A class emerges from a sequence – not from a single pixel
Use the slider to step through a schematic time window. The visualization shows why a temporal model can evaluate changes across several radar frames.
What does a radar classifier see?
A model does not receive an optical image, but a representation derived from radar processing. Different representations preserve different information.
Where + how fast radially?
Range-Doppler maps combine range and Doppler structure. Time sequences of such maps can make motion evolution visible.
Fine structure of motion
Time-frequency representations can highlight periodic or rotating micro-motion that is not visible in bulk Doppler alone.
Multiple perspectives
Depending on the system, several representations or hand-crafted features can be combined. More input does not, however, automatically mean better generalisation.
Why CNN and temporal modelling fit together
A CNN is strong at recognising local patterns in a representation. A temporal model adds the question of how these patterns change across the observation sequence.
Extracting local structure
Convolutional layers can learn recurring local patterns in range-Doppler or time-frequency data. The exact meaning of individual filters is not automatically physically interpretable.
Modelling temporal dependency
Recurrent models process a sequence step by step. Alternatively, temporal CNNs or transformer-like approaches can also model temporal information.
CRNN is an architecture family, not a quality promise. Whether a model works outside its training dataset depends more on data diversity, splits, ground truth and holdout design than on the architecture's acronym.
Class evidence is not the same as certainty
A classifier typically supplies scores or probability-like quantities. A robust system should not automatically interpret these values as calibrated real-world probabilities.
Known classes
Training and test contain the same class types. This is methodically simpler, but says little about how unfamiliar flying objects are handled.
Allowing uncertainty
An Unknown path is useful when evidence is insufficient or the pattern lies outside the known training space. Unknown is not an error state, but a conservative output.
Checking scores
Calibration examines whether model confidence and observed hit frequency match. A high softmax score alone does not prove reliability.
The hardest problem often begins after training
Radar features change with range, SNR, aspect angle, site, hardware, frequency band, clutter and target type. A model can appear strong on a test set and still become markedly weaker at a new site.
Schematic class evidence with a clearer margin.
Schematically flatter evidence – a sound Unknown mechanism becomes more important.
The bars are didactic and not Ozense measurements.
A good split design matters more than a spectacular accuracy figure
With temporally correlated radar data, random frame splitting can distribute near-identical sequences across train and test. This produces easily over-optimistic results.
Random frame split
Neighbouring frames from the same recording can end up in different splits. The model may then recognise recording conditions instead of generalisable target features.
Target-/session-/site-disjoint
Entire flights, target instances, sessions or sites are separated. This is harder, but closer to the actual question: does the model work on genuinely new data?
Accuracy alone is not enough. Class-wise errors, confusion matrix, calibration, Unknown behaviour, false-alarm burden and holdout tests under shifted conditions are also relevant.
Inference is part of the system – not just the model
For local real-time processing, latency, memory footprint, deterministic preprocessing and the exact match between training and runtime pipeline matter alongside model quality.
Measurement data.
Track-centred crop.
Identical to training.
Sequence inference.
Assessing confidence.
Class or Unknown.
A neural network does not create information the sensor never saw
If the relevant target signal disappears under clutter or noise, a classifier cannot reliably reconstruct the missing physical evidence.
Weak SNR
Less visible structure can worsen class separation.
Unknown targets
New drones, birds or other flying objects can lie outside the training space.
Site clutter
New background conditions can shift the feature distribution.
Aspect angle
The observed signature can change substantially with orientation and flight path.
Timing contract
A sequence architecture assumes that the temporal sampling matches the training assumption.
No product guarantee
Published model results do not automatically transfer to other hardware, ranges or sites.
This page explains the model – not the physics of drone and bird
The neighbouring page "Drone or bird?" explains which physical and temporal radar properties can provide distinguishable evidence in the first place. This page is about how a model processes such representations.
Frequently asked questions about CRNN and radar classification
What does CRNN mean?
Convolutional Recurrent Neural Network: convolutional layers for feature extraction are typically combined with a temporal model. The concrete architecture can look very different.
Why not just a CNN?
A CNN can also process single or stacked radar frames. An explicit temporal model can be useful when the temporal evolution itself carries relevant information.
Is micro-Doppler always the best input?
No. Micro-Doppler is important, but range-Doppler, HRRP, complex raw data or combined representations can also be relevant depending on sensor, SNR and task.
What is domain shift?
A change in the data distribution between training and deployment, for example through a new site, new hardware, different ranges or unfamiliar target types.
Why is an Unknown path needed?
Because real environments are not limited to the classes known during training. A conservative Unknown path can reduce overconfident class assignments.
Sources
- Rahman, S.; Robertson, D. A. (2020). Classification of drones and birds using convolutional neural networks applied to radar micro-Doppler spectrogram images. IET Radar, Sonar & Navigation. DOI:
10.1049/iet-rsn.2019.0493. - Han, S.-K.; Lee, J.-H.; Jung, Y.-H. (2024). Convolutional Neural Network-Based Drone Detection and Classification Using Overlaid FMCW Range-Doppler Images. Sensors 24(17), 5805. DOI:
10.3390/s24175805. - Remote Sensing (2024). Realizing Small UAV Targets Recognition via Multi-Dimensional Feature Fusion of High-Resolution Radar. Remote Sensing 16(15), 2710. DOI:
10.3390/rs16152710.