General Feature Extraction In SAR Target Classification: A Contrastive Learning Approach Across Sensor Types

Chengfang Ren (SONDRA); J. Frontera-Pons; J.-P. Ovarlez; M. Muzeau (SONDRA)

REVIEW 4 major objections 6 minor 24 references

General Feature Extraction In SAR Target Classification: A Contrastive Learning Approach Across Sensor Types

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A contrastive ViT trained only on unlabeled SETHI SAR data clusters MSTAR targets well enough that k-NN reaches 95.9% accuracy with ten labels per class, beating a ResNet-34 trained on MSTAR (52.2%).

desk verdict A fresh cross-sensor SAR transfer result worth checking, but the SETHI/MSTAR overlap assumption and thin few-shot evidence need to be addressed before the 95.9% claim can be trusted. read the letter →

arxiv 2502.01162 v1 pith:LIMYBZUD submitted 2025-02-03 eess.SP

M. Muzeau (SONDRA) , J. Frontera-Pons , Chengfang Ren (SONDRA) , J.-P. Ovarlez This is my paper · ORCID

classification eess.SP

keywords contrastivelearningSARtargetclassificationself-supervisedVisionTransformerfew-shotcross-sensorgeneralizationMSTARSETHI

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

The paper aims to establish that a single self-supervised feature extractor can serve SAR target classification across different sensors, without ever seeing the target dataset or its labels. It trains a Vision Transformer with a contrastive prototype objective on 199,040 unlabeled SETHI patches, then uses k-NN on the extracted features to classify MSTAR vehicle images. With ten labeled images per class, the method reaches 95.9% accuracy, far above a ResNet-34 trained from scratch on MSTAR (52.2%) and a k-NN on PCA features (70.16%). If this holds, unlabeled SAR collections from one sensor could build a reusable encoder for another sensor's scarce labeled data.

What carries the argument

The method is a student-teacher Siamese ViT trained with a prototype-based masked siamese objective (the loss of reference [15]). A student encoder sees global and local crops, random patch masking, subband subsampling, and a random mean shift; a teacher encoder sees a despeckled global crop and is updated by a moving average. Both project an image into a feature vector, then into a representation matched against a set of learned prototypes through a softmaxed cosine similarity, and the student is trained to match the teacher's sharper prototype distribution while an entropy regularizer forces all prototypes to be used. The prototypes act as cluster centers, so the encoder learns to map similar SAR images into the same regions of feature space, and this clustering is what later lets a k-NN classifier on the frozen student features label MSTAR images.

What would settle it

Compute raw-patch or simple-feature nearest-neighbor distances between the 199,040 SETHI training patches and the MSTAR test images; if a nontrivial fraction of MSTAR images have a SETHI patch at near-duplicate distance, the cross-sensor claim collapses into retrieval rather than generalization.

Watch

Extended reading notes

Core claim

The central claim is that a ViT feature extractor trained with a masked siamese, prototype-based contrastive objective on unlabeled SETHI SAR imagery generalizes to MSTAR target classification without fine-tuning. The learned embeddings separate MSTAR classes well enough that a simple k-NN classifier reaches 74% accuracy with one labeled image per class and 95.9% with ten, outperforming a ResNet-34 trained on MSTAR from scratch (52.2%) and a k-NN on PCA-reduced features (70.16%), and reaching 99.1% with 100 labels per class. The authors interpret this as evidence that contrastive pretraining on one sensor captures transferable SAR structure, and that the prototype clustering objective is well matched to target recognition.

Load-bearing premise

The training imagery must contain no test targets or near-duplicates; the paper asserts in its feature-visualization section that the network was not trained on MSTAR data or any military vehicle, but it provides no overlap statistics to verify this.

Editorial extensions

If this is right

With ten labeled images per class, the frozen features give 95.9% MSTAR accuracy, 43.7 points above a ResNet-34 trained from scratch on the same few labels, and the gap narrows but does not reverse when labels increase to 100.
One labeled image per class already yields 74% accuracy, so the feature space is organized before any target-domain labels are seen.
Because the extractor was trained on SETHI patches and evaluated on MSTAR without fine-tuning, the same pretrained encoder could in principle be applied to other sensors and resolutions with comparable domain shifts.
The t-SNE projections show class separation that aligns with k-NN performance, with some classes split into sub-clusters, which suggests the remaining errors concentrate in those fragmented classes.

Reading between the lines

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

A natural next experiment would be to pretrain the same architecture on other large unlabeled SAR collections, such as spaceborne data, and test on MSTAR to map how far the cross-sensor generalization extends.
Publishing patch-level similarity statistics between SETHI and MSTAR would make the cross-sensor claim independently checkable, since the method assumes no near-duplicate content.
An ablation study that removes each augmentation in turn (masking, subband subsampling, despeckling, mean shift) would show which SAR-specific augmentation carries the transfer, since the paper presents them as a bundle.
If the prototype-clustering mechanism is the key, then the method should also transfer to other SAR recognition tasks such as ship or building classification, where labels are equally scarce.

Share X Bluesky LinkedIn Reddit HN

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

Short version: this is a plausible and potentially useful empirical result — a ViT trained with MSN on unlabeled SETHI imagery, frozen, and evaluated on MSTAR with a k-NN head reaches 95.9% ten-shot accuracy without ever seeing MSTAR labels. That is new, as far as the cited literature goes, and it is the kind of result that would matter for SAR applications where labels are scarce. The method itself is not new: MSN, ViT, subband augmentation, MERLIN despeckling are all existing pieces. The contribution is the cross-sensor transfer setup and the specific positive result. Code is released, which is good.

The soft spots are real but not fatal. The empirical support is thin: every accuracy is a single number, no error bars, no multiple random label draws, no seed variation. The k=2 in k-NN is fixed without sensitivity analysis. The baselines (ResNet-34 from scratch for 200 epochs, PCA k-NN) are weak; a better supervised or self-supervised baseline would make the comparison fairer. Also the evaluation uses 8 MSTAR classes, excluding the standard BMP2, BTR70, T72; the headline number should be scoped to the 8-class subset or reproduced on the full 10-class benchmark.

The bigger concern is the unverified semantic disjointness between SETHI training patches and MSTAR targets. The paper asserts in Section 3.3 that the network was not trained on MSTAR data or any military vehicle, but it gives no scene statistics, no example patches, no overlap check. SETHI is an airborne sensor that images terrain with man-made structures; 199,040 random 100x100 patches could plausibly contain vehicles or target-like objects. Under contrastive pretraining, even a small fraction of such patches can shape the feature space, so the 95.9% could partly reflect content familiarity rather than a general sensor-agnostic representation. This is the load-bearing assumption, and the manuscript should verify it with data statistics or at least show representative patches and a targeted analysis.

That said, the central argument is not internally circular: no MSTAR labels are used to train the extractor, and the k-NN uses only a small labeled set for evaluation. The paper is coherent and honest about its components. The missing pieces are experiment rigor and dataset hygiene, not the overall idea.

Who is this for? People working on SAR target recognition, self-supervised learning for remote sensing, and few-shot classification. It deserves a serious referee: the idea is fresh, the result is interesting, and the weaknesses are fixable with more careful evaluation. I would send it to review rather than desk reject, with a request for error bars, a fairer baseline suite, a k sensitivity check, and explicit verification that the SETHI training set is semantically disjoint from MSTAR targets.

Referee Report

4 major / 6 minor

Summary. The paper proposes a self-supervised feature extractor (SFE) for SAR target classification. A ViT-T/8 is trained with a masked siamese contrastive objective on 199,040 unlabeled SETHI patches, then evaluated on the MSTAR dataset without any fine-tuning. Classification is performed with a k-NN on the frozen features, and the paper reports 95.9% accuracy with ten labeled images per class, outperforming a ResNet-34 trained from scratch on MSTAR and a PCA k-NN baseline. The paper also shows t-SNE visualizations and argues that the method demonstrates cross-sensor transfer of SAR features.

Significance. If the cross-sensor transfer claim holds, this is a valuable contribution to label-efficient SAR automatic target recognition, since it would show that an unlabeled, non-MSTAR SAR corpus can yield a feature space that transfers to a different sensor and target set. The method is simple, the code is provided, and the results are potentially useful for few-shot SAR classification. However, the current evidence is not sufficient to establish the central claim: the SETHI training data composition is asserted but never verified, the evaluation uses a single random draw of labeled examples with no error bars, and the baseline comparisons may be understated. The paper's significance therefore hinges on additional verification that the training and test data are truly disjoint in content.

major comments (4)

[3.1, 3.3] The central claim of cross-sensor transfer requires that the SETHI training patches contain no military vehicles or semantically similar objects to MSTAR targets. The paper asserts this in Section 3.3 ('the network not being trained on MSTAR data or any military vehicle') but provides no verification: no scene description, patch content statistics, visual examples, or overlap analysis. Since contrastive pretraining can be shaped by even a small fraction of target-like patches, the reported 95.9% ten-shot accuracy could be explained by content overlap rather than sensor-agnostic transfer. Please add a quantitative description of the SETHI patches (e.g., location, terrain type, presence of man-made structures) and an explicit overlap check, such as nearest-neighbor distances between SETHI patches and MSTAR chips or manual labeling of a random subset of training patches.
[3.4, Table 1] The evaluation uses only 8 MSTAR classes (2S1, BRDM2, BTR60, D7, T62, ZIL131, ZSU23-4, and the SLICY calibration object) and omits BMP2, BTR70, and T72, which are part of the standard 10-class MSTAR benchmark commonly used in the literature. The headline accuracy of 95.9% should either be explicitly scoped to this 8-class subset in the abstract and conclusion, or the experiment should be repeated on the full 10-class set to allow comparison with prior work.
[3.4, Fig. 3] All few-shot accuracies are reported as single numbers, with no error bars, no multiple random draws of the labeled set, and no variation of random seeds. In few-shot classification, the choice of which specific images are labeled can strongly affect accuracy. Report the mean and standard deviation over at least 5–10 independent trials and state the random seed policy. Additionally, the choice of k=2 for the k-NN classifier is not justified; provide a sensitivity analysis over k.
[3.4] The ResNet-34 baseline is trained from scratch for 200 epochs, but no training curves or convergence checks are shown. If the baseline is undertrained, the comparison in Fig. 3 (labeled 'ResNet-34 specifically trained for the task') would be misleading. Provide training convergence or a stronger baseline (e.g., longer training with a learning-rate schedule, or a supervised ViT) to support the claim that the SFE outperforms a reasonably trained supervised model.

minor comments (6)

[1] In the Introduction, 'quatity' should be 'quantity'.
[2.3] The subsection title 'Performances assesment' contains a typo; it should be 'Performance assessment'.
[2.1] The architecture name is written inconsistently as 'Vit' and 'ViT'; please use one consistent capitalization throughout.
[3.1] The values of the temperature τ, the entropy weight λ, and the momentum m are not reported; they are only referenced to [15]. For reproducibility, report these hyperparameters or explicitly state that they exactly follow [15].
[3.2] The preprocessing of MSTAR images before feature extraction with the SFE (e.g., resizing, normalization, patch cropping) is not described. Please clarify how the 128×128 or 192×192 MSTAR chips are fed to the ViT.
[5, ref [10]] Reference [10] has formatting errors in the author list; please correct the citation style.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SETHI-trained contrastive encoder is evaluated on MSTAR with an untrained k-NN, and no target-derived fit is back-propagated into the feature extractor.

full rationale

The paper's derivation chain is not circular. A feature extractor (ViT-T/8) is trained with contrastive learning on 199,040 unlabeled SETHI patches; the MSTAR data are used only for evaluation, via t-SNE visualization and k-NN classification with a small labeled support set. The k-NN classifier is not used to update the encoder, and no fitted constant is extracted from the MSTAR test accuracy into the training procedure. The reported 95.9% accuracy with ten labeled images per class is therefore a genuine out-of-distribution evaluation, assuming the asserted domain gap holds. The hyperparameters (600 epochs, ViT-T/8 patch 8, 256 prototypes, temperature schedule) are inherited from the external Masked Siamese Networks paper [15], not from the authors' own prior work, and the only self-citations ([5], [11]) are contextual references to the authors' anomaly-detection work and the SETHI sensor description, neither of which supplies a load-bearing premise or an imported uniqueness claim. The skeptical concern raised in the reader's take — that SETHI patches may contain target-like objects, making the transfer less clean than claimed — is an empirical validity issue about dataset composition and domain overlap, not a circularity. Section 3.3 states that the network was 'not being trained on MSTAR data or any military vehicle,' but this is a factual assertion about training-data provenance, not a definitional equivalence or a fitted parameter renamed as a prediction. The choice of k=2 and the absence of sensitivity analysis are methodological limitations, but they do not make the central result equivalent to its own input. Consequently, no circular step meets the evidentiary bar of quoting a specific reduction, and the appropriate score is 0.

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

No new physical entities or forces are introduced. The central assumptions are dataset-level: no target leakage between SETHI and MSTAR, and the hand-designed augmentations capture the cross-sensor invariance needed for transfer. The free parameters are the k-NN setting and the learning hyperparameters inherited from prior work.

free parameters (3)

k in k-NN classifier = 2
Reported in Section 3.4 without justification or sensitivity analysis; the headline accuracy depends on this choice.
Contrastive learning hyperparameters (temperature tau, entropy weight lambda, number of prototypes)
Inherited from Masked Siamese Networks [15] and not tuned on MSTAR; no sensitivity study is reported, but they control the learned representation.
Augmentation parameters (crop sizes, subband cropping, mean shift)
Hand-designed for SAR in Section 2.2; their values affect which invariances the extractor learns, and no ablation isolates their contribution.

assumptions (3)

domain assumption SETHI training patches contain no MSTAR targets or military vehicles
Asserted in Section 3.3 but not verified with overlap statistics; if violated, the transfer result is inflated by target leakage.
domain assumption The augmentation set, including masking, subband subsampling, and mean shift, teaches invariances appropriate for the SETHI-to-MSTAR domain shift
Section 2.2 defines augmentations to teach feature invariance; if the augmentations do not match the actual cross-sensor shift, transfer fails.
domain assumption MERLIN despeckling trained on SETHI images generalizes to MSTAR images
Section 2.2 applies despeckling as a teacher augmentation; no evaluation of despeckling quality on MSTAR is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of General Feature Extraction In SAR Target Classification: A Contrastive Learning Approach Across Sensor Types." pith.science (2026). https://pith.science/paper/LIMYBZUD

@misc{pith2026250201162,
  author       = {Pith},
  title        = {Pith review of: General Feature Extraction In SAR Target Classification: A Contrastive Learning Approach Across Sensor Types},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LIMYBZUD}},
  note         = {Machine review of arXiv:2502.01162}
}

read the original abstract

The increased availability of SAR data has raised a growing interest in applying deep learning algorithms. However, the limited availability of labeled data poses a significant challenge for supervised training. This article introduces a new method for classifying SAR data with minimal labeled images. The method is based on a feature extractor Vit trained with contrastive learning. It is trained on a dataset completely different from the one on which classification is made. The effectiveness of the method is assessed through 2D visualization using t-SNE for qualitative evaluation and k-NN classification with a small number of labeled data for quantitative evaluation. Notably, our results outperform a k-NN on data processed with PCA and a ResNet-34 specifically trained for the task, achieving a 95.9% accuracy on the MSTAR dataset with just ten labeled images per class.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 21 canonical work pages

[1]

Widely employed to monitor various activities, SAR plays a crucial role in tracking urban development [2], assessing biomass changes [3], and detect- ing ships [4], for example

INTRODUCTION Synthetic Aperture Radar (SAR) is a remote sensing tech- nology that utilizes microwave signals to capture images of the Earth’s surface [1], providing a unique advantage in all weather conditions due to its ability to penetrate clouds and other atmospheric obstructions. Widely employed to monitor various activities, SAR plays a crucial role ...

work page
[2]

PROPOSED METHOD In this section, we first briefly describe the goal of Siamese networks, the reason ViT is the architecture chosen to extract features from SAR images, the training method and the aug- mentations used, knowing that there is a significant difference between SAR and optical images. The objective is to train a network on a specific sensor wit...

work page Pith review arXiv 2025
[3]

Then, the model is evaluated with qualitative and quantitative tests

EXPERIMENTS First, this section explains the training data and model pa- rameters. Then, the model is evaluated with qualitative and quantitative tests. For the quantitative part, we compare the results of the trained network with a ResNet-34 architecture trained on the MSTAR dataset and with a k-NN applied on the data reduced with PCA. This is done in th...

work page
[4]

It underscores great adaptability across sensor types

CONCLUSION In this paper, we propose a new deep-learning framework based on contrastive learning for SAR feature extraction in the case of classification. It underscores great adaptability across sensor types. Although the feature extractor has not seen a single image of the MSTAR dataset, it can cluster each class accurately. This method leads to great c...

work page
[5]

Self-supervised learning based anomaly de- tection in synthetic aperture radar imaging,

M. Muzeau, C. Ren, S. Angelliaume, M. Datcu, and J.- P. Ovarlez, “Self-supervised learning based anomaly de- tection in synthetic aperture radar imaging,”IEEE Open Journal of Signal Processing, pp. 1–9, 2022

work page 2022
[6]

A tutorial on synthetic aperture radar,

A. Moreira, P. Prats-Iraola, M. Younis, G. Krieger, I. Ha- jnsek, and K. P. Papathanassiou, “A tutorial on synthetic aperture radar,” IEEE Geoscience and Remote Sensing Magazine, vol. 1, no. 1, pp. 6–43, 2013

work page 2013
[7]

New robust statistics for change detection in time se- ries of multivariate sar images,

A. Mian, G. Ginolhac, J.-P. Ovarlez, and A. M. Atto, “New robust statistics for change detection in time se- ries of multivariate sar images,” IEEE Transactions on Signal Processing, vol. 67, no. 2, pp. 520–534, 2019

work page 2019
[8]

Application of l-band sar for mapping tundra shrub biomass, leaf area index, and rainfall interception,

Q. Chang, S. Zwieback, B. DeVries, and A/ Berg, “Application of l-band sar for mapping tundra shrub biomass, leaf area index, and rainfall interception,” Re- mote Sensing of Environment , vol. 268, pp. 112747, 2022

work page 2022

Show all 24 references

[9]

Ship detection in sar im- ages based on an improved faster r-cnn,

J. Li, C. Qu, and J. Shao, “Ship detection in sar im- ages based on an improved faster r-cnn,” in 2017 SAR in Big Data Era: Models, Methods and Applications (BIGSARDATA), 2017, pp. 1–6

2017
[10]

A cookbook of self- supervised learning,

R. Balestriero, M. Ibrahim, V . Sobal, A. Morcos, S. Shekhar, T. Goldstein, F. Bordes, A. Bardes, G. Mi- alon, Y . Tian, A. Schwarzschild, Wilson A. G., Geip- ing J., Garrido Q., Fernandez P., Bar A., Pirsiavash H., Le Cun Y ., and Goldblum M., “A cookbook of self- supervised ...

2023 arXiv
[11]

Beets or cotton? blind extraction of fine agricultural classes using a convolutional autoen- coder applied to temporal sar signatures,

T. Di Martino, R. Guinvarc’h, L. Thirion-Lefevre, and E. Colin Koeniguer, “Beets or cotton? blind extraction of fine agricultural classes using a convolutional autoen- coder applied to temporal sar signatures,” IEEE Trans- actions on Geoscience and Remote Sensing , vol. 60, pp...

2022
[12]

Target recognition in sar images by deep learning with training data augmentation,

Z. Geng, Y . Xu, B.-N. Wang, X. Yu, D.-Y . Zhu, and G. Zhang, “Target recognition in sar images by deep learning with training data augmentation,” Sensors, vol. 23, no. 2, pp. 941, 2023

2023
[13]

Self- supervised feature representation for sar image target classification using contrastive learning,

H. Pei, M. Su, G. Xu, M. Xing, and W. Hong, “Self- supervised feature representation for sar image target classification using contrastive learning,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023

2023
[14]

An image is worth 16x16 words: Trans- formers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weis- senborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Trans- formers for image recognition at scale,” in 9th Interna- tional Conference o...

2021
[15]

Each image is encoded by a ViT (fθstudent or fθteacher) in a feature z ∈ Rde, then projected with a multilayer perceptron head ( Wθstudent or Wθteacher) in a representation h ∈ Rdh

to train the network. Each image is encoded by a ViT (fθstudent or fθteacher) in a feature z ∈ Rde, then projected with a multilayer perceptron head ( Wθstudent or Wθteacher) in a representation h ∈ Rdh. The result is then projected onto a set of prototypes Q = [q1, q2, . . . ...
[16]

SETHI: Review of 10 years of development and experimentation of the remote sensing platform,

R. Baqu ´e, P. Dreuillet, and H. Oriot, “SETHI: Review of 10 years of development and experimentation of the remote sensing platform,” in International Radar Con- ference, 2019, pp. 1–5

2019
[17]

Gen- eralizing from a few examples: A survey on few-shot learning,

Y . Wang, Q. Yao, J. T. Kwok, and L. M. Ni, “Gen- eralizing from a few examples: A survey on few-shot learning,” ACM computing surveys (csur) , vol. 53, no. 3, pp. 1–34, 2020

2020
[18]

Signature verification using a

J. Bromley, I. Guyon, Y . LeCun, E. S ¨ackinger, and R. Shah, “Signature verification using a” siamese” time delay neural network,” Advances in Neural Information Processing Systems, vol. 6, 1993

1993
[19]

Emerging properties in self-supervised vision transformers,

M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 9650–9660

2021
[20]

Masked siamese networks for label-efficient learning,

M. Assran, M. Caron, I. Misra, P. Bojanowski, F. Bor- des, P. Vincent, A. Joulin, M. Rabbat, and N. Ballas, “Masked siamese networks for label-efficient learning,” in European Conference on Computer Vision . Springer, 2022, pp. 456–473

2022
[21]

A simple framework for contrastive learning of visual rep- resentations,

T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual rep- resentations,” in International Conference on Machine Learning. PMLR, 2020, pp. 1597–1607

2020
[22]

Subband ex- traction strategies in ship detection with the subaperture cross-correlation magnitude,

C. Brekke, S. N. Anfinsen, and Y . Larsen, “Subband ex- traction strategies in ship detection with the subaperture cross-correlation magnitude,” IEEE Geoscience and Re- mote Sensing Letters, vol. 10, no. 4, pp. 786–790, 2013

2013
[23]

Some fundamental properties of speckle,

J. W. Goodman, “Some fundamental properties of speckle,” JOSA, vol. 66, no. 11, pp. 1145–1150, 1976

1976
[24]

As if by magic: self-supervised training of deep despeckling networks with MERLIN,

E. Dalsasso, L. Denis, and F. Tupin, “As if by magic: self-supervised training of deep despeckling networks with MERLIN,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–13, 2021

2021

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Widely employed to monitor various activities, SAR plays a crucial role in tracking urban development [2], assessing biomass changes [3], and detect- ing ships [4], for example

INTRODUCTION Synthetic Aperture Radar (SAR) is a remote sensing tech- nology that utilizes microwave signals to capture images of the Earth’s surface [1], providing a unique advantage in all weather conditions due to its ability to penetrate clouds and other atmospheric obstructions. Widely employed to monitor various activities, SAR plays a crucial role ...

work page

[2] [2]

PROPOSED METHOD In this section, we first briefly describe the goal of Siamese networks, the reason ViT is the architecture chosen to extract features from SAR images, the training method and the aug- mentations used, knowing that there is a significant difference between SAR and optical images. The objective is to train a network on a specific sensor wit...

work page Pith review arXiv 2025

[3] [3]

Then, the model is evaluated with qualitative and quantitative tests

EXPERIMENTS First, this section explains the training data and model pa- rameters. Then, the model is evaluated with qualitative and quantitative tests. For the quantitative part, we compare the results of the trained network with a ResNet-34 architecture trained on the MSTAR dataset and with a k-NN applied on the data reduced with PCA. This is done in th...

work page

[4] [4]

It underscores great adaptability across sensor types

CONCLUSION In this paper, we propose a new deep-learning framework based on contrastive learning for SAR feature extraction in the case of classification. It underscores great adaptability across sensor types. Although the feature extractor has not seen a single image of the MSTAR dataset, it can cluster each class accurately. This method leads to great c...

work page

[5] [5]

Self-supervised learning based anomaly de- tection in synthetic aperture radar imaging,

M. Muzeau, C. Ren, S. Angelliaume, M. Datcu, and J.- P. Ovarlez, “Self-supervised learning based anomaly de- tection in synthetic aperture radar imaging,”IEEE Open Journal of Signal Processing, pp. 1–9, 2022

work page 2022

[6] [6]

A tutorial on synthetic aperture radar,

A. Moreira, P. Prats-Iraola, M. Younis, G. Krieger, I. Ha- jnsek, and K. P. Papathanassiou, “A tutorial on synthetic aperture radar,” IEEE Geoscience and Remote Sensing Magazine, vol. 1, no. 1, pp. 6–43, 2013

work page 2013

[7] [7]

New robust statistics for change detection in time se- ries of multivariate sar images,

A. Mian, G. Ginolhac, J.-P. Ovarlez, and A. M. Atto, “New robust statistics for change detection in time se- ries of multivariate sar images,” IEEE Transactions on Signal Processing, vol. 67, no. 2, pp. 520–534, 2019

work page 2019

[8] [8]

Application of l-band sar for mapping tundra shrub biomass, leaf area index, and rainfall interception,

Q. Chang, S. Zwieback, B. DeVries, and A/ Berg, “Application of l-band sar for mapping tundra shrub biomass, leaf area index, and rainfall interception,” Re- mote Sensing of Environment , vol. 268, pp. 112747, 2022

work page 2022