Radar signal processing with AI: target classification and clutter rejection at the edge

Radar has been the primary all-weather sensor for air and ground surveillance since the Second World War, but the classical signal processing chain -- pulse compression, range-Doppler FFT, constant false alarm rate detection, and track filtering -- was designed for a world of cooperative targets and predictable clutter. Modern battlefields introduce low-observable drones that spend most of their energy in clutter bins, electronic attack waveforms that saturate the receiver, and heterogeneous terrain backgrounds that violate the homogeneous clutter assumptions underpinning conventional CFAR. Neural networks trained on large libraries of measured and synthetic radar data can learn to characterize these environments directly from the I/Q signal stream, augmenting the classical pipeline where it is weakest. This article walks through where AI enters the radar signal processing chain, what hardware runs it, and how on-device target recognition and classification change the operational calculus for edge-deployed radar systems.

Why radar signal processing needs AI augmentation

The fundamental limitation of classical radar signal processing is that every algorithm assumes a statistical model for clutter and interference. CA-CFAR assumes clutter power is uniform across the reference window. The Kalman filter assumes linear target dynamics and Gaussian process noise. Pulse-Doppler waveform design assumes the target resides in a clean Doppler bin uncontaminated by ground returns. Each assumption is a reasonable approximation in the scenario the algorithm was designed for, and a source of performance degradation everywhere else. When a small quadrotor drone hovers in the main-lobe clutter region of a ground surveillance radar, or when a jammer injects a coherent repeater signal that mimics a real target, the classical chain has no principled way to recover -- it was not designed for those conditions.

AI augmentation addresses this gap by replacing or supplementing the fixed statistical models with learned functions trained on data from the actual operating environment. A neural CFAR trained on heterogeneous terrain data learns a detection threshold function that adapts to clutter gradients, discrete clutter scatterers, and rain bands without requiring an explicit model of each. A micro-Doppler classifier trained on drone signatures learns the distinction between a quadrotor, a fixed-wing UAS, and a bird from the characteristic frequency modulation patterns each produces -- patterns that no fixed threshold can cleanly separate. The key insight is that AI does not replace the radar's physics; it replaces the human-designed heuristics that map physics to decisions, replacing them with learned mappings that generalize better across the real diversity of the operating environment.

The practical case for running these neural modules at the edge -- on the radar platform itself rather than at a remote processing node -- rests on latency and availability. A ground surveillance radar that must datalink every range-Doppler map to a cloud processor for neural inference introduces 200-500 ms of round-trip latency that is unacceptable for cueing interceptors or alerting ground forces. More critically, the datalink is a single point of failure: electronic attack, terrain masking, or network congestion disables the AI capability exactly when it is most needed. Edge inference preserves the AI capability in denied and degraded environments, which is precisely the scenario that motivated the AI augmentation in the first place.

Doppler processing and moving target indication: classical baselines

The range-Doppler map is the fundamental data product of a pulsed Doppler radar. A coherent processing interval of N pulses is collected, pulse-compressed in range, and then transformed along the slow-time (pulse) dimension by an N-point FFT to produce a two-dimensional map of radar cross section as a function of range and radial velocity. The range resolution is determined by the signal bandwidth (typically 1-100 MHz for ground surveillance radars, giving 1.5-150 m resolution), and the Doppler resolution is determined by the coherent processing interval duration (typically 0.05-2 seconds, giving 0.5-10 m/s velocity resolution at X-band). This map is the input to all subsequent detection and classification processing, and its quality -- sidelobe level, dynamic range, sensitivity -- sets a hard ceiling on what any downstream algorithm can recover.

Moving target indication (MTI) and pulse-Doppler filtering are the classical tools for separating moving targets from stationary clutter. MTI uses delay-line cancellers or finite impulse response filters to null the zero-Doppler clutter region in the Doppler spectrum, typically achieving 30-50 dB of clutter cancellation in benign conditions. The limitation is the clutter notch width: a radar on a moving platform (airborne, vehicle-mounted, or shipborne) must apply space-time adaptive processing (STAP) to handle the fact that platform motion smears the clutter Doppler spectrum across a range of Doppler bins that depends on aspect angle. STAP solves a joint space-time optimization problem across the array's spatial channels and the slow-time dimension, requiring an estimate of the clutter covariance matrix from training data -- an estimate that degrades when the clutter is non-stationary or when the training data contains target contamination.

Neural augmentation of the Doppler processing stage typically takes one of two forms. In the first, a neural network replaces the STAP weight computation, learning to predict optimal spatial-temporal weights from a compact representation of the local clutter environment rather than from an explicit covariance matrix inversion. This approach is particularly valuable in low-sample-support regimes where the covariance estimate is poorly conditioned -- a persistent problem for radars with large arrays and rapidly changing clutter environments. In the second form, the neural network operates downstream of classical STAP as a post-filter, examining the clutter-cancelled range-Doppler map and suppressing residual clutter returns that the linear filter failed to cancel, reducing the false-alarm density fed to the CFAR stage by 30-50% in heterogeneous terrain.

CFAR detection and its limitations: where neural detectors help

Constant False Alarm Rate detection is the standard method for setting a range-Doppler detection threshold that maintains a specified probability of false alarm regardless of local clutter power. The classical cell-averaging CFAR (CA-CFAR) estimates clutter power from a window of reference cells surrounding the cell under test (CUT), excluding a small guard region around the CUT to prevent target energy from biasing the clutter estimate. The threshold is set as a multiple of the estimated clutter power, where the multiple is chosen to achieve the desired false-alarm probability under the assumption of Rayleigh-distributed clutter. This is a sound approach when clutter is spatially homogeneous -- but the battlefield is not a homogeneous medium.

Three failure modes of CA-CFAR are well documented. At clutter edges -- range bins where the radar transitions from low-reflectivity ground to a higher-reflectivity feature such as a treeline, building, or river bank -- the reference window straddles the edge, and the clutter estimate is biased high on the low-clutter side, causing masking of real targets, and biased low on the high-clutter side, causing a burst of false alarms. In the presence of strong discrete scatterers (metal fences, vehicles, tall structures), a single high-power return in the reference window inflates the clutter estimate and masks nearby weaker targets. And in rain or chaff, volume clutter fills many range-Doppler bins simultaneously, producing a non-stationary, spatially correlated interference environment in which the i.i.d. assumptions of CA-CFAR are badly violated.

Neural CFAR variants address these failure modes by learning the local clutter distribution from a wider spatial and Doppler context than a fixed-size reference window can capture. A CNN applied to a patch of the range-Doppler map centered on the CUT learns to recognize clutter edge patterns, discrete scatterer constellations, and volume clutter textures, and outputs a locally adapted threshold that accounts for the observed heterogeneity. In field evaluations against ground targets in mixed terrain, neural CFAR reduces false-alarm density by 40-60% compared to CA-CFAR at matched probability of detection, translating directly to reduced operator workload and lower track initiation latency. The same architectural patterns used in ATR edge systems -- compact CNNs with INT8 quantization -- apply directly to neural CFAR deployment on radar hardware.

Neural target classifiers on radar micro-Doppler signatures

Once a target passes CFAR detection, the radar has a range-Doppler cell containing a return from something moving. Knowing that something is there is valuable; knowing what it is -- drone, vehicle, person, bird -- is operationally decisive. Micro-Doppler analysis extracts this classification information from the modulations superimposed on the target's main Doppler shift by rotating or vibrating substructures. A quadrotor drone's four spinning rotors produce periodic Doppler sidebands at multiples of the rotor blade passing frequency. A walking person's arms and legs produce a characteristic sinusoidal modulation pattern. A wheeled vehicle's engine produces harmonic vibrations modulated by vehicle speed. Each of these signatures appears as a distinct structure in the short-time Fourier transform (STFT) of the complex radar return over a dwell period of 0.1-2 seconds.

The standard pipeline for micro-Doppler classification collects a dwell on the detected target, computes the STFT of the complex range-bin time series to produce a time-frequency map, and passes the map through a CNN trained to distinguish target classes. For a four-class problem (pedestrian, wheeled vehicle, drone, bird), MobileNetV2-scale CNNs trained on large datasets of measured micro-Doppler signatures achieve 90-97% classification accuracy across representative operating conditions. The input feature map is typically normalized to zero mean and unit variance per-pixel to remove the dependence on absolute signal power, making the classifier robust to range variation and aspect-dependent RCS fluctuations. Inference time on a Jetson Orin Nano is 8-15 ms per dwell, well within the radar's track update cycle.

The hardest classification problem is the drone-versus-bird discrimination that is now operationally critical for counter-UAS radar systems. Small commercial drones and large birds occupy overlapping RCS ranges (0.001-0.1 m2 at X-band) and have similar flight speed envelopes. The discriminant is in the micro-Doppler: birds produce an asymmetric flapping pattern with a slow downstroke and fast upstroke, while quadrotors produce symmetric periodic sidebands at the rotor RPM. Trained CNN classifiers achieve 92-95% discrimination accuracy in clear air; accuracy degrades in rain (water-coated wing surfaces change the bird micro-Doppler signature) and at long range (SNR falls below the level needed to resolve the modulation structure). Combining micro-Doppler classification with multimodal sensor fusion from EO and acoustic channels substantially improves overall discrimination accuracy in these marginal conditions.

Clutter rejection: separating returns from weather, terrain, and chaff

Clutter in radar is any return that is not the target of interest. The three dominant clutter sources in ground and air surveillance radar are terrain (ground clutter from the illuminated surface), weather (rain, snow, hail producing distributed volume returns), and chaff (metallic dipole clouds deployed by aircraft as a countermeasure). Each requires a different rejection approach because each has a distinct spectral and spatial structure, and a clutter rejection filter tuned for one type can actually worsen performance against another.

Ground clutter is spectrally concentrated near zero Doppler for a stationary radar. The classical rejection approach -- a Doppler notch filter -- works well for slow-moving targets at speeds above the minimum detectable velocity (MDV), but fails for targets moving at low speed or hovering in the clutter Doppler band. AI-augmented ground clutter rejection uses a neural network that learns the spatial texture of ground clutter in the range-Doppler map -- the characteristic ridge structures from extended terrain features, the discrete spikes from buildings and trees -- and suppresses them while preserving target energy in overlapping Doppler bins. This is particularly important for detecting slow-moving vehicles and dismounts that conventional MTI filtering would cancel along with the clutter.

Weather clutter and chaff present different challenges. Rain produces a distributed, spatially smooth return whose Doppler spectrum is centered on the wind velocity rather than zero, making Doppler notching ineffective without knowledge of the wind field. Chaff clouds drift with the wind and bloom to cover a large range-Doppler volume over tens of seconds. Neural classifiers trained on the spatial and temporal evolution of chaff clouds can track the cloud boundary and suppress returns within it while maintaining detection in adjacent clear airspace. The classifier exploits features that human-designed filters cannot easily encode: the characteristic ellipsoidal shape of a blooming chaff cloud, the gradual Doppler drift as the cloud equilibrates with the wind, and the power density profile that differs from both aircraft returns and weather. For each of these clutter types, the neural approach achieves rejection improvement ratios comparable to or exceeding the best classical adaptive filters, while generalizing across a wider range of environmental conditions.

ECCM integration: detecting and characterizing jamming at the edge

Electronic Counter-Countermeasures (ECCM) are the techniques a radar uses to maintain detection performance under deliberate electronic attack. Classical ECCM draws on a toolkit of techniques -- frequency agility, pulse compression gain, low sidelobe antenna design, sidelobe blanking, and coherent excision of narrowband interference -- each designed to counter a specific jammer type. The limitation of this approach is that it requires the radar to know which jammer type it faces before selecting the appropriate countermeasure. Against a sophisticated adversary using an adaptive jammer that changes its waveform in response to the radar's behavior, a fixed ECCM strategy is insufficient.

Edge AI changes the ECCM architecture by enabling closed-loop jammer characterization within the radar's own processing chain. A neural classifier monitoring the radar's receiver output -- specifically the noise floor, sidelobe region, and cancelled-clutter residual -- learns to recognize the signatures of different jamming waveforms: broadband noise jamming (elevated noise floor across all range bins), swept spot jamming (a narrow-band interference ridge that sweeps through the Doppler dimension), coherent repeater jamming (false targets that appear at physically consistent range and Doppler but fail velocity cross-checks), and deceptive range-gate pull-off (targets that gradually drift in range away from their initial position). Each signature maps to a neural class label and triggers a corresponding ECCM response within the radar's mode controller.

The operational advantage of AI-driven ECCM identification over a manual human-in-the-loop process is reaction time. A skilled electronic warfare officer analyzing a scope can identify a jammer type in 5-30 seconds. A neural classifier running on the radar's signal processor identifies the jamming type in under 100 ms from the onset of the attack, selects the optimal ECCM mode, and reports the jammer's estimated direction of arrival to the C2 network for cueing of other sensors. For short-dwell acquisition radars used in air defense, a 30-second identification lag can mean the difference between a successful intercept and a missed engagement window. The neural ECCM module also provides a persistent characterization record -- every jamming event is logged with its waveform parameters and classifier confidence -- that feeds the electronic order of battle database maintained by the edge AI hardware stack supporting the sensor network.

Hardware deployment: DSP platforms, FPGAs, and GPU inference for radar AI

The choice of hardware for AI-augmented radar signal processing is more constrained than for other edge AI domains because the radar signal processor must also perform the real-time FFT, pulse compression, and CFAR processing that cannot be deferred to a general-purpose GPU. The three main hardware categories each make different tradeoffs between determinism, power, and neural inference throughput.

FPGAs occupy the deterministic, low-latency end of the spectrum. A Xilinx Zynq UltraScale+ MPSoC integrates the full range-Doppler pipeline in programmable logic alongside INT8 neural network inference implemented on the fabric's DSP slice array, delivering end-to-end latency from ADC sample to detection event under 1 ms for a 256-range-bin, 64-Doppler-bin map. The programming complexity is substantial -- implementing even a compact CNN on an FPGA requires high-level synthesis tools such as Vitis AI -- but the result is a processing chain with no operating system jitter, deterministic worst-case latency, and power consumption in the 5-25 W range appropriate for vehicle-mounted or fixed ground radar. The Agilex 7 generation extends this capability with embedded AI tensor blocks that further accelerate INT8 convolution without consuming general-purpose DSP slice resources.

GPU-class SoCs such as the NVIDIA Jetson Orin AGX (275 TOPS at 60 W) and the lower-power Orin NX (100 TOPS at 25 W) provide the highest neural inference throughput for complex classifiers and multi-dwell micro-Doppler analysis. TensorRT-optimized INT8 models run at 5-15 ms per inference call for MobileNetV2-scale classifiers, and the Orin's DLA (Deep Learning Accelerator) hardware engine can process one inference in parallel with the iGPU running another, enabling true concurrent classification of multiple tracked targets without pipeline bubbles. The hardware comparison for defense edge AI shows Jetson Orin as the preferred platform when classification throughput and model diversity matter more than absolute worst-case latency. For radar applications, a hybrid architecture pairing an FPGA front-end (range-Doppler map generation, CA-CFAR) with a Jetson Orin back-end (neural CFAR, micro-Doppler classification, ECCM characterization) captures the strengths of both: deterministic signal processing pipeline latency at the front, high-throughput flexible neural inference at the back.

DSP platforms from Texas Instruments (TMS320C7x series) and Analog Devices (SHARC+) remain relevant for legacy radar programs and for applications where the deterministic, single-core programming model is mandated by certification requirements. The C7x architecture includes a 512-bit SIMD vector unit and a matrix-multiply accelerator that can execute INT8 convolution at competitive throughput for networks up to 2-3 million parameters. The practical ceiling is the memory bandwidth available to feed the convolution engine: the C7x's on-chip L2 cache can sustain the bandwidth needed for layers with small spatial dimensions (1x1 and 3x3 convolutions on small feature maps) but falls short for the wider feature maps produced by the initial layers of a full MobileNet applied to a radar range-Doppler map. This makes DSP platforms best suited to the final classification layers operating on compact feature vectors extracted upstream by the FPGA, rather than to full end-to-end neural inference from raw range-Doppler data.