Radar Target Classification with Deep Learning

Review the core concepts you'll need: the radar range equation (received power as a function of range, RCS, and system parameters), range-Doppler processing (how a pulsed radar converts time-delay to range and frequency shift to velocity), and radar cross-section (RCS) — the effective reflective area of a target, measured in square meters (dBsm).

Understand what a range-Doppler map (RDM) looks like: a 2D image where the x-axis is Doppler frequency (proportional to target radial velocity), the y-axis is range (distance from radar), and pixel intensity represents received power. Each target appears as a bright spot at its range-Doppler coordinates, potentially spread by its physical extent and motion components. Study published examples of RDMs for different target types to develop intuition.

Create physics-based models for four target classes. Commercial aircraft: RCS ~10–100 m² (large, strong return), Doppler corresponding to 200–500 knots, minimal micro-Doppler. Small drone: RCS ~0.001–0.01 m² (very small return), speed 20–60 knots, strong micro-Doppler from spinning propellers (periodic modulation at blade-pass frequency). Bird: RCS ~0.001–0.01 m² (similar to drone), speed 15–40 knots, distinctive wing-beat micro-Doppler (sinusoidal modulation at 3–10 Hz).

Clutter: ground and weather returns that create false targets. Model ground clutter as a band of returns at zero Doppler across all ranges, and weather clutter as a diffuse return with a Gaussian Doppler spectrum. The key classification challenge is separating drones from birds — they have similar RCS and speed, but their micro-Doppler signatures differ (rotating propellers vs. flapping wings).

Implement the radar signal processing chain in NumPy. For each simulated target: (1) generate a sequence of radar pulses, (2) add the target return at the appropriate range bin with phase shift corresponding to its velocity, (3) add micro-Doppler modulation based on target type, (4) add thermal noise (complex Gaussian, scaled by SNR), (5) apply a 2D FFT (range compression + Doppler processing) to produce the range-Doppler map.

Generate 1,000–2,000 RDMs per class with randomized parameters: target range (1–20 km), velocity (within class-appropriate bounds), SNR (5–30 dB), and clutter level. Save as 64×64 or 128×128 grayscale images. The randomization ensures the classifier learns the signature shape rather than memorizing specific range-velocity combinations. Apply a CFAR (Constant False Alarm Rate) detector to normalize the background level across different noise conditions.

Range-Doppler maps have different statistical properties than natural images, so standard ImageNet architectures may not be optimal. Design a CNN specifically for RDM classification: start with smaller kernels (3×3) in early layers to capture fine micro-Doppler structure, followed by larger receptive fields in deeper layers to integrate across range-Doppler extent.

Implement a 5-block CNN: each block is Conv2d → BatchNorm2d → LeakyReLU → MaxPool2d, with channel progression 1 → 32 → 64 → 128 → 256 → 512. Use AdaptiveAvgPool2d followed by fully connected layers (512 → 256 → 4). Add dropout (0.4) between FC layers. Also implement a ResNet-18 adapted for single-channel input as a comparison baseline.

Split data 70/15/15 with stratification. Train with Adam optimizer, initial learning rate 1e-3, and OneCycleLR scheduler. Use cross-entropy loss with class weights if any class is underrepresented.

Critically, ensure your training and test sets contain a uniform distribution of SNR values. If you train only on high-SNR examples, the model will fail on noisy real-world data. After initial training, evaluate accuracy as a function of SNR — plot a classification accuracy vs. SNR curve for each target type. This operational performance curve tells a radar engineer exactly when the classifier can be trusted: "above 12 dB SNR, drone vs. bird classification is 90% accurate."

The hardest and most valuable classification task is drone vs. bird. Analyze what the CNN learned about micro-Doppler: use Grad-CAM to visualize which regions of the range-Doppler map drive the classification decision. For drones, the CNN should attend to the periodic propeller harmonics (discrete spectral lines). For birds, it should attend to the broader, sinusoidal wing-beat modulation.

Generate a confusion matrix focused on the drone/bird pair across different SNR levels. At what SNR does discrimination break down? Try an ablation: remove the micro-Doppler component from your synthetic data and re-train — how much does accuracy drop? This quantifies the information value of micro-Doppler for target classification, which is a publishable research finding.

Test the classifier's robustness to conditions not seen in training. Generate test cases with: multiple simultaneous targets (overlapping returns), maneuvering targets (time-varying Doppler), and mismatched clutter models (weather clutter intensity different from training). Report accuracy degradation for each condition.

Document the complete system: synthetic data generation methodology (with enough detail for reproduction), CNN architecture and training procedure, performance results across all conditions, and a clear-eyed discussion of the sim-to-real gap — the difference between synthetic and real radar data that remains the fundamental challenge in this field. This project is publishable at IEEE Radar Conference or SPIE Defense + Commercial Sensing, and represents the type of work done at radar research labs nationwide.