Radar Jamming Recognition Feature Extraction: Methods, Models, and Performance

Radar jamming recognition feature extraction turns raw jamming signals into discriminative representations that classifiers can separate, even at low jamming-to-noise ratios. This guide walks through handcrafted features, deep learning pipelines, parameter-invariant learning, and the accuracy numbers recent studies actually report.

What Is Radar Jamming Recognition and Why Does Feature Extraction Matter?

Feature extraction is the field that defined how accurately any radar jamming recognition system could work. A receiver receives some combination of target echoes and hostile emissions, an extractor extracts the waveform to numbers that discriminate one type of jamming from another. No downstream classifier can retrieve that lost info if the features are weak. I think as I read the literature this is why so many papers devote most of their pages to representation and secondarily describe a classifier.

Radar jamming covers way more ground than most people assume when they first get into this field. There's a systematic review out of IEEE JSTARS—Zan Wang, Zhengwei Guo, Gaofeng Shu, and Ning Li—that sorts jamming models into three buckets: deception jamming, suppression jamming, and smart jamming. From there, it breaks recognition methods into traditional machine learning and neural network camps, with the latter further divided into shallow and deep architectures. That kind of framework is genuinely helpful as a mental map, because it makes one thing clear: feature extraction isn't a single trick you apply everywhere. It's a whole set of choices, and which ones make sense depends on the jamming mechanism you're actually trying to catch.

Which Handcrafted Features Are Used for Jamming Signal Recognition?

Before the era of deep networks, engineers designed features by taking hand-crafted signal statistics and geometry. The multi-dimensional feature extraction; Information dimension, box dimension in time and frequency domain [2025]G Hao. Tie those fractal measures with energy entropy, information entropy and standard deviation in the related studies while simply establish baselines applying mean value and its deviation. All three capture different properties: entropy is disorder, fractal dimension waveform complexity and deviation amplitude spread.

Time-frequency and decomposition tools augment the handcrafted toolbox. The short-time Fourier transform (STFT) is used to convert one-dimensional capture data into a two-dimensional time-frequency image, variational mode decomposition (VMD), which signal decomposes into some modes whose energy entropy and information entropy use as features. One more established yet still referenced work by X Tian in 2013 employed a product spectrum matrix (SPM) based on non-negative matrix factorization (NMF) to detect deception jamming. Common handcrafted families and what they measure are summarized in the table below.

How Do Deep Learning Models Extract Jamming Features Automatically?

With the filter learned over training of data, Deep Learning removes manual design. Results are easy to miss. In March 2026, a cross-modal approach was proposed that employs multimodal convolution network along with multilevel feature extraction module to jointly capture the relevant features from both horizontal (1D tensor) and vertical (2D representations), powered by robust classifiers despite varying degrees of jamming-to-noise ratios within each modality [144]. The idea is that a raw time series and its image in the (time, frequency) plane contains complementary information and we can let network fuse both instead of forcing engineer to choose one.

Architecture choices are as important to the input format. The first one published on March 31, 2026 is a design using attention-based network called P-MSGC-AM-CNN which integrates the use of bottleneck-structured multi-scale ghost convolution with message passing integration and channel attention (MSGC-CA) and spatial attentional feature channels utilizing combined spatial-channel comparative optimization along multiple scales. Increased (up to some point) model size yields a higher margin, but takes the trade off here run faster small models on ginormous datasets are high in parameter count and over-fit its only us that suffers. That network averaged over 89.8% accuracy across the same jamming types (28 active), beating ResNet18 and three other published architectures at -10 dB JNR, and is presented as suitable for resource-constrained real-time applications by its authors [6].

Simpler convolutional baselines still show up in the comparisons, and they're worth paying attention to. You'll see studies that benchmark 1D-CNN and CV-CNN across 12 jamming types, or pair STFT with a CNN on nine LFM jamming types. What's interesting is that these setups perform reasonably well, which tells you something important: how you represent the input often matters more than how deep you make the network.

What Is Parameter-Invariant Feature Learning for Jamming Recognition?

One of the recurring common problems for early feature sets is that they drift according to jammer parameters. M Luo et al. (2025) directly address this with parameter-invariant feature learning that decouples intrinsic features from variations due to parameters so recognition remains consistent, even when the emitter moves its frequency, power or modulation settings. The experimental dataset included six types of radar jamming signal, each being stored as a time series with dimensions 8001x2. That dimension count is a good reminder for just how much raw data can be contained in one sample before any sort of compression.

Parameter dependence is not an issue of aesthetics. When a classifier learns that a pulse width always implies one jamming type, the adversary can then break the model invisibly by adjusting some parameters ever so slightly. In other words, decoupling intrinsic structure from nuisance parameters is not just a preprocessing trick but actually fits like hand and glove with the practical limited-sample strategies detailed in the IEEE JSTARS review.

How Are Compound and Deception Jamming Recognized Under Low JNR?

Compound jamming is the situation in which single-label classifiers fail, when multiple jsmming mechanisms overlap within a same capture. Deep Residual Network with Spatial-Channel Attention for Compound Jamming Recognition Based on Image Processing (2025): P Li et al., a compound jamming signal recognition using ResNet optimized model. RF jamming sends high-power radio frequency signals, which produce noise in the radar working band (this is also a reason for segmentation instead of taking all things through classification), according to the paper. W Lu et al recognize complex radar jamming signals from a low snr compound complicated data set by utilizing deep learning methods in 2025 [6].

Deception jamming needs a different lens because the jammer imitates the target rather than drowning it. L Du's 2024 work from the Beijing Institute of Technology, published at IEEE ICSIDP 2024 in Zhuhai, China, analyzes scattering point angle and polarization differences between target echoes and deception jamming, then computes jamming clustering thresholds per feature from measured radar data. The method maintains performance under low jamming-to-noise ratios, and its keywords include dense false target jamming, feature extraction, and radar jamming identification. A related 2025 study on SAR jamming recognition via dual-channel feature extraction targets low jamming-to-signal ratio conditions.

Few-shot and open-set settings round out the picture. H Zhou's 2025 prototype-feature-driven few-shot work adds prototype feature extraction and fusion to a feature pyramid network, raising overall recognition accuracy by about 2.3% over the baseline. Y Liu's 2025 work on open-set recognition for active radar jamming uses deep learning to adaptively filter and represent data, while Y Zhang's 2023 open-world study reached 91.93% accuracy with nine known targets and one unknown target. W Zha's 2026 lightweight multi-feature fusion network reported over 87% recognition accuracy for seven compound jamming types on simulated datasets.

What Accuracy Do Recent Radar Jamming Recognition Methods Achieve?

Reported accuracy varies widely because datasets, jamming counts, and noise conditions differ, so cross-paper comparison needs care. The table below collects the headline numbers from the studies cited here, along with the condition each number was measured under. I treat these as directional benchmarks rather than a single leaderboard, since a 91.93% result on ten targets is not directly comparable to 89.8% across 28 jamming types.

What Limits Current Feature Extraction Approaches?

The honest answer is that most published gains come with conditions attached. Limited sample data forces researchers into few-shot and prototype strategies, strong electromagnetic jamming degrades any feature that depends on clean amplitude statistics, and low JNR or JSR conditions remain the hardest test bed. Open-world deployments add unknown targets that a closed-set classifier will confidently mislabel, and large models risk overfitting when the training set is small. Parameter variation, as M Luo's work shows, can quietly erode features that looked strong in simulation.

These limits explain the direction of travel in the field. The IEEE JSTARS review by Wang, Guo, Shu, and Li devotes explicit attention to limited-sample strategies alongside feature extraction and neural network design, which signals that data scarcity is treated as a first-class problem rather than an afterthought. For anyone building a real system, the practical takeaway is to validate under the JNR and JSR ranges you actually expect, not the comfortable ones.

How Should You Choose a Feature Extraction Pipeline?

Start from the jamming type you must catch, not from the newest architecture. Deception jamming rewards polarization and angle-domain features, compound jamming rewards segmentation plus multi-domain fusion, and suppression jamming often responds to entropy and fractal measures. If your dataset is small, prototype fusion or parameter-invariant decoupling will likely help more than adding depth. If latency matters, lightweight multi-scale convolution with attention is a reasonable default, since the P-MSGC-AM-CNN results suggest you can stay under 90% accuracy loss while cutting parameters.

Finally, benchmark against a simple baseline before committing. A 1D-CNN or an STFT plus CNN pipeline is cheap to build and tells you whether your problem is representation-limited or data-limited. Only after that diagnosis does it make sense to invest in cross-modal fusion or attention-heavy designs.

Frequently Asked Questions

What features are extracted for radar jamming recognition?

Methods extract information dimension and box dimension in the time and frequency domains, plus energy entropy, information entropy, and standard deviation. Statistical metrics include mean value and deviation. Some approaches use product spectrum matrix features with non-negative matrix factorization, or multiscale features pulled from both 1D tensors and 2D time-frequency representations.

How accurate is deep learning for radar jamming recognition?

A lightweight multi-feature fusion network reached over 87% accuracy for seven compound jamming types. The P-MSGC-AM-CNN algorithm achieved over 89.8% average accuracy across 28 active jamming types at -10 dB jamming-to-noise ratio. An open-world method reached 91.93% with nine known targets and one unknown target.

What is parameter-invariant feature learning in radar jamming recognition?

It decouples intrinsic features from parameter variations so recognition stays robust when jamming parameters change. The experimental dataset contains six types of radar jamming signals, each consisting of 8001x2-dimensional time series, and the approach directly addresses the problem of parameter-dependent feature degradation that weakens conventional handcrafted feature sets.

What are the main jamming model categories?

A systematic review framework groups jamming models into deception jamming, suppression jamming, and smart jamming, explaining their operational mechanisms. Recognition methods are then split into traditional machine learning approaches and neural network approaches, including both shallow and deep networks, which helps explain why feature extraction choices differ so much across studies.