Synthetic aperture radar (SAR) foundation models provide transferable representations for diverse perception tasks in all-weather Earth observation. However, their development often relies on costly large-scale SAR pretraining, raising the question of whether such extensive domain-specific pretraining is necessary. We find that modern multimodal large language models (MLLMs) already provide visual encoders that can outperform dedicated SAR foundation models as downstream initializations. This motivates a new paradigm: reuse general visual capabilities and devote domain training to brief specialization. We therefore introduce SARVIA, an adaptive post-pretraining method that specializes modern MLLMs using self-supervised visual questions with computable answers. These questions are constructed from unlabeled observations through general image tasks that capture spatial relations and SAR-specific tasks that exploit native measurements and SAR--optical correspondence. A shared answer-prediction objective jointly trains low-rank adapters across visual and language components, with loss-driven task sampling. We then export the adapted visual encoder as a standalone backbone. With 600 adaptation steps and 2 GPU-hours, SARVIA outperforms the compared foundation models on all six SAR detection and segmentation benchmarks. Compared with dedicated SAR pretraining, SARVIA reduces domain-training costs by orders of magnitude while improving visual transfer across architectures and scales. This new paradigm shifts the emphasis from repeated, costly pretraining to a reusable specialization process that can benefit from advances in open MLLMs. Dataset and code: https://github.com/zcablii/SARVIA.
Figures & tables
Figure 1: Paradigm shift from conventional dedicated pretraining to our proposed efficient adaptation. SARVIA improves SAR foundation models at lower pretraining cost and offers a reusable adaptation route as open MLLMs advance.
Figure 2: ViT-L downstream performance across six SAR benchmarks. SARVIA achieves the best score on every benchmark.
Figure 3: SARVIA overview. Unlabeled SAR observations yield tasks with deterministic answers. Answer loss jointly trains LoRA in the ViT, merger, and language model, while detached task losses adjust future QA draws through the loss-driven task allocation. We merge the visual adapters and export only the ViT for downstream perception.
Dataset
Task
Categories
Metric
Train
Val.
Test
Prediction head
Input / crop
Schedule
RSAR
Oriented det.
6
AP 50
78,837
–
8,538
Oriented R-CNN [ 48 ]
8002
12 epochs
M4-SAR
Oriented det.
6
AP 50
56,116
22,112
33,946
Oriented R-CNN [ 48 ]
5122
12 epochs
SARDet-100K
Horizontal det.
6
mAP
94,493
10,492
11,613
Cascade R-CNN [ 4 ]
10242
12 epochs
OpenEarthMap-SAR
Segmentation
8
mIoU
4,333
210
490
UPerNet [ 46 ]
5122
40k steps
AIR-PolSAR-Seg-2.0
Segmentation
6
mIoU
1,233
309
–
UPerNet [ 46 ]
5122
40k steps
WHU-OPT-SAR
Segmentation
7
mIoU
17,640
5,880
5,880
UPerNet [ 46 ]
2562
40k steps
Table 1: Downstream benchmark settings. Evaluation uses the listed test partitions, except AIR-PolSAR-Seg-2.0 (validation).
Encoder source
Backbone
Vanilla
+SARVIA
GPU-hours
SAR / Earth-observation foundation models
TerraMind [ 20 ]
ViT-L
34.69
—
≈ 5,122
SARMAE [ 31 ]
ViT-L
33.07
—
≈ 320
MLLM visual encoders
Qwen3-VL-2B [ 1 ]
ViT-L
28.42
32.66 (+4.24)
1.87
InternVL2.5-2B [ 8 ]
ViT-L
33.79
35.37 (+1.58)
2.71
Table 2: Standalone encoder transfer on OEM-SLC-Amplitude (mIoU, %). GPU-hours of pretraining are estimated or measured on a single RTX PRO 6000.
Method
Backbone
AP 50 ↑
O-RCNN [ 48 ]
ResNet-50
64.82
SARMAE [ 31 ]
ViT-B
66.80
RoI Transformer [ 12 ]
ResNet-50
66.95
GSDet [ 13 ]
ResNet-50
68.00
SARVIA (Ours)
ViT-B
70.29
OlmoEarth [ 18 ]
ViT-L
57.93
Table 3: Results of RSAR [ 51 ] oriented detection.
Method
Backbone
mAP ↑
SUMMIT [ 15 ]
ViT-B
50.86
SARMAE [ 31 ]
ViT-B
52.00
SARATR-X [ 22 ]
HiViT-B
54.20
DINO (UniSAR-7M) [ 5 , 41 ]
ViT-B
54.22
SARVIA (Ours)
ViT-B
55.10
SatMAE [ 10 ]
ViT-L
40.98
Table 4: Results of SARDet-100K [ 28 ] horizontal detection.
Method
Backbone
AP 50 ↑
OlmoEarth [ 18 ]
ViT-B
21.86
TerraMind [ 20 ]
ViT-B
39.12
SUMMIT [ 15 ]
ViT-B
40.18
SARMAE [ 31 ]
ViT-B
40.73
SARVIA (Ours)
ViT-B
41.23
SatMAE [ 10 ]
ViT-L
23.26
Table 5: Results of M4-SAR [ 42 ] oriented detection.
Method
Backbone
mIoU ↑
SARATR-X-v2 [ 21 ]
iTPN-B
31.34
VMamba [ 35 ]
VSS
34.74
U-Net [ 40 ]
EfficientNet-B4
35.13
SARVIA (Ours)
ViT-B
35.76
OlmoEarth [ 18 ]
ViT-L
27.85
SatMAE [ 10 ]
ViT-L
27.91
Table 6: Results of OpenEarthMap-SAR [ 45 ] segmentation.
Method
Backbone
mIoU ↑
SARATR-X-v2 [ 21 ]
iTPN-B
83.43
TerraSegNet [ 44 ]
EfficientNetV2-S
92.10
SARVIA (Ours)
ViT-B
93.37
SatMAE [ 10 ]
ViT-L
78.75
OlmoEarth [ 18 ]
ViT-L
79.25
TerraMind [ 20 ]
ViT-L
88.12
Table 7: Results of AIR-PolSAR-Seg-2.0 [ 43 ] segmentation.
Method
Backbone
mIoU ↑
PIDNet [ 49 ]
PIDNet-S
44.30
OCRNet [ 50 ]
HRNetV2-W48
44.80
SegFormer [ 47 ]
MiT-B4
45.30
SARATR-X-v2 [ 21 ]
iTPN-B
46.10
Mask2Former [ 9 ]
ResNet-101
46.80
TransUNet [ 6 ]
ResNet-50 + ViT
47.10
Table 8: Results of WHU-OPT-SAR [ 24 ] segmentation.
Figure 4: Adaptation length and rank on Qwen3.5-2B. OEM-SLC-Amplitude mIoU for four LoRA ranks. The dashed line marks the original visual encoder.
Figure 5: How adaptive sampling changes training. (a) Cumulative changes in task exposure, colored by task family. (b) Loss and sampling trajectories for two SARLO tasks. Fast NLL is the fast exponential moving average of answer-token negative log-likelihood. Consumed p/p0 is the ratio of the sampling probability used in training to its initial value. (c) Question-answering (QA) accuracy and OEM-SLC-Amplitude mIoU can diverge.
Sampling
Task families
mIoU ↑
Base encoder
—
36.29
Static
General image only
37.33
SAR-specific only
37.94
Both
38.21
TSCL [ 37 ]
Both
37.75
DWA [ 33 ]
37.33
Table 9: Task-family ablations and sampling strategies on OEM-SLC-Amplitude.
Update
QA draws
Last-used TV
Budget shifted
400
51,200
21.85%
5.46%
600
76,800
22.97%
10.55%
800
102,400
24.53%
13.04%
Table 10: Reallocation relative to the initial mixture. Last-used TV is the total variation distance 21∑g∣pg−pg0∣ , where pg is the last-used probability of task variant g and pg0 its initial probability. Budget shifted measures the fraction of cumulative expected draws reassigned.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Source
Train
Eval
TerraMesh
703,460
257,276
SARLO-80
101,699
8,700
KuroSiwo
53,889
51,152
Total
859,048
317,128
Appendix
Table S1: Source records used for SAR adaptation post-pretraining and the evaluation pool from which diagnostic questions are sampled.
Task group
Task and answer
Var.
T
K
S
General image tasks
Basic spatial tasks
Recover SAR patch order, match crops, identify flips or missing patches, and locate polygonal crops. Return a permutation, candidate index, flip descriptor, or box.
6
6
6
6
Single-modality ordering
Reorder SAR or optical patches using a same-modality reference, one fixed anchor, or no reference. Return the permutation restoring a 2×2 or 3×3 grid.
12
6
6
12
Reference flips
Compare a SAR image with its unchanged reference to identify the flipped cell in a 2×2 or 3×3 grid. Return its index and flip direction.
2
2
2
2
Multiscale matching
Match a SAR or optical query resized by 0.5× or 1.5× to three same-modality candidate regions. Return the matching index.
4
2
2
4
Subtotal
24
16
16
24
Appendix
Table S2: Tasks used in the main post pretraining, with their inputs, required operation, and answer form. Var. counts distinct configured variants; T, K, and S count active variants for TerraMesh, KuroSiwo, and SARLO. The 14 task groups contain 87 variants and 131 source–task combinations. The two families match the taxonomy used in the sampling analysis.
Figure S1 : Representative self-supervised questions from real KuroSiwo and SARLO observations. (a) The known patch permutation determines the answer. (b) The fraction of native power carried by each quadrant’s brightest 1% of pixels determines its scattering concentration. TL, TR, BL, and BR denote the top-left, top-right, bottom-left, and bottom-right quadrants. (c) Paired crop coordinates identify the optical region containing the SAR query. Grids, labels, and highlighted answers are explanatory overlays.
Vision foundation model adaptation tailors pretrained image and video models to new tasks and domains while balancing accuracy, efficiency, and generalization.
School of Engineering, University of Edinburgh, United Kingdom · Leonardo UK, United Kingdom · School of Engineering and Physical Sciences, Heriot-Watt University, United Kingdom