Multi-Domain Features Guided Supervised Contrastive Learning for Radar Target Detection

Dongying Li; Junjie Wang; Wenxian Yu; Yuze Gao

REVIEW 4 major objections 5 minor 11 references

Multi-Domain Features Guided Supervised Contrastive Learning for Radar Target Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that guiding supervised contrastive learning with six weighted multi-domain radar features improves small-target detection in sea clutter, cutting realized false alarms below mainstream contrastive detectors and improving…

desk verdict Sensible method, sloppy evaluation: unmatched false alarm rates across methods make the reported mIoU advantage unreliable until re-evaluated at common operating points. read the letter →

arxiv 2412.12620 v2 pith:TDVER42O submitted 2024-12-17 cs.CV

Junjie Wang , Yuze Gao , Dongying Li , Wenxian Yu This is my paper · ORCID

classification cs.CV

keywords radartargetdetectionseacluttersupervisedcontrastivelearningmulti-domainfeaturesGinicoefficientsmallIPIXdatasetmatchingloss

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that radar target detection in sea clutter is improved by guiding supervised contrastive learning with statistical features drawn from multiple domains. The authors propose MDFG_SCL, which computes six shallow echo features in the time, frequency, and time-frequency domains, weights them by decision-tree Gini impurity, and aligns deep encoder representations to them through a matching loss during pre-training. On real IPIX 1993 data, the method reports higher detection mIoU at a preset false alarm rate of $P_{fa}=0.001$ and an average realized false alarm rate of 0.0010, below the compared unsupervised and supervised contrastive detectors. The practical point is that combining low-level domain-specific cues with high-level semantic features may fix the generalization failure of deep detectors across changing sea conditions.

What carries the argument

The load-bearing mechanism is the matching loss $L_{align}$, which projects deep features and weighted shallow features into a common normalized space and applies a symmetric cross-entropy over their similarity matrix to pull deep representations toward the shallow ones. The shallow feature vector $F^{shallow}$ concatenates six statistical descriptors from three domains: relative average amplitude, relative Doppler peak height, relative Doppler vector entropy, ridge integral, and the number and maximum size of connected regions, each normalized and weighted by its Gini-impurity contribution to clutter-target separation. During pre-training the total loss is the supervised contrastive loss plus $\alpha L_{align}$, optimized over a 1D-ResNet50 encoder and a two-layer projection head; during fine-tuning the encoder is frozen and only the projection head and classifier are trained.

What would settle it

Replace the six computed shallow features with six random noise vectors of the same shape during pre-training, keeping all other hyperparameters and the same matching loss, then measure mIoU at $P_{fa}=0.001$ on the IPIX 1993 datasets; if performance stays near the reported 97.42%, the shallow-feature content is not the active ingredient.

Watch

Extended reading notes

Core claim

The paper's central claim is that weighted multi-domain shallow features can reliably guide deep supervised contrastive representation learning, and that the resulting MDFG_SCL detector separates small targets from sea clutter better than existing shallow-feature detectors, unsupervised contrastive learning, and plain supervised contrastive learning. The supporting experiments on the IPIX 1993 dataset show MDFG_SCL achieving the best mIoU across HH, HV, VH, and VV polarizations, reaching 97.42% mIoU in the HH ablation, and realizing an average true false alarm rate of 0.0010 at the preset $P_{fa}=0.001$; the supervised contrastive baseline without shallow guidance realizes 0.0033, and the unsupervised contrastive baseline 0.0765. The paper also reports that after pretraining on six datasets and testing on four unseen ones, the shallow-guided method detects nearly all target samples at $P_{fa}=0.01$, while supervised learning and plain supervised contrastive learning generalize poorly.

Load-bearing premise

The claimed advantage rests on the six hand-crafted multi-domain features remaining informative enough to separate sea clutter from targets, so that aligning deep features to them improves generalization; if any sea state makes these features uninformative, the method's edge over plain supervised contrastive learning disappears.

Editorial extensions

If this is right

If the central claim holds, radar detectors can operate at a preset false alarm probability while realizing far fewer false alarms than supervised contrastive learning alone: the reported average true $P_{fa}$ drops from 0.0033 to 0.0010.
Shallow physical features can serve as knowledge-driven regularization for representation learning, meaning deep radar detectors do not have to rely solely on labels to separate clutter from targets.
The alignment coefficient $\alpha$ has a working range between 0.1 and 0.3; beyond 0.5 the shallow guidance hurts performance, so the method's benefit depends on calibrating how strongly physical features override deep features.
Cross-dataset pretraining with shallow guidance transfers to unseen sea states at $P_{fa}=0.01$, which would make the detector usable in new maritime environments without per-scene retraining.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An implication the authors leave implicit is that the Gini weighting scheme should be recomputed per sea state; a direct testable extension is to measure the method's margin over plain supervised contrastive learning as the six Gini weights flatten.
Because the matching loss pushes deep features toward the shallow set, a failure of any of the six physical features would propagate into the learned representation; a learned or adaptive shallow-feature set could make the guidance more robust in regimes the paper does not test.
The cross-dataset result suggests a deployment recipe the paper does not state: freeze the shallow-guided pretrained encoder and adapt only the projection head and classifier to a new environment, avoiding full retraining while keeping the false alarm advantage.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: This is a plausible incremental method — Gini-weighted multi-domain shallow features plus supervised contrastive learning, with a matching loss to align them. The idea is coherent and the ablations indicate each component helps. The empirical comparison, however, is undercut by unmatched false alarm rates across methods, so the headline mIoU advantage is not reliably established.

What's actually new: The specific combination isn't in the cited literature, and the matching loss (Eq. 4) is a reasonable mechanism to inject shallow physical insight into deep representation learning. The Gini weighting is simple and interpretable. The paper does a decent job of ablation and uses a public dataset (IPIX1993). It also shows the proposed method achieves the target false alarm rate, whereas competitors run at looser thresholds — that's a legitimate practical point in its favor.

The soft spots: The big one is Table III. True false alarm rates range from 0.0010 (ours) to 0.1676 (TF-Tri-Feature). So Fig. 2 at nominal Pf_a=0.001 is not a matched operating point comparison. Competitors operating at 7–76x higher false alarm rates likely get inflated recall, and we can't tell from the paper whether mIoU differences are real or threshold effects. This needs fixing with matched thresholds or ROC-style reporting. Also, there are no error bars or multiple runs, so significance is unknown. Finally, the six shallow features (RAA, RDPH, etc.) are never defined with equations or references; for a methods paper that's a reproducibility barrier.

Who it's for: Radar target detection community, specifically people using contrastive learning on sea clutter. The idea is worth knowing about, but the evidence needs to be cleaned up.

Recommendation: Send to peer review, but expect major revision. The evaluation protocol must be corrected, feature definitions added, and ideally code released. If that happens, the paper could be a solid, citable contribution.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes MDFG_SCL, a radar target detection method that combines six hand-crafted multi-domain shallow features (RAA, RDPH, RVE, RI, NR, MS) with a 1D-ResNet50 encoder trained by supervised contrastive learning. A Gini-importance weighting is applied to the shallow features, and a matching loss aligns deep and shallow representations in a common space. The method is evaluated on ten IPIX1993 datasets in four polarization channels, with comparisons against traditional feature-based detectors, supervised learning, unsupervised contrastive learning (RAVA-CL), and supervised contrastive learning. The paper reports higher mIoU at a nominal false-alarm probability Pf_a=0.001, ablation results over the matching-loss coefficient, and qualitative generalization experiments.

Significance. If the performance claims survive a properly matched operating-point evaluation, the idea of guiding supervised contrastive representation learning with physically motivated multi-domain features is a useful contribution to radar small-target detection. The paper uses a real-world benchmark dataset, includes several baselines, and tests sensitivity to the matching-loss coefficient. However, the central quantitative claim is currently compromised by unmatched true false-alarm rates across methods, and the method is not fully reproducible because the six shallow features are not defined. The core idea is plausible and the paper is within the scope of the journal, but the experimental evidence needs substantial revision.

major comments (4)

[Table III and Fig. 2] The central comparison of detection performance is not made at a common operating point. Fig. 2 reports mIoU at a nominal Pf_a=0.001, but Table III shows that the average true false-alarm rates differ by up to two orders of magnitude across methods: Ours=0.0010, Supervised Contrastive Learning [11]=0.0033, Supervised Learning=0.0047, MDCCNN [10]=0.0073, RAVA-CL [6]=0.0765, Tri-Feature [1]=0.1155, and TF-Tri-Feature [2]=0.1676. A detector evaluated at a higher true false-alarm rate can report higher mIoU simply because its decision threshold is more permissive. The claim that MDFG_SCL outperforms the baselines is therefore not established. Please report ROC curves, or evaluate all methods at matched true Pf_a values, or provide per-method mIoU at several Pf_a levels so the operating point is controlled.
[Section II-B.3, Eq. (5)] Equation (5) is not the standard supervised contrastive loss of [11]. In the denominator, the summation is over the A negative samples only, while the standard SupCon loss sums over all samples except the anchor, including the positive samples. Since the text explicitly defines A as the number of negative samples, this is not merely a typographical ambiguity. As written, the loss changes the gradient and the claimed equivalence to supervised contrastive learning is inaccurate. Please correct the equation or clarify the intended modified loss and justify its use.
[Section II-A] The six multi-domain features RAA, RDPH, RVE, RI, NR, and MS are named but never defined by equations or citations. These features enter directly into the matching loss in Eq. (4) and are the basis of the claimed improvement in Table IV, so the method cannot be reproduced or independently tested without their definitions. Please provide formulas or precise references for each feature, and specify how the Gini-weighted shallow feature vector f_s is formed from the six individual features.
[Table IV and Section III-B] The ablation claim that "only the proposed method meets the specified Pf_a" is not supported by Table IV, which reports only Accuracy, Recall, and mIoU and contains no true false-alarm column. Since thresholds are calibrated using pure clutter validation samples, the table should report the true Pf_a on test data for every ablation row, together with the threshold-setting procedure. Without this, the 1-2% improvements attributed to the weighted shallow features may reflect differences in operating point rather than detection ability.

minor comments (5)

[Section II-B.2, Eq. (2)] The dimensions of the input, hidden layers, and output of the projection head are not specified; please provide them for reproducibility.
[Eq. (5)] The notation z_i z_p and z_i z_a should use an explicit inner product or dot product symbol, as the current notation is ambiguous.
[Table II] The mIoU formula in Table II appears to include TN in the denominator of the target IoU term (TP/(TP+FN+FP+TN)); if this is the case, it is not the standard per-class IoU and should be corrected.
[References] Reference [9] is listed as "I. Radar, The McMaster IPIX radar sea clutter database"; please use the actual institutional citation for the IPIX dataset.
[Section II-B.1] The heading contains a typo: "Data Augmetation" should be "Data Augmentation."

Circularity Check

0 steps flagged · score 0.0 of 10

No self-referential or definitional circularity: the training objective optimizes deep representations toward labels and hand-crafted features, while the reported detection gains are evaluated on held-out test data and external baselines.

full rationale

The chain from method to result is a supervised training pipeline rather than a derivation that encodes its own output. The total loss Ltotal = Lsup_cl + alpha*Lalign (Eq. 6) uses labels for supervised contrastive learning (Eq. 5) and uses weighted shallow features for the alignment loss (Eq. 4), but these terms only define the training objective. The reported mIoU and false-alarm rates come from applying the trained detector to test segments and from thresholds calibrated on a pure-clutter validation set; no equation in the paper maps the loss or the Gini-based feature weights directly to the reported test metrics, so the empirical claims are not forced by construction. The Gini weighting is fitted to training labels, but it is an internal feature-weighting mechanism, not a parameter that is then rebranded as a prediction of test performance. The paper contains self-citations ([3] and [7] share a J. Wang author name), but these appear only in the introduction as context about prior CNN and semi-supervised work; the proposed method is not justified by appealing to those papers, and no uniqueness claim is imported from them. The main legitimate concern is experimental fairness: Table III reports unequal average true Pf_a values across methods (e.g., 0.0010 for the proposed method versus 0.0033 for supervised contrastive learning), so the comparison in Fig. 2 may not be at matched operating points. That is a correctness or evaluation-protocol issue, not a circularity issue, because the detector outputs are not derived from the competitors' operating points. Overall, the paper's central claim is an empirical comparison against external methods and datasets, and there is no load-bearing self-citation or definitional reduction of the predicted quantity to its inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities or forces. It relies on existing dataset, features, architecture, and loss functions. The main free parameters are the matching loss coefficient alpha and the temperature t, while the Gini weights are computed from the training data rather than fitted as independent parameters. The axiomatic burden is the assumed informativeness of the six shallow features and the representativeness of the IPIX1993 dataset.

free parameters (3)

alpha (matching loss coefficient) = 0.1
Tuned via experiments over {0, 0.1, 0.2, 0.3, 0.4, 0.7, 1}; the claimed benefit of shallow feature guidance depends on this choice.
temperature coefficient t = 0.07
Hyperparameter for supervised contrastive loss, set following standard practice; not fitted to the target metric but affects the representation.
observation time / segment length = 0.512 s / 512 segments per cell
The letter segments radar echoes into 0.512s windows and 512 segments per cell; this choice fixes the temporal support of all features and the comparison baseline.

assumptions (4)

domain assumption IPIX1993 dataset is representative of varied sea conditions and sufficient to demonstrate generalization.
All experiments use the 1993 IPIX dataset from McMaster University; no validation on other radar systems or newer sea clutter data supports the claimed generalization.
domain assumption The six multi-domain features are informative and correctly computed.
The paper invokes RAA, RDPH, RVE, RI, NR, MS without defining them or citing precise formulas; the method assumes they separate clutter from targets and that Gini weighting improves their use.
domain assumption Data augmentation methods from [6] preserve class labels and improve invariance.
RCRS-DAM, AD-DAM, and F-DAM are taken from prior work on radar echoes; the authors do not re-validate that augmented samples retain target or clutter identity under the supervised contrastive loss.
domain assumption 1D-ResNet50 and supervised contrastive loss provide useful representations for this task.
The encoder and loss are standard choices borrowed from image and radar literature; no evidence is given that this architecture is optimal or that the supervised contrastive loss behaves well on small radar datasets.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-Domain Features Guided Supervised Contrastive Learning for Radar Target Detection." pith.science (2026). https://pith.science/paper/TDVER42O

@misc{pith2026241212620,
  author       = {Pith},
  title        = {Pith review of: Multi-Domain Features Guided Supervised Contrastive Learning for Radar Target Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TDVER42O}},
  note         = {Machine review of arXiv:2412.12620}
}

read the original abstract

Detecting small targets in sea clutter is challenging due to dynamic maritime conditions. Existing solutions either model sea clutter for detection or extract target features based on clutter-target echo differences, including statistical and deep features. While more common, the latter often excels in controlled scenarios but struggles with robust detection and generalization in diverse environments, limiting practical use. In this letter, we propose a multi-domain features guided supervised contrastive learning (MDFG_SCL) method, which integrates statistical features derived from multi-domain differences with deep features obtained through supervised contrastive learning, thereby capturing both low-level domain-specific variations and high-level semantic information. This comprehensive feature integration enables the model to effectively distinguish between small targets and sea clutter, even under challenging conditions. Experiments conducted on real-world datasets demonstrate that the proposed shallow-to-deep detector not only achieves effective identification of small maritime targets but also maintains superior detection performance across varying sea conditions, outperforming the mainstream unsupervised contrastive learning and supervised contrastive learning methods.

Figures

Figures reproduced from arXiv: 2412.12620 by the authors.

**Figure 1.** Supervised contrast learning framework based on multidomain shallow features guidance. A. Multi-domain features extraction Six commonly used distinctive features in sea surface target detection are selected, which include RAA in the time domain, RDPH and RVE in the frequency domain, and RI, NR, MS in the time-frequency domain. The six distinctive features mentioned above are used to calculate the multi-domain shal… view at source ↗

**Figure 2.** Detection performance mIoU at Pfa=0.001. As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Evaluation metrics change under different α. combined the first six datasets from IPIX1993 for pretraining, enabling the model to learn radar echo representations across multiple scenes. The pretrained model was then evaluated on classification tasks using the remaining four datasets. Comparative experiments with traditional supervised learning, unsupervised contrastive learning, and supervised contrastive learning … view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Detection visualization results at Pfa=0.01. (a)~(d)Detection results of traditional supervised learning, unsupervised contrastive learning, supervised contrastive learning, and the proposed method on dataset #280. (e)~(h)Detection results on dataset #310. (i)~(l)Detec…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 9 canonical work pages

[11]

Supervised contrastive learn- ing,

P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learn- ing,” Advances in neural information processing systems , vol. 33, pp. 18 661–18 673, 2020

2020
[10]

False-alarm-controllable radar detection for marine target based on multi features fusion via cnns,

X. Chen, N. Su, Y . Huang, and J. Guan, “False-alarm-controllable radar detection for marine target based on multi features fusion via cnns,” IEEE Sensors Journal , vol. 21, no. 7, pp. 9099–9111, 2021

work page 2021
[6]

Target detection in sea clutter via contrastive learning,

S. Xia, Y . Kong, K. Xiong, and G. Cui, “Target detection in sea clutter via contrastive learning,” IEEE Transactions on Instrumentation and Measurement, 2023

work page 2023
[1]

Tri-feature-based detection of floating small targets in sea clutter,

P.-L. Shui, D.-C. Li, and S.-W. Xu, “Tri-feature-based detection of floating small targets in sea clutter,” IEEE Transactions on Aerospace and Electronic Systems , vol. 50, no. 2, pp. 1416–1430, 2014

work page 2014
[2]

Sea-surface floating small target detection by one-class classifier in time-frequency feature space,

S.-N. Shi and P.-L. Shui, “Sea-surface floating small target detection by one-class classifier in time-frequency feature space,” IEEE Transactions on Geoscience and Remote Sensing , vol. 56, no. 11, pp. 6395–6411, 2018

work page 2018
[3]

Maritime radar target detection in sea clutter based on cnn with dual-perspective attention,

J. Wang and S. Li, “Maritime radar target detection in sea clutter based on cnn with dual-perspective attention,” IEEE Geoscience and Remote Sensing Letters , vol. 20, pp. 1–5, 2022

work page 2022
[4]

A pointnet-based cfar detection method for radar target detection in sea clutter,

X. Chen, K. Liu, and Z. Zhang, “A pointnet-based cfar detection method for radar target detection in sea clutter,” IEEE Geoscience and Remote Sensing Letters , 2024

work page 2024
[5]

Radar target detection with multi-task learning in heterogeneous environment,

H. Jing, Y . Cheng, H. Wu, and H. Wang, “Radar target detection with multi-task learning in heterogeneous environment,” IEEE Geoscience and Remote Sensing Letters , vol. 19, pp. 1–5, 2022

work page 2022

Show all 11 references

[7]

Maritime radar target detection model self-evolution based on semi-supervised learning,

J. Wang and S. Li, “Maritime radar target detection model self-evolution based on semi-supervised learning,” IEEE Transactions on Geoscience and Remote Sensing , 2023

2023
[8]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

2016
[9]

The mcmaster ipix radar sea clutter database,

I. Radar, “The mcmaster ipix radar sea clutter database,” [Online], 2021, http://soma.ece.mcmaster.ca/ipix/index.html

2021

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [11]

Supervised contrastive learn- ing,

P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learn- ing,” Advances in neural information processing systems , vol. 33, pp. 18 661–18 673, 2020

2020

[2] [10]

False-alarm-controllable radar detection for marine target based on multi features fusion via cnns,

X. Chen, N. Su, Y . Huang, and J. Guan, “False-alarm-controllable radar detection for marine target based on multi features fusion via cnns,” IEEE Sensors Journal , vol. 21, no. 7, pp. 9099–9111, 2021

work page 2021

[3] [6]

Target detection in sea clutter via contrastive learning,

S. Xia, Y . Kong, K. Xiong, and G. Cui, “Target detection in sea clutter via contrastive learning,” IEEE Transactions on Instrumentation and Measurement, 2023

work page 2023

[4] [1]

Tri-feature-based detection of floating small targets in sea clutter,

P.-L. Shui, D.-C. Li, and S.-W. Xu, “Tri-feature-based detection of floating small targets in sea clutter,” IEEE Transactions on Aerospace and Electronic Systems , vol. 50, no. 2, pp. 1416–1430, 2014

work page 2014

[5] [2]

Sea-surface floating small target detection by one-class classifier in time-frequency feature space,

S.-N. Shi and P.-L. Shui, “Sea-surface floating small target detection by one-class classifier in time-frequency feature space,” IEEE Transactions on Geoscience and Remote Sensing , vol. 56, no. 11, pp. 6395–6411, 2018

work page 2018

[6] [3]

Maritime radar target detection in sea clutter based on cnn with dual-perspective attention,

J. Wang and S. Li, “Maritime radar target detection in sea clutter based on cnn with dual-perspective attention,” IEEE Geoscience and Remote Sensing Letters , vol. 20, pp. 1–5, 2022

work page 2022

[7] [4]

A pointnet-based cfar detection method for radar target detection in sea clutter,

X. Chen, K. Liu, and Z. Zhang, “A pointnet-based cfar detection method for radar target detection in sea clutter,” IEEE Geoscience and Remote Sensing Letters , 2024

work page 2024

[8] [5]

Radar target detection with multi-task learning in heterogeneous environment,

H. Jing, Y . Cheng, H. Wu, and H. Wang, “Radar target detection with multi-task learning in heterogeneous environment,” IEEE Geoscience and Remote Sensing Letters , vol. 19, pp. 1–5, 2022

work page 2022