SAR-VesselBench: 32 controlled vessel-detection checkpoints

Fine-tuned dark-vessel detectors for Sentinel-1 SAR from the SAR-VesselBench label-efficiency study (code and evidence). Two size-matched tracks (ViT-B/16 and ConvNeXt-V2-Base) each compare four initializations — random, optical remote sensing, SAR, and ImageNet — at four nested label budgets (12, 28, 56, and 111 xView3-SAR training scenes). Every cell shares one CenterNet-style point detector, optimizer, schedule, scene-level split, and seed, trained in strict FP32 on one eight-GPU H100 node, one single-GPU process per cell.

Each file here is the cohort-bound best development checkpoint of one cell: a PyTorch Lightning checkpoint containing the full detector (encoder, adapter, and 256-channel head; ~90M parameters ViT track, ~94M CNN track). TRAINING_COHORT.json binds every checkpoint's SHA-256, selected epoch, and development-selected operating threshold; verify any download against it.

Results (seed-0 point estimates)

Development-selection F1 / once-scored held-out 16-scene test F1 at each cell's checkpoint-bound operating threshold:

Track	Init	10% (12)	25% (28)	50% (56)	100% (111)
ViT-B/16	Random	0.8030 / 0.7042	0.8563 / 0.7372	0.8912 / 0.8101	0.8958 / 0.8101
ViT-B/16	Optical (SatDINO)	0.8902 / 0.8198	0.8802 / 0.8279	0.9136 / 0.8495	0.9268 / 0.8465
ViT-B/16	SAR (SARMAE)	0.8857 / 0.8041	0.8624 / 0.7715	0.9161 / 0.8672	0.9289 / 0.8748
ViT-B/16	ImageNet (AugReg)	0.8761 / 0.8103	0.8716 / 0.8051	0.9284 / 0.8545	0.9399 / 0.8569
ConvNeXt-V2-B	Random	0.7992 / 0.6571	0.8167 / 0.7065	0.8664 / 0.7807	0.8919 / 0.8124
ConvNeXt-V2-B	Optical (BigEarthNet S2)	0.7992 / 0.6712	0.8063 / 0.7024	0.8586 / 0.7616	0.8636 / 0.7948
ConvNeXt-V2-B	SAR (BigEarthNet S1)	0.8762 / 0.8000	0.8868 / 0.8025	0.8998 / 0.8521	0.9068 / 0.8221
ConvNeXt-V2-B	ImageNet (FCMAE+sup)	0.8856 / 0.7810	0.9005 / 0.8481	0.9221 / 0.8807	0.9387 / 0.8966

Headline findings: transfer concentrates its value under label scarcity; the SAR-versus-optical contrast is architecture-dependent (SAR wins the CNN track at every budget, optical leads the ViT track below half data, and the held-out split confirms the sign at every budget); and held-out scoring reverses one development-selection conclusion — the full-data ViT winner flips from ImageNet to SAR. Two SAR test curves violate a 0.02 monotonicity tolerance as the budget grows; treat the grid as descriptive seed-0 point estimates. Test scenes are Sentinel-1 revisits of regions near training scenes, so absolute values measure in-region generalization.

File naming

<arm>-f<fraction>-s0/checkpoints/best.ckpt, where <arm> is one of vitrand, satdino, sarmae, vitin1k (ViT track) or cnnrand, beS2, beS1, cnnin1k (CNN track), and <fraction> is the label budget in percent (10, 25, 50, 100).

Loading

Checkpoints load through the study's code (HeatmapLitModule.load_from_checkpoint from the SAR-VesselBench repository, which also defines the whole-scene tiled-inference path and the frozen scorer). Inputs are three-channel decibel tensors [VH, VV, VH-VV] normalized with the repository's committed training statistics; each cell's operating threshold is recorded in TRAINING_COHORT.json.

License — read before use

These checkpoints are fine-tuned derivatives of released pretrained encoders, and each cell inherits its source's terms:

Cells	Source encoder	Inherited terms
sarmae-*	SARMAE (SAR-1M)	CC BY-NC 4.0 — noncommercial use only
satdino-*	SatDINO (fMoW-RGB)	Apache-2.0
beS1-*, beS2-*	BigEarthNet-v2 ConvNeXt	upstream model-card terms
vitin1k-*, cnnin1k-*	timm ImageNet-1K	Apache-2.0
vitrand-*, cnnrand-*	none (random init)	MIT (this project)

The four sarmae-* checkpoints may not be used commercially. Fine-tuning data is xView3-SAR (Paolo et al., NeurIPS 2022); its labels and imagery retain their own distribution terms.

Citation

John Roth and Kyle Wagner. Label-Efficient Dark-Vessel Detection in SAR: Does SAR-Domain Pretraining Outperform Optical and ImageNet Transfer Across ViT and CNN? Johns Hopkins University, EN.705.643, August 2026.