Signal-to-Noise Ratio (SNR)

SNR is the ratio of signal power (P_signal) to noise power (P_noise) , typically expressed in decibels (dB). The fundamental formula is SNR_dB = 10 * log10(P_signal / P_noise) . In digital communication systems, a related metric Eb/N0 (energy per bit to noise power spectral density ratio) is often used to normalize for different modulation orders and data rates. A higher SNR indicates a cleaner signal with less corruption from thermal noise, interference, or quantization errors.

Classification accuracy is directly proportional to SNR. At high SNR (>20 dB), modern deep learning classifiers routinely achieve >95% accuracy across common modulation schemes. Performance degrades sharply below 0 dB, where noise power equals or exceeds signal power. The SNR wall is the threshold below which reliable classification becomes impossible regardless of algorithm sophistication. Key performance benchmarks include:

• High SNR regime (>15 dB) : Near-perfect classification
• Mid SNR regime (0-15 dB) : Gradual accuracy decline, scheme-dependent
• Low SNR regime (<0 dB) : Rapid degradation, requires robust feature extraction

Blind SNR estimation is critical when ground truth is unavailable. Common methods include:

• M2M4 estimator: Uses second-order and fourth-order moments of the received signal to separate signal and noise power without prior knowledge of the constellation
• Maximum Likelihood (ML) estimation: Assumes known constellation geometry for optimal estimation
• Data-aided methods: Exploit known pilot symbols or training sequences embedded in the transmission
• Deep learning-based estimation: Neural networks trained to directly regress SNR from raw IQ samples or constellation diagrams, often outperforming classical estimators in non-AWGN channels

In deep learning for AMC, multi-SNR training is essential for robust deployment. Models are trained on signals spanning a wide SNR range (e.g., -20 dB to +30 dB) to learn noise-invariant features. Techniques include:

• Uniform SNR sampling: Equal probability across the training range
• Curriculum learning: Progressively decreasing SNR during training to improve low-SNR performance
• SNR-aware architectures: Models that receive estimated SNR as an auxiliary input to condition their classification logic on current channel quality This approach prevents catastrophic failure when deployed in dynamic electromagnetic environments.

SNR alone is insufficient in fading channels. The instantaneous SNR fluctuates due to multipath propagation, Doppler shifts, and shadowing. Effective SNR metrics include:

• Average SNR: Mean signal power over the fading distribution
• Effective SNR: Accounts for the bit error rate (BER) degradation caused by fading relative to AWGN
• Outage probability: The probability that instantaneous SNR falls below a minimum threshold required for reliable classification Classifiers must be evaluated across the full SNR distribution, not just at fixed points, to guarantee real-world performance.

In MIMO (Multiple-Input Multiple-Output) systems, SNR is defined per spatial stream. The post-processing SNR after equalization or beamforming determines classification performance for each individual stream. Key concepts include:

• Spatial multiplexing gain: Increases data rate but reduces per-stream SNR
• Diversity gain: Improves effective SNR through combining multiple received copies of the same signal
• Condition number of the channel matrix: A high condition number indicates ill-conditioned channels where some spatial streams experience significantly lower SNR, complicating modulation recognition