HRRPLLM: LLM-Based ATR for Radar Imaging

• HRRPLLM is a framework that transforms high-resolution range profile (HRRP) scattering center data into textual prompts for LLM-based automatic target recognition, achieving notable accuracy in 1-shot setups.
• The approach highlights the limitations of a single averaged prototype per class, as it can mask intra-class variations due to different viewing angles.
• The Aspect-Distributed Prototype (ADP) strategy clusters support samples by aspect, yielding multiple prototypes per class, which improves robustness and performance in few-shot scenarios.

HRRPLLM (High-Resolution Range Profile LLM) refers to a methodology for automatic target recognition (ATR) in radar imaging that harnesses LLMs via in-context learning (ICL). This approach textualizes radar-derived scattering center (SC) distributions from high-resolution range profiles (HRRPs) and structures them as prompts for LLM inference. HRRPLLM predates the Aspect-Distributed Prototype (ADP) strategy, which further extends few-shot HRRP ATR robustness by modeling intra-class, aspect-dependent variance (Bi et al., 7 Dec 2025).

1. HRRP and ATR Fundamentals

High-resolution range profiles (HRRPs) are one-dimensional projections of a radar target’s echo energy, capturing the coherent superposition from multiple dominant scattering centers (SCs) distributed over range cells. Mathematically, the baseband HRRP vector $p \in \mathbb{C}^N$ is expressed as

$p = \Phi\alpha + \epsilon$

where $\Phi = [\phi(R_1), \dots, \phi(R_N)]$ is the Fourier basis for range-cell positions $R_n$, $\alpha \in \mathbb{C}^N$ denotes complex SC amplitudes, and $\epsilon$ models noise. The geometric and electromagnetic structure of the target is encoded by the number, position, and amplitudes of SCs, making HRRPs critical for ATR tasks.

2. HRRPLLM Framework

HRRPLLM, introduced by Chen et al. (2025), maps the HRRP ATR problem to the LLM domain by converting SC information into textual features and leveraging ICL (Bi et al., 7 Dec 2025). The canonical workflow includes:

1. SC Extraction: Dominant peaks are detected in HRRPs, each represented by (range index, normalized amplitude).
2. Prototype Generation: For each class $c$, support samples $x_i$ with label $y_i = c$ are encoded to SC text features $f(x_i)$. The class prototype is computed as:

$p^{(c)} = \frac{1}{K_c} \sum_{i: y_i = c} f(x_i)$

where $K_c$ is the number of class $c$ support examples.

1. Prompt Construction for ICL: The ATR task, SC semantics, reference prototypes (as JSON-like snippets), and a query SC list are assembled as a natural-language prompt.
2. Inference via LLM: The prompt is fed to a frozen LLM (e.g., GPT-4.1 or GPT-04-mini), which predicts the target class and optionally provides an explanation.

The similarity metric and classification logic are implicit, emerging from the LLM’s attention and next-token prediction mechanisms without backpropagated loss.

3. Performance Characteristics and Limitations

HRRPLLM demonstrated substantial improvements over classical SVM and random forest (RF) baselines for one-shot HRRP ATR. For example, GPT-04-mini HRRPLLM achieved ≈76.6% mean accuracy in 1-shot classification on a 12-class simulated HRRP benchmark, outperforming SVM and RF by 8–20% (Bi et al., 7 Dec 2025).

However, in few-shot settings (K > 1), especially on measured datasets with substantial aspect (viewing-angle) variation, HRRPLLM’s performance plateaued or degraded as support size increased. This phenomenon was traced to the strategy of pooling support SCs into a single “monolithic” prototype per class, which fails to capture systematic intra-class HRRP variations induced by changing viewing aspect. As a result, the averaged prototype $p^{(c)}$ became a poor representative when samples spanned diverse aspects.

4. Aspect-Distributed Prototype (ADP) Augmentation

To remedy aspect sensitivity, the ADP strategy clusters support SC features by aspect-dependent patterns and constructs multiple prototypes per class, each reflecting a distinct aspect cluster. The process is:

1. Extract SC features $\{f(x_i)\}_{i: y_i = c}$ for class $c$.
2. Cluster features into $A$ aspect bins $S_1, \dots, S_A$ via unsupervised methods (e.g., K-means).
3. Compute aspect-distributed prototypes:

$$p_a^{(c)} = \frac{1}{|S_a|} \sum_{x_i \in S_a} f(x_i), \quad a = 1\dots A$$

At inference, a query $x$ yields $f(x)$, which is compared (e.g., using cosine similarity) to all $p_a^{(c)}$. For each class, the aggregated score is

$$\text{score}(c) = \sum_{a=1}^A \text{sim}(p_a^{(c)}, f(x))$$

and the predicted label is $\hat{y} = \arg\max_c \text{score}(c)$.

Prompt construction for ADP explicitly lists each class’s $A$ aspect prototypes, facilitating LLM comparison and alleviating prototype dilution.

5. Experimental Evaluation

5.1 Datasets

• Simulated Aircraft Dataset: 12 classes (F-18, F-15, EP-3E, etc.), HRRPs sampled over 0°–60° azimuth, four polarizations, yielding profiles with up to 10 SCs per profile.
• Measured Aircraft Dataset: 3 classes (An-26, Yark-42, Cessna Citation), C-band HRRPs with significant real-world aspect variation.

5.2 Benchmarking Results

	SVM-HRRP	SVM-SC	RF-SC	GPT-4.1 HRRPLLM	GPT-4.1 ADP	GPT-04-mini HRRPLLM	GPT-04-mini ADP
1-shot	75.6/68.7	72.2/65.4	71.1/65.9	72.2/72.1	83.3/83.9	76.6/69.7	87.8/87.7
5-shot	90.0/86.7	72.2/64.4	78.9/73.0	83.3/82.4	83.3/83.9	87.8/86.5	88.9/88.4
10-shot	97.8/97.0	54.4/47.2	87.8/83.7	87.5/87.6	85.6/86.3	87.8/87.7	88.9/88.4

(Entries: “MeanAcc/F1” percent; (Bi et al., 7 Dec 2025))

ADP exhibits improved accuracy and F1, especially for GPT-04-mini, where 1-shot performance is +11 points higher than HRRPLLM. In measured datasets, HRRPLLM accuracy declines as $K$ increases, while ADP recovers or improves accuracy and F1 score (e.g., GPT-4.1 ADP at 20-shot yields +5.8 F1, +3.3 Acc over HRRPLLM).

A plausible implication is that explicit aspect-distributed prototype generation counteracts prototype confusion induced by aspect variance, which affects the reliability of averaging support features in classical HRRPLLM.

6. Implementation Aspects and Practical Considerations

• SC Peak Extraction: Careful tuning of peak detection (prominence threshold, spacing) is mandated to match radar characteristics and noise statistics.
• Clustering: K-means clustering on concatenated feature vectors (range index, normalized amplitude) determines aspect bins. $A$ (number of clusters) should be selected to ensure adequate sample count per bin, typically $A = \lceil K/2 \rceil$ for $K$-shot scenarios.
• Similarity: Cosine similarity on normalized SC-feature histograms.
• Prompt Formatting: Consistent, JSON-like syntax for SC lists, clear prototype labeling for aspect clusters.
• Training-Free: Both HRRPLLM and ADP operate without fine-tuning the LLM and rely on prompt engineering and classification via in-context matching.
• No Regularizers Required: Although one could add cluster separation regularization during prototype formation, empirical evidence did not necessitate this.

7. Significance and Outlook

HRRPLLM constitutes a domain-adaptive application of LLMs in radar ATR, bridging signal processing and NLP through semantic feature embedding and prompt-based inference. The ADP extension provides enhanced resilience to intra-class variance, notably aspect-induced variations that degrade performance in classical, one-prototype frameworks. Both frameworks are compatible with off-the-shelf LLMs, with minimal computational overhead and no requirement for model retraining.

This suggests opportunities for further integration of domain-specific clustering or feature engineering in LLM-based semantic ATR pipelines, particularly under few-shot settings where conventional methods suffer from overfitting and poor generalization.