Labeling Conventions in the RFML Field

Warning

Pre-implementation. The architectural-implications section of this page describes proposed conventions and planned extensions to the label layer. Class signatures, parameter types, and ABC names are subject to change before code lands. Once implementation exists, content here will be reconciled with generated API docs and running tests.

Why This Page Exists

rfgen needs a labeling architecture that supports current canonical formats and can absorb new research-driven modalities without breaking existing consumers. This page surveys what the RFML and spectrum-monitoring communities actually use, with individual citations, so that the BaseLabeler hierarchy and Record schema are informed by evidence rather than assumption. New modalities (instance segmentation, anchor-free bins, foundation-model targets, track-level pulse trains) must be able to plug in as additional BaseLabeler subclasses without disrupting the bbox, segmentation, and per-emitter-metadata paths that ship in the first release.

Summary

The RFML-detection community has converged on a two-format canonical for wideband scene labeling: per-signal time-frequency bounding boxes (using TorchSig’s SignalMetadataObject vocabulary or SigMF core:annotation fields) combined with per-emitter metadata rows. The WRIST paper demonstrated 90 mAP on ~1M emissions using this scheme, and TorchSig’s WidebandSig53 ships the same pairing. SigMF’s unlimited-overlapping-annotations model and omittable sample_count field make it the interchange format of record.

A parallel segmentation-based branch has grown since 2021. DeepSig’s West/Roy/O’Shea work applied U-Net to per-bin spectral segmentation, deriving bboxes from connected components. Northeastern’s Uvaydov 2024 paper pushed this to I/Q-level semantic segmentation at 96.7% mIoU over five protocols and explicitly critiques bbox-only granularity for dense and overlapping scenes. These approaches address a genuine failure mode of bboxes: two co-channel emitters with overlapping time-frequency extents cannot be unambiguously separated by rectangle-based labels.

The spectrum-monitoring community (NTIA, ITU) works with occupancy ratios rather than per-signal localization. NTIA TR-20-548 defines occupancy as per-(time, frequency)-cell threshold exceedance, aggregated to statistical ratios. These labels are structurally different from detection labels and require their own storage and evaluation conventions.

Anchor-free and per-bin attribute prediction represents a third direction. Li et al. (IEEE SPL 2022) predict center-frequency and shape attributes per frequency bin without anchor boxes, citing anchor mismatch and limited receptive fields as failure modes for slender HF signals. Recent RF foundation-model work (SpectrumFM 2025, RIS-MAE 2025, Zheng et al. 2026) has converged on masked reconstruction and contrastive multi-view pairs over IQ sequences as pretraining targets, using no spatial-box labels at all. Time-evolving signals such as radar pulse trains (Mottier et al. 2023) are labeled with per-pulse Pulse Descriptor Words (PDWs) carrying emitter IDs rather than bounding boxes. Military and EW radar work uses segmentation masks for interleaved pulse activity (Huang et al. 2023/2024) or flat/multi-class threat-intent labels.

The field has not converged on a single standard. Bbox, segmentation, occupancy, PDW-track, and SSL pretraining targets coexist, each optimized for a different downstream task. rfgen’s architecture must accommodate all of them.

Canonical Formats

Bounding Boxes (per-signal time-frequency rectangles)

TorchSig SignalMetadataObject defines per-signal fields: start_in_samples, stop_in_samples, duration_in_samples, center_freq, bandwidth, lower_freq, upper_freq, class_name [1]. SigMF core schema annotations carry the equivalent fields: core:sample_start, core:sample_count, core:freq_lower_edge, core:freq_upper_edge [2]. SigMF supports unlimited overlapping annotations and treats a missing sample_count as a continuous (open-ended) emitter, which handles always-on signals without hacks.

WRIST (arXiv:2107.05114) trained a YOLO/DETR-style detector on spectrograms, framing RF detection as a computer-vision object-detection problem [3]. On the SPREAD dataset (~1M emissions) it achieved 90 mAP, establishing bbox conventions as workable for moderate-density comms scenes.

Failure modes. Bbox labels break down in three scenarios:

  1. Dense overlap. Two co-channel emitters have intersecting time-frequency rectangles. A single-label segmentation mask is ambiguous; the bbox ground-truth is inherently multi-valued.

  2. Continuous emitters. A signal that never stops has no stop_in_samples; SigMF’s omitted-sample_count convention handles this, but most detection pipelines assume finite boxes.

  3. Slender HF signals. Li et al. (IEEE SPL 2022, doc 9788039) show that anchor mismatch and limited receptive field cause bbox detectors to miss narrow-bandwidth high-frequency signals [4].

Per-Cell Semantic Segmentation

DeepSig’s “Wideband Signal Localization with Spectral Segmentation” (West/Roy/O’Shea 2021, arXiv:2110.00583) applied U-Net to per-frequency-bin spectral segmentation of multi-emitter scenes [5]. Bboxes are recovered via connected components on the output mask. This sidesteps anchor design entirely and handles overlap via multi-label per-cell classification.

Northeastern’s “Stitching the Spectrum” (Uvaydov 2024, arXiv:2402.03465) extended segmentation to the raw I/Q domain, achieving 96.7% mIoU over five protocols [6]. The paper explicitly critiques bbox-only granularity: rectangles cannot represent irregular frequency occupancy (e.g., LTE guard bands, OFDM null subcarriers) or two overlapping emitters at the same time-frequency cell.

Segmentation label cost is higher: each record requires a full time-frequency mask rather than a short list of rectangle coordinates. For dense wideband scenes, this is the right tradeoff.

Per-Emitter Metadata

Per-emitter metadata rows carry classification labels and physical attributes alongside or separately from spatial labels. TorchSig’s SignalMetadataObject includes class_name, SNR, and frequency parameters as a structured record per emitter [1]. rfgen extends this to include channel profile, device identity, CFO, SFO, IQ imbalance, and provenance fields.

Multi-task radar characterization work (Huang et al. 2023, arXiv:2306.13105) demonstrated a framework with simultaneous classification and regression heads, combining categorical signal class labels with continuous-valued physical parameters [7]. This is the natural multi-task extension of per-emitter metadata: a single emitter record carries both its class label and its measured attribute values.

Spectrum Occupancy Masks

NTIA TR-20-548 (Sanders et al.) defines occupancy as per-(time, frequency)-cell binary threshold exceedance, aggregated to statistical occupancy ratios over the observation window [8]. This is not detection labeling. The unit of analysis is a (time, frequency) cell rather than a named signal. Cells do not carry emitter identity or class; they carry an occupancy ratio (fraction of time the cell exceeds the threshold).

NTIA uses SigMF plus the sigmf-ns-ntia extension namespace to store occupancy metadata alongside captures. This format is appropriate for spectrum management, regulatory monitoring, and database-driven channel allocation. It is not directly compatible with RFML detection metrics (mAP requires emitter identity) but maps cleanly to occupancy-curve evaluation.

Anchor-Free Per-Bin Attribute Prediction

Li et al. (IEEE SPL 2022, doc 9788039) propose an anchor-free framework that predicts center-frequency and shape attributes independently for each frequency bin, avoiding the anchor-mismatch problem that degrades bbox detectors on slender HF signals [4]. The label format is a per-bin attribute vector rather than a list of box coordinates.

This direction is relevant when the signal catalog includes narrow-bandwidth emitters (HF/MF communications, FHSS hops at a single dwell) where a rectangle’s aspect ratio is too extreme for anchor-based detectors to learn reliably.

Track-Level Labeling

Time-evolving signals require labels that span multiple time steps. Radar pulse trains, FHSS frequency hops, ADS-B cadence, and mobile transmitters all move in time-frequency space across a recording.

Mottier et al. (arXiv:2312.11178, 2023) label interleaved radar pulse trains using per-pulse Pulse Descriptor Words (PDWs) [9]. Each PDW carries: frequency, pulse width, signal level, and time of arrival. Each pulse also carries an emitter ID (true emitter assignment), forming a sequence of (PDW, emitter-ID) pairs over the recording. Evaluation uses Adjusted Rand Index over the clustering of pulses by emitter rather than mAP over spatial boxes. This is structurally a track-labeling problem: the ground truth is a set of per-emitter pulse sequences rather than independent spatial labels.

WidebandSig53 (Boegner et al. 2022, arXiv:2211.10335) labels each signal’s time and frequency extent plus its modulation class, which accommodates time-varying signals as a sequence of per-hop bbox annotations if the signal changes center frequency between hops [10]. However, no explicit track-ID linking consecutive hops is defined in the public schema.

Huang et al. (arXiv:2312.09489, 2023/2024) use segmentation masks to label interleaved radar pulse activity over extended time horizons in an EW context [11]. Per-channel masks capture pulse boundaries without requiring per-pulse identity, at the cost of losing emitter-level ground truth within overlapping pulses.

Gap. No public RF dataset ships an explicit track-level label format (emitter-ID + hop-sequence + dwell metadata) for FHSS or mobile transmitters in a standardized schema. PDW-style per-pulse records from radar deinterleaving are the closest analogue.

Hierarchical Taxonomy

Multi-level signal class taxonomies appear in both modulation-recognition and open-set identification literature.

G-AMC (Yu et al. 2026, arXiv:2604.06402) partitions the RadioML 2018.01A 24-class label space into a two-level tree [12]. Level 1 has four coarse classes (amplitude-related, frequency, phase-related, amplitude+phase mixed). Level 2 provides three refinement classifiers (R0 for amplitude, R1 for phase, R2 for QAM/APSK). The leaf classes include BPSK, QPSK, PSK8/16/32, QAM16/32/64/128/256, APSK16/32/64/128, OOK, ASK4/8, FM, AM variants, and GMSK. The tree is organized by modulation property rather than by protocol or standard.

The rfgen design-decisions page documents the same motivation: Sig53, RadioML, DroneRF, WiSig, and NIST CBRS all use different label vocabularies [see design-decisions.md]. The framework uses hierarchical tags (Qwen-Audio pattern) to unify them without lossy mapping to a single flat label space.

Open-set RF identification papers (e.g., Karunaratne et al. 2021, arXiv:2108.13099; Liu et al. 2026, arXiv:2603.24268) use two- or three-tier schemes: known authorized, known unauthorized, and unknown/rejected [13, 14]. Some add an incremental-discovery layer that promotes clusters of unknowns into new named classes over time. These are not hierarchical taxonomies in the signal-property sense but represent a different kind of nested label structure based on authorization status.

Foundation-Model Formats

Recent RF SSL and foundation-model papers have converged on two pretraining target formats, neither of which is a spatial label.

Masked reconstruction. SpectrumFM (Zhou et al. 2025, arXiv:2505.06256) pretrains on masked reconstruction of IQ sequences combined with next-slot signal prediction [15]. The target is the reconstructed IQ segment, not a class label or spatial coordinate. Downstream fine-tuning (AMC, WTC, spectrum sensing, anomaly detection) uses standard classification or binary detection heads. RIS-MAE (Liu et al. 2025, arXiv:2508.00274) applies the same masked-autoencoder idea to raw IQ, recovering amplitude and phase features from masked patches [16].

Contrastive multi-view pairs. Zheng et al. (arXiv:2604.10906, 2026) construct contrastive pairs from four information-lossless equivalent transformations of the same signal (time domain, instantaneous domain, frequency domain, time-frequency domain) [17]. The pretraining target is alignment of representations across these four views; no external class label is required. A unified foundation model for technology recognition and localization (Cheraghinia et al. 2025, arXiv:2505.19390) uses IQ + Channel Impulse Response timeseries with input patching for masked SSL pretraining, then fine-tunes on technology classification, LOS detection, and ranging correction [18].

These formats require no bbox, segmentation, or per-emitter-metadata labels during pretraining. The record still carries IQ and physical metadata, but the labeler for SSL pretraining produces token masks or augmented-view pairs rather than spatial annotations.

Spectrum-Monitoring vs RFML-Detection Branches

The spectrum-monitoring community and the RFML-detection community share the same raw material (RF captures) but use structurally different ground-truth conventions.

RFML-detection. Labels are per-emitter and per-signal: bboxes, segmentation masks, or metadata rows identify individual transmissions. Evaluation is detection-oriented: mAP, mIoU, classification accuracy. Sources: TorchSig, SigMF, WRIST, DeepSig, Northeastern.

Spectrum monitoring. Labels are per-(time, frequency)-cell: occupancy ratios from threshold exceedance, with no emitter identity. Evaluation is occupancy-curve accuracy and false-alarm rate. Sources: NTIA TR-20-548, ITU spectrum monitoring guidelines, sigmf-ns-ntia. The focus is regulatory compliance and interference management, not emitter identification.

rfgen’s JointLabeler is designed for the detection branch. Occupancy-mask labeling requires a separate path that produces NTIA-style per-cell arrays, not per-emitter box lists. Both paths can coexist in the same Record if stored under separate label modality keys.

Evaluation Metrics

Label format

Primary evaluation metric

Bounding boxes

mAP (mean Average Precision) at IoU thresholds

Segmentation masks

mIoU (mean Intersection over Union)

Spectrum occupancy

Occupancy ratio error, ROC / false-alarm rate

Per-emitter metadata

Per-attribute accuracy or RMSE (SNR, CFO, class accuracy)

Hierarchical taxonomy

Top-k accuracy at each tree level; hierarchical precision/recall

Track / PDW labeling

Adjusted Rand Index (clustering); per-emitter precision/recall

SSL / foundation-model targets

NMI, contrastive accuracy, downstream task accuracy after fine-tuning

Anchor-free per-bin

Per-bin attribute accuracy; AP on derived detections

Field Disagreements and Gaps

Bbox vs. hybrid. The RFML community is split between bbox-only (dominant in detection benchmarks such as WRIST, WidebandSig53) and hybrid segmentation+bbox approaches (DeepSig 2021, Northeastern 2024). The segmentation branch argues bboxes cannot represent dense or irregular occupancy; the detection branch argues segmentation labels are expensive to compute at scale and hard to use with standard detection heads.

No standard track schema. There is no public, standardized track-level label format for FHSS or mobile transmitters. PDW-style per-pulse records from radar deinterleaving (Mottier et al. 2023) represent the closest analogue but are radar-specific. SigMF does not define a track-linkage field. This is a genuine gap.

Foundation models predate the need for spatial labels. Current RF SSL papers do not produce spatial annotations and are not yet integrated with detection pipelines. The connection between masked-IQ pretraining and downstream bbox fine-tuning has not been standardized.

Military / EW coverage is thin in the open literature. Most operational EW signal classification and jamming datasets are classified. Public work uses flat multi-class threat-intent labels (binary jammed/normal, or three-class station-keeping/shadowing/jamming) rather than fine-grained EW signal taxonomies. Radar pulse segmentation for EW context (Huang et al. 2023) is the most detailed public example, but the specific class taxonomy for hostile radar modes is not published.

Hierarchical taxonomy fragmentation. Each public dataset defines its own label vocabulary. G-AMC’s property-based hierarchy (amplitude/phase/frequency/mixed) differs from protocol-based hierarchies (comms.cellular.5g.nr.qpsk). No community standard maps across them.

Implications for rfgen’s Architecture

  • Multi-modal labels in one Record. A single record should be able to carry bbox, segmentation mask, per-emitter metadata, occupancy mask, PDW-track sequence, and SSL pretraining targets simultaneously, stored under separate modality keys. Downstream consumers read the modalities they need.

  • Pluggable label modalities via BaseLabeler subclasses. The bbox, segmentation, and per-emitter paths are the initial implementations. Track, occupancy, anchor-free, and SSL-target modalities should plug in as additional subclasses without modifying JointLabeler’s core contract.

  • Cross-modality consistency invariants. The current four invariants in concepts/labels.md (bbox/segmentation/metadata share the same component list in the same order; labels are derived from component metadata, not from IQ inference; coordinates are absolute; re-running with the same inputs yields identical labels) should be extended when new modalities land. PDW-track labels must respect the same component ordering invariant. SSL targets must be deterministic given a fixed random seed and masking schedule.

  • Coordinate system for track labels. Per-pulse PDW records need a time axis (absolute sample offset or relative-to-window-start) and a frequency axis (Hz, not normalized). Consistency with the existing absolute-coordinate convention in concepts/labels.md is required.

  • Recommendation: add BaseTrackLabeler. A per-slot ABC for time-evolving signals (radar pulse trains, FHSS hops, ADS-B cadence) should be planned for when the emitter library includes FHSS or radar emitters. The labeler would produce a sequence of (time_of_arrival, frequency, pulse_width, emitter_id) records per recording window.

  • Recommendation: add BaseFoundationTargetLabeler. Masked-modeling and contrastive-pair targets are structurally separate from detection labels. A dedicated ABC for SSL pretraining target generation allows the same IQ record to produce both detection labels (for supervised training) and SSL targets (for pretraining or multi-task learning) without coupling the two.

  • Hierarchical class field. Per-emitter metadata rows should carry a hierarchical class tag (e.g., comms.cellular.lte.qpsk) not just a flat class name. The G-AMC and Qwen-Audio patterns confirm two levels (coarse group + fine class) is the minimum useful granularity. rfgen should define a canonical tag vocabulary that covers Sig53, RadioML, and protocol-specific classes.

  • Occupancy-mask path as a separate labeler. NTIA-style occupancy labels are not per-emitter and cannot share the JointLabeler code path without a awkward per-cell vs. per-signal mismatch. A SpectrumOccupancyLabeler is the clean solution, writing a separate Zarr array under record["occupancy"].

  • Evaluation metric infrastructure. The diversity of metrics (mAP, mIoU, ARI, NMI, occupancy-ratio error) requires that each BaseLabeler subclass declares which metric applies, so that the validation harness can dispatch correctly.

References

[1] TorchDSP. TorchSig: SignalMetadataObject and dataset toolkit. GitHub. https://github.com/TorchDSP/torchsig

[2] SigMF Contributors. Signal Metadata Format (SigMF) Specification. sigmf.org. https://sigmf.org/sigmf-spec.pdf and https://github.com/sigmf/SigMF

[3] Hanna et al. WRIST: Wideband RF Inference and Signal Type detection (arXiv:2107.05114). arXiv, 2021. https://arxiv.org/abs/2107.05114

[4] Li et al. Anchor-free framework for per-frequency-bin attribute prediction. IEEE Signal Processing Letters, 2022, doc 9788039.

[5] West, Roy, O’Shea. Wideband Signal Localization with Spectral Segmentation (arXiv:2110.00583). arXiv, 2021. https://arxiv.org/abs/2110.00583

[6] Uvaydov. Stitching the Spectrum (arXiv:2402.03465). arXiv, 2024. https://arxiv.org/abs/2402.03465

[7] Huang, Pemasiri, Denman, Fookes, Martin. Multi-task Learning for Radar Signal Characterisation (arXiv:2306.13105). arXiv, 2023. https://arxiv.org/abs/2306.13105

[8] Sanders et al. NTIA Technical Report TR-20-548: Spectrum Occupancy Measurements. NTIA, 2020. (SigMF + sigmf-ns-ntia extension used for storage.)

[9] Mottier, Chardon, Pascal. Deinterleaving RADAR emitters with optimal transport distances (arXiv:2312.11178). arXiv, 2023. https://arxiv.org/abs/2312.11178

[10] Boegner et al. Large Scale Radio Frequency Wideband Signal Detection & Recognition (arXiv:2211.10335). arXiv, 2022. https://arxiv.org/abs/2211.10335

[11] Huang, Pemasiri, Denman, Fookes, Martin. Multi-stage Learning for Radar Pulse Activity Segmentation (arXiv:2312.09489). arXiv, 2023/2024. https://arxiv.org/abs/2312.09489

[12] Yu, Chen, Kuo. G-AMC: A Green Automatic Modulation Classification Method (arXiv:2604.06402). arXiv, 2026. https://arxiv.org/abs/2604.06402

[13] Karunaratne, Hanna, Cabric. Open Set RF Fingerprinting using Generative Outlier Augmentation (arXiv:2108.13099). arXiv, 2021. https://arxiv.org/abs/2108.13099

[14] Liu et al. Incremental Learning-Based Open-Set Classification of Unknown UAVs via RF Signal Semantics (arXiv:2603.24268). arXiv, 2026. https://arxiv.org/abs/2603.24268

[15] Zhou et al. SpectrumFM: A Foundation Model for Intelligent Spectrum Management (arXiv:2505.06256). arXiv, 2025. https://arxiv.org/abs/2505.06256

[16] Liu et al. RIS-MAE: A Self-Supervised Modulation Classification Method Based on Raw IQ Signals and Masked Autoencoder (arXiv:2508.00274). arXiv, 2025. https://arxiv.org/abs/2508.00274

[17] Zheng et al. Unsupervised Equivalent Contrastive Learning for Radio Signal Recognition (arXiv:2604.10906). arXiv, 2026. https://arxiv.org/abs/2604.10906

[18] Cheraghinia et al. A Unified Foundation Model for Wireless Technology Recognition and Localization (arXiv:2505.19390). arXiv, 2025. https://arxiv.org/abs/2505.19390

[19] Huang, Denman, Pemasiri, Martin, Fookes. RadDet: A Wideband Dataset for Real-Time Radar Spectrum Detection (arXiv:2501.10407). arXiv, 2025. https://arxiv.org/abs/2501.10407

[20] Tandiya et al. Deep Predictive Coding Neural Network for RF Anomaly Detection (arXiv:1803.06054). arXiv, 2018. https://arxiv.org/abs/1803.06054

[21] Boumeftah, Kurt. Leveraging Orbital Dynamics with RF Signal Features for Satellite Multi-Orbit Proximity Threat Detection (arXiv:2509.01802). arXiv, 2025. https://arxiv.org/abs/2509.01802

Verification Methodology

This survey was assembled from two prior parallel-agent research workflows, which established the canonical findings on TorchSig SignalMetadataObject, SigMF core annotations, WRIST bbox conventions, the DeepSig 2021 segmentation paper, the Northeastern 2024 I/Q segmentation paper, Li et al. 2022 anchor-free work, and NTIA TR-20-548 spectrum occupancy labeling. A final sequential gap-filling research pass covered four areas not addressed by prior runs: (1) RF foundation models and SSL pretraining targets, sourced from arXiv queries on self-supervised wireless learning (SpectrumFM, RIS-MAE, Zheng et al., Cheraghinia et al.); (2) time-evolving signal track labeling, sourced from radar deinterleaving literature (Mottier et al. 2023) and wideband radar datasets (RadDet, WidebandSig53); (3) hierarchical modulation taxonomy, sourced from G-AMC and open-set RF identification papers; and (4) military and EW signal datasets, sourced from public EW radar and jamming-detection literature (Huang et al. 2023/2024, RadDet, Tandiya et al., Boumeftah et al.). All claims are individually cited. Sources that returned 403/404 or whose full text did not confirm a claim were dropped rather than assumed.

See Also