A Comprehensive Benchmark of Recurrent, Convolutional, and Attention-Based Models for Radar Emitter Classification Under Photon-Starved Conditions

The identification of radar scanning patterns from
pulse amplitude sequences is a critical function in electronic in
telligence and electronic warfare. While deep learning has proven
highly effective for this task under additive white Gaussian noise,
its performance in photon-starved sensing environments remains
largely unexplored. This work presents a systematic benchmark
of eight deep learning architectures—LSTM, BiLSTM, GRU,
xLSTM, CNN, TCN, InceptionTime, and Transformer—under
identical quantum-noise-limited conditions. A physically realistic
Poisson noise model, incorporating pulse amplitude distortions,
missing pulse events, spurious outliers, and frequency-dependent
gain variations, is used to generate nine distinct noise scenarios.
Each architecture is trained and evaluated over ten indepen
dent runs with consistent preprocessing and training protocols.
Experimental results show that all models achieve 86–89%
accuracy under classical conditions but suffer a significant
degradation to 47–60% accuracy under quantum noise. Among
the evaluated architectures, InceptionTime achieves the highest
mean quantum accuracy (56.48%), while Transformer exhibits
the lowest (45.82%). Statistical analysis reveals no significant
performance differences across architectures, indicating that the
observed degradation is architecture-independent and fundamen
tally driven by photon-counting statistics. These findings establish
a clear performance bound for amplitude-only classification
under quantum-limited sensing and underscore the urgent need
for phase-aware or coherent detection strategies in future radar
emitter classification systems.