that produce false positives. This regime is arguably harder than either full daylight or full darkness with thermal.

Three Regimes, Three Datasets

A CUAS program that only intends to operate in daylight can get away with a single training modality. Any program with a 24-hour operational requirement has three distinct problems to solve:

  • Daylight EO — the standard baseline, best-covered in commercial datasets
  • Low-light and transitional EO — dawn, dusk, urban glow; often skipped, high-value
  • Thermal (MWIR or LWIR) — required for full-dark and heavily-obscured environments

Each regime needs its own collection plan, its own class balance, its own edge-case library. And ideally the collections are time-synchronized so multi-modal fusion has aligned data to train on.

An adversary who knows your detector was trained on daylight footage will fly at night. Anyone building CUAS without a night-and-thermal collection plan is training a system that works half the time.

The good news is that thermal collection is not exotic. What it requires is a vendor with the sensor payload, a crew that knows how to operate it, and a program office willing to fund the second and third dataset rather than assuming the first one generalizes. The first two are procurable. The third is a decision.