Simulated radar echoes trained a drone-and-bird classifier.
A drone and a bird can return similar radar signals. Their moving blades and wings leave different patterns in the echo. Simulation creates those patterns for a model to learn from.
Simulate radar echoes from drones and birds as their blades and wings move. Train a model to distinguish them, then test its predictions on separate synthetic examples with noise and clutter.
Edited demonstration briefThe result
Synthetic echoes supported recognition, with classification errors still visible.
The classifier distinguishes drones from birds in synthetic tests, while identifying the exact target type remains harder. The recorded approach includes misclassifications, and simplified bird motion makes the task easier than it would be with real targets.
Synthetic radar data only. Simplified bird motion makes classification easier; no real-target detection validation.
Recorded simulation
Blade and wing motion produce different radar patterns.
Compare recorded drone and bird signals, and follow the classifier's changing assessment.
Recorded outputs, replayed in your browser. No new simulation is running.
A similar radar echo can hide a different object.
The world we created
An electromagnetic solver computes how radio waves scatter from virtual drones and birds as their blades or wings move. The resulting echoes are combined with simulated motion, noise and clutter.
What this enables
A classifier learns to distinguish drones from birds and estimate the target type from the changing echo. A separate learned model predicts radar scattering at viewing directions and rotor phases.
Simulated echoes had known targets and motion.
Vary the target, viewing direction and blade or wing position. Combine the computed scattering with simulated motion, noise and clutter.
A classifier learned patterns in the synthetic echoes.
Train a classifier on the simulated radar patterns. Train a separate model to approximate how targets scatter the signal.
Separate synthetic tests exposed recognition errors.
Evaluate classification on separate synthetic examples and scattering predictions on held-out rotor phases. Compare the recorded decisions with the known target.
Physics creates labelled radar examples whose source and motion are known. Testing the learned models on simulated examples exposes confusion and the limits of the synthetic data.
How the model was tested
Physical model. The bempp-cl boundary-element solver uses the electric-field integral equation to model electromagnetic scattering. Targets are rigid perfect electric conductors, with blades and wings represented by thin screens. Phase-dependent scattering generates micro-Doppler patterns for a convolutional classifier; a separate neural surrogate predicts complex scattering.
Training experience. Simulated drones and birds are sampled across viewing directions, frequency bands and rotor or wing phases. Synthetic radar observations include randomized trajectories, thermal noise and statistical clutter. Stored validation predictions and approach tracks retain the known targets.
Evaluation. Classification is checked on a separate synthetic validation split. The scattering surrogate is checked on rotor phases excluded from training. Sphere and plate comparisons test the electromagnetic calculation against analytic references, rather than validating detection on real drones or birds.
- Bird motion is undersampled, suppressing part of its Doppler signature and making drone-versus-bird classification easier. A drone rotor sequence also marginally undersamples its highest harmonic.
- Perfect-conductor targets, rigid bodies and thin blades or wings simplify real materials and motion.
- Clutter is statistical rather than a simulated terrain scene. Ground multipath, antenna effects and operational detection performance were not validated.
- The demonstration supplies synthetic perception data and learned inference. It does not establish autonomous flight control or deployment-ready detection.
Original research artifacts
The original files retain the recorded measurements and model details.