SAR

Synthetic Aperture Radar

Join the Competition Contribute to Benchmark

Standard reconstruction benchmark — forward model perfectly known, no calibration needed. Score = 0.5 × clip((PSNR−15)/30, 0, 1) + 0.5 × SSIM

#	Method	Score	PSNR (dB)	SSIM	Source
🥇	DiffusionSAR DiffusionSAR Wei et al., NeurIPS 2024 35.42 dB SSIM 0.955 Checkpoint unavailable	0.818	35.42	0.955	✓ Certified	Wei et al., NeurIPS 2024
🥈	PanSharpener++ PanSharpener++ Zhang et al., ICCV 2024 34.58 dB SSIM 0.945 Checkpoint unavailable	0.799	34.58	0.945	✓ Certified	Zhang et al., ICCV 2024
🥉	SARFormer SARFormer Li et al., CVPR 2024 33.85 dB SSIM 0.932 Checkpoint unavailable	0.780	33.85	0.932	✓ Certified	Li et al., CVPR 2024
4	ScoreSAR ScoreSAR Johnson et al., ECCV 2025 31.9 dB SSIM 0.942 Checkpoint unavailable	0.753	31.9	0.942	✓ Certified	Johnson et al., ECCV 2025
5	SAR-CAM SAR-CAM Cross-attention SAR, 2024 32.1 dB SSIM 0.912 Checkpoint unavailable	0.741	32.1	0.912	✓ Certified	Cross-attention SAR, 2024
6	SARDenoiserViT SARDenoiserViT Wang et al., ICCV 2024 30.2 dB SSIM 0.920 Checkpoint unavailable	0.713	30.2	0.920	✓ Certified	Wang et al., ICCV 2024
7	SAR-DRN SAR-DRN Zhang et al., RS 2018 30.6 dB SSIM 0.882 Checkpoint unavailable	0.701	30.6	0.882	✓ Certified	Zhang et al., RS 2018
8	SAR-ResNet SAR-ResNet Chen et al., IEEE TGRS 2022 28.84 dB SSIM 0.897 Checkpoint unavailable	0.679	28.84	0.897	✓ Certified	Chen et al., IEEE TGRS 2022
9	Lee Filter Lee Filter Lee, IEEE TGRS 1980 28.75 dB SSIM 0.896 Try in SpecLab →	0.677	28.75	0.896	✓ Certified	Lee, IEEE TGRS 1980
10	SAR-BM3D SAR-BM3D Parrilli et al., IEEE TGRS 2012 27.2 dB SSIM 0.790 Try in SpecLab →	0.598	27.2	0.790	✓ Certified	Parrilli et al., IEEE TGRS 2012
11	Range-Doppler Range-Doppler SAR signal processing baseline 25.9 dB SSIM 0.829 Try in SpecLab →	0.596	25.9	0.829	✓ Certified	SAR signal processing baseline
12	Chirp Scaling Chirp Scaling Raney et al., IEEE TGRS 1994 22.68 dB SSIM 0.718 Try in SpecLab →	0.487	22.68	0.718	✓ Certified	Raney et al., IEEE TGRS 1994
13	Matched Filter Matched Filter Standard SAR focusing 23.5 dB SSIM 0.640 Try in SpecLab →	0.462	23.5	0.640	✓ Certified	Standard SAR focusing

Dataset: PWM Benchmark (13 algorithms)

Blind Reconstruction Challenge — forward model has unknown mismatch, must calibrate from data. Score = 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖)

#	Method	Overall Score	Public PSNR / SSIM	Dev PSNR / SSIM	Hidden PSNR / SSIM	Trust	Source
🥇	DiffusionSAR + gradient DiffusionSAR + gradient Wei et al., NeurIPS 2024 Score 0.731 Correct & Reconstruct →	0.731	0.789 33.01 dB / 0.953	0.729 29.75 dB / 0.913	0.676 27.68 dB / 0.874	✓ Certified	Wei et al., NeurIPS 2024
🥈	SARFormer + gradient SARFormer + gradient Li et al., CVPR 2024 Score 0.693 Correct & Reconstruct →	0.693	0.762 30.88 dB / 0.929	0.690 26.85 dB / 0.854	0.626 24.88 dB / 0.798	✓ Certified	Li et al., CVPR 2024
🥉	SAR-CAM + gradient SAR-CAM + gradient Cross-attention SAR, 2024 Score 0.689 Correct & Reconstruct →	0.689	0.738 29.47 dB / 0.908	0.704 28.2 dB / 0.885	0.624 24.86 dB / 0.798	✓ Certified	Cross-attention SAR, 2024
4	PanSharpener++ + gradient PanSharpener++ + gradient Zhang et al., ICCV 2024 Score 0.669 Correct & Reconstruct →	0.669	0.803 33.58 dB / 0.958	0.654 25.7 dB / 0.823	0.551 21.32 dB / 0.660	✓ Certified	Zhang et al., ICCV 2024
5	ScoreSAR + gradient ScoreSAR + gradient Johnson et al., ECCV 2025 Score 0.647 Correct & Reconstruct →	0.647	0.732 28.9 dB / 0.898	0.627 24.14 dB / 0.773	0.582 23.23 dB / 0.740	✓ Certified	Johnson et al., ECCV 2025
6	Lee Filter + gradient Lee Filter + gradient Lee, IEEE TGRS 1980 Score 0.646 Correct & Reconstruct →	0.646	0.705 27.18 dB / 0.862	0.626 24.87 dB / 0.798	0.607 24.07 dB / 0.771	✓ Certified	Lee, IEEE TGRS 1980
7	SAR-DRN + gradient SAR-DRN + gradient Zhang et al., RS 2018 Score 0.621 Correct & Reconstruct →	0.621	0.717 28.46 dB / 0.890	0.617 24.23 dB / 0.777	0.529 20.68 dB / 0.631	✓ Certified	Zhang et al., RS 2018
8	Range-Doppler + gradient Range-Doppler + gradient SAR signal processing baseline Score 0.610 Correct & Reconstruct →	0.610	0.617 23.74 dB / 0.759	0.613 23.97 dB / 0.767	0.600 23.52 dB / 0.751	✓ Certified	SAR signal processing baseline
9	SARDenoiserViT + gradient SARDenoiserViT + gradient Wang et al., ICCV 2024 Score 0.609 Correct & Reconstruct →	0.609	0.732 28.51 dB / 0.891	0.606 24.03 dB / 0.770	0.488 19.33 dB / 0.566	✓ Certified	Wang et al., ICCV 2024
10	SAR-ResNet + gradient SAR-ResNet + gradient Chen et al., IEEE TGRS 2022 Score 0.596 Correct & Reconstruct →	0.596	0.708 27.36 dB / 0.867	0.572 22.08 dB / 0.693	0.508 20.7 dB / 0.632	✓ Certified	Chen et al., IEEE TGRS 2022
11	SAR-BM3D + gradient SAR-BM3D + gradient Parrilli et al., IEEE TGRS 2012 Score 0.572 Correct & Reconstruct →	0.572	0.645 24.9 dB / 0.799	0.555 21.54 dB / 0.670	0.516 20.28 dB / 0.612	✓ Certified	Parrilli et al., IEEE TGRS 2012
12	Chirp Scaling + gradient Chirp Scaling + gradient Raney et al., IEEE TGRS 1994 Score 0.509 Correct & Reconstruct →	0.509	0.558 21.01 dB / 0.646	0.495 19.98 dB / 0.598	0.473 19.44 dB / 0.572	✓ Certified	Raney et al., IEEE TGRS 1994
13	Matched Filter + gradient Matched Filter + gradient Standard SAR focusing Score 0.496 Correct & Reconstruct →	0.496	0.542 20.79 dB / 0.636	0.499 19.98 dB / 0.598	0.447 18.38 dB / 0.519	✓ Certified	Standard SAR focusing

Complete score requires all 3 tiers (Public + Dev + Hidden).

Join the competition →

Scoring: 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖) PSNR 40% · SSIM 40% · Consistency 20%

Public 3 scenes

Full-access development tier with all data visible.

What you get & how to use

What you get: Measurements (y), ideal forward operator (H), spec ranges, ground truth (x_true), and true mismatch spec.

How to use: Load HDF5 → compare reconstruction vs x_true → check consistency → iterate.

What to submit: Reconstructed signals (x_hat) and corrected spec as HDF5.

Public Leaderboard

#	Method	Score	PSNR	SSIM
1	PanSharpener++ + gradient	0.803	33.58	0.958
2	DiffusionSAR + gradient	0.789	33.01	0.953
3	SARFormer + gradient	0.762	30.88	0.929
4	SAR-CAM + gradient	0.738	29.47	0.908
5	ScoreSAR + gradient	0.732	28.9	0.898
6	SARDenoiserViT + gradient	0.732	28.51	0.891
7	SAR-DRN + gradient	0.717	28.46	0.89
8	SAR-ResNet + gradient	0.708	27.36	0.867
9	Lee Filter + gradient	0.705	27.18	0.862
10	SAR-BM3D + gradient	0.645	24.9	0.799
11	Range-Doppler + gradient	0.617	23.74	0.759
12	Chirp Scaling + gradient	0.558	21.01	0.646
13	Matched Filter + gradient	0.542	20.79	0.636

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
motion_error	-2.0	4.0	cm
phase_error	-0.3	0.6	rad
range_cell_migration	-0.5	1.0	cells

View Details →

Dev 3 scenes

Blind evaluation tier — no ground truth available.

What you get & how to use

What you get: Measurements (y), ideal forward operator (H), and spec ranges only.

How to use: Apply your pipeline from the Public tier. Use consistency as self-check.

What to submit: Reconstructed signals and corrected spec. Scored server-side.

Dev Leaderboard

#	Method	Score	PSNR	SSIM
1	DiffusionSAR + gradient	0.729	29.75	0.913
2	SAR-CAM + gradient	0.704	28.2	0.885
3	SARFormer + gradient	0.690	26.85	0.854
4	PanSharpener++ + gradient	0.654	25.7	0.823
5	ScoreSAR + gradient	0.627	24.14	0.773
6	Lee Filter + gradient	0.626	24.87	0.798
7	SAR-DRN + gradient	0.617	24.23	0.777
8	Range-Doppler + gradient	0.613	23.97	0.767
9	SARDenoiserViT + gradient	0.606	24.03	0.77
10	SAR-ResNet + gradient	0.572	22.08	0.693
11	SAR-BM3D + gradient	0.555	21.54	0.67
12	Matched Filter + gradient	0.499	19.98	0.598
13	Chirp Scaling + gradient	0.495	19.98	0.598

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
motion_error	-2.4	3.6	cm
phase_error	-0.36	0.54	rad
range_cell_migration	-0.6	0.9	cells

Submit Reconstruction View Details →

Hidden 3 scenes

Fully blind server-side evaluation — no data download.

What you get & how to use

What you get: No data downloadable. Algorithm runs server-side on hidden measurements.

How to use: Package algorithm as Docker container / Python script. Submit via link.

What to submit: Containerized algorithm accepting y + H, outputting x_hat + corrected spec.

Hidden Leaderboard

#	Method	Score	PSNR	SSIM
1	DiffusionSAR + gradient	0.676	27.68	0.874
2	SARFormer + gradient	0.626	24.88	0.798
3	SAR-CAM + gradient	0.624	24.86	0.798
4	Lee Filter + gradient	0.607	24.07	0.771
5	Range-Doppler + gradient	0.600	23.52	0.751
6	ScoreSAR + gradient	0.582	23.23	0.74
7	PanSharpener++ + gradient	0.551	21.32	0.66
8	SAR-DRN + gradient	0.529	20.68	0.631
9	SAR-BM3D + gradient	0.516	20.28	0.612
10	SAR-ResNet + gradient	0.508	20.7	0.632
11	SARDenoiserViT + gradient	0.488	19.33	0.566
12	Chirp Scaling + gradient	0.473	19.44	0.572
13	Matched Filter + gradient	0.447	18.38	0.519

Spec Ranges (3 parameters)

Parameter	Min	Max	Unit
motion_error	-1.4	4.6	cm
phase_error	-0.21	0.69	rad
range_cell_migration	-0.35	1.15	cells

Submit Algorithm View Details →

Blind Reconstruction Challenge

Challenge

Given measurements with unknown mismatch and spec ranges (not exact params), reconstruct the original signal. A method must be evaluated on all three tiers for a complete score. Scored on a composite metric: 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖).

Input

Measurements y, ideal forward model H, spec ranges

Output

Reconstructed signal x̂

About the Imaging Modality

SAR synthesizes a large antenna aperture by combining coherent radar returns collected as the platform (satellite/aircraft) moves along its flight path. The azimuth resolution is achieved by coherent integration of the Doppler history, while range resolution comes from pulse compression (chirp). The forward model is a 2D convolution with the SAR impulse response in range and azimuth. SAR images exhibit speckle noise (multiplicative, fully developed) from coherent interference of distributed scatterers. Applications include Earth observation, terrain mapping, and interferometric displacement measurement.

Principle

Synthetic Aperture Radar achieves fine azimuth resolution by coherently processing radar echoes collected as the antenna moves along its flight path, synthesizing an aperture much larger than the physical antenna. The SAR signal processor applies matched filtering (pulse compression) in both range and azimuth to form a high-resolution complex image. SAR operates through clouds, at night, and in all weather conditions.

How to Build the System

Mount a microwave transmitter/receiver (C-band 5.4 GHz, L-band 1.3 GHz, or X-band 9.6 GHz) on a satellite (Sentinel-1, RADARSAT) or aircraft. The antenna illuminates a strip on the ground as the platform moves. Record the complex (I/Q) echo data with precise pulse timing and platform position/velocity from GNSS/INS. Range resolution is set by pulse bandwidth (1-200 MHz); azimuth resolution equals L_ant/2 (half the antenna length).

Common Reconstruction Algorithms

Range-Doppler algorithm (range compression + azimuth compression)
Chirp scaling algorithm for wide-swath SAR
Omega-K (wavenumber domain) algorithm for high-resolution spotlight SAR
InSAR (Interferometric SAR) for DEM generation and deformation mapping
PolSAR decomposition (Cloude-Pottier, Freeman-Durden) for land classification

Common Mistakes

Incorrect motion compensation causing azimuth defocusing
Range cell migration not properly corrected for squinted geometries
Phase errors from atmospheric delay (troposphere, ionosphere) in InSAR
Ambiguities (range or azimuth) from incorrect PRF selection
Speckle noise mistaken for real features in SAR imagery

How to Avoid Mistakes

Use precise INS/GNSS data for autofocus and motion compensation
Apply appropriate RCMC (Range Cell Migration Correction) for the imaging geometry
Use atmospheric phase screens (from weather models or GNSS delays) for InSAR correction
Design PRF to avoid range and azimuth ambiguity constraints for the swath geometry
Apply multi-look or speckle filtering (Lee, refined-Lee) before interpretation

Forward-Model Mismatch Cases

The widefield fallback produces a real-valued blurred image, but SAR acquires complex-valued (I/Q) radar echoes that require coherent pulse compression in range and azimuth — the phase information essential for InSAR and coherent processing is lost
SAR image formation requires matched filtering with the transmitted chirp waveform and Doppler history — the widefield spatial blur cannot model microwave scattering, range-Doppler processing, or speckle statistics

How to Correct the Mismatch

Use the SAR operator that models coherent radar echo formation: each pixel's complex return includes amplitude (backscatter cross-section) and phase (range + Doppler history), requiring range and azimuth compression
Process using range-Doppler, chirp scaling, or omega-K algorithms for image formation; preserve complex data for InSAR, PolSAR, and coherence-based applications

Experimental Setup — Signal Chain

Experimental setup diagram for Synthetic Aperture Radar

Reconstruction Gallery — 4 Scenes × 3 Scenarios

Method: CPU_baseline | Mismatch: nominal (nominal=True, perturbed=False)

I Ideal II Mismatch (uncorrected) III Oracle correction

I — Ideal

Ground Truth

Measurement

Reconstruction

15.503504478809369 dB SSIM 0.2341635915614249

II — Mismatch

Ground Truth

Measurement

Reconstruction

15.447084365158803 dB SSIM 0.2373800670650224

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

15.545546774161242 dB SSIM 0.23052417070261594

I — Ideal

Ground Truth

Measurement

Reconstruction

29.027661217474883 dB SSIM 0.8684057272970452

II — Mismatch

Ground Truth

Measurement

Reconstruction

28.58935657255064 dB SSIM 0.8712144155448451

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

31.592415537777523 dB SSIM 0.9075777438597679

I — Ideal

Ground Truth

Measurement

Reconstruction

29.227212879824922 dB SSIM 0.8516233712949678

II — Mismatch

Ground Truth

Measurement

Reconstruction

28.334155500148754 dB SSIM 0.8570647803988047

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

31.36543035296549 dB SSIM 0.8997771391179263

I — Ideal

Ground Truth

Measurement

Reconstruction

28.22777093634853 dB SSIM 0.8776792969503701

II — Mismatch

Ground Truth

Measurement

Reconstruction

29.575687028393908 dB SSIM 0.8479275774293132

III — Oracle

Ground Truth

Measurement (perturbed)

Reconstruction

32.56408008462443 dB SSIM 0.8954795385690928

Mean PSNR Across All Scenes

25.496537378114425 dB

I — Ideal

25.486570866563028 dB

II — Mismatch

27.766868187382173 dB

III — Oracle

Degradation: -0.0 dB Recovery: +2.3 dB

Per-scene PSNR breakdown (4 scenes)

Scene	I (PSNR)	I (SSIM)	II (PSNR)	II (SSIM)	III (PSNR)	III (SSIM)
scene_00	15.503504478809369	0.2341635915614249	15.447084365158803	0.2373800670650224	15.545546774161242	0.23052417070261594
scene_01	29.027661217474883	0.8684057272970452	28.58935657255064	0.8712144155448451	31.592415537777523	0.9075777438597679
scene_02	29.227212879824922	0.8516233712949678	28.334155500148754	0.8570647803988047	31.36543035296549	0.8997771391179263
scene_03	28.22777093634853	0.8776792969503701	29.575687028393908	0.8479275774293132	32.56408008462443	0.8954795385690928
Mean	25.496537378114425	0.707967996775952	25.486570866563028	0.7033967101094963	27.766868187382173	0.7333396480623507

Experimental Setup

Instrument: Sentinel-1 (ESA Copernicus) / TerraSAR-X

Frequency Band: C-band (5.405 GHz)

Wavelength Cm: 5.6

Mode: IW (Interferometric Wide Swath)

Spatial Resolution M: 5 (range) x 20 (azimuth)

Swath Km: 250

Polarization: VV + VH (dual-pol)

Incidence Angle Deg: 29.1-46.0

Revisit Days: 6

Key References

Cumming & Wong, 'Digital Processing of Synthetic Aperture Radar Data', Artech House (2005)
Torres et al., 'GMES Sentinel-1 mission', Remote Sensing of Environment 120, 9-24 (2012)

Canonical Datasets

SEN12MS (Schmitt et al., multi-modal Sentinel-1/2)
SpaceNet 6 (SAR building footprints)

Spec DAG — Forward Model Pipeline

F(azimuth×range) → D(g, η₁)

F Range-Doppler Sampling (azimuth×range)

D Radar Receiver (g, η₁)

Mismatch Parameters

Symbol	Parameter	Description	Perturbed
Δr_a	motion_error	Platform motion error (cm)	2.0
Δφ	phase_error	Autofocus phase error (rad)	0.3
ΔRCM	range_cell_migration	Range cell migration error (cells)	0.5

Credits System

40%

Platform Profit Pool

Revenue allocated to benchmark rewards

30%

Winner Share

Top algorithm receives from pool

$100

Min Withdrawal

Minimum payout threshold

Spec Primitives Reference (11 primitives)

P Propagation

Free-space or medium propagation kernel (Fresnel, Rayleigh-Sommerfeld).

M Mask / Modulation

Spatial or spatio-temporal amplitude modulation (coded aperture, SLM pattern).

Π Projection

Geometric projection operator (Radon transform, fan-beam, cone-beam).

F Fourier Sampling

Sampling in the Fourier / k-space domain (MRI, ptychography).

C Convolution

Shift-invariant convolution with a point-spread function (PSF).

Σ Summation / Integration

Summation along a physical dimension (spectral, temporal, angular).

D Detector

Sensor readout with gain g and noise model η (Gaussian, Poisson, mixed).

S Structured Illumination

Patterned illumination (block, Hadamard, random) applied to the scene.

W Wavelength Dispersion

Spectral dispersion element (prism, grating) with shift α and aperture a.

R Rotation / Motion

Sample or gantry rotation (CT, electron tomography).

Λ Wavelength Selection

Spectral filter or monochromator selecting a wavelength band.