Organizations: National Key Laboratory of Microwave Imaging Technology, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China, and also with the School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100049, China.
Synthetic aperture radar (SAR) object detection is an important part of remote sensing interpretation. However, because of variations in frequency band, resolution, background clutter, and target scattering responses, the performance of existing detectors often degrades when training and testing data are acquired from different SAR domains. Although domain adaptation methods offer a promising paradigm for solving this problem, most of them mainly pursue domain-invariant feature alignment and suppress sensor-dependent scattering characteristics that are useful for object detection. This problem becomes more challenging in few-shot scenarios, where only a few fully annotated target-domain SAR images are available. To address this issue, we propose a scattering-aware shared-specific feature decomposition framework for few-shot SAR domain adaptation object detection. We decompose detection features into a shared path and several soft-gated scattering-specific expert paths. The shared path learns transferable object structural information and is used for asymmetric domain alignment, while the scattering-specific experts adaptively compensate heterogeneous SAR responses. In addition, routing-domain auxiliary loss is introduced to encourage specific experts to capture sensor-dependent routing preferences, and an expert balancing loss is used to prevent routing collapse. Extensive experiments on four bidirectional heterogeneous SAR detection tasks between FARAD-X/FARAD-Ka and MiniSAR under different few-shot settings have been conducted and experimental results demonstrate that the proposed method achieves superior performance in both forward and reverse adaptation directions.
Figures & tables
Fig. 1: Comparison between optical and cross-sensor SAR domain shifts. (a) Cityscapes and Foggy Cityscapes illustrate low-level visual variations caused by fog, which provide limited target-discriminative cues. (b) FARAD and MiniSAR show cross-sensor SAR discrepancies in clutter and target scattering responses. (c) Enlarged vehicle patches highlight shared target structures with green boxes and domain-specific scattering cues with blue dots.
Fig. 2: Overall framework of the proposed method. Source and few-shot target SAR images are fed into a shared feature extractor, followed by an MoE feature learning module composed of an always-on shared path and router-weighted scattering-specific experts. The shared representation is used for asymmetric domain alignment, while the aggregated expert features provide domain-sensitive compensation for object detection.
Fig. 3: Motivating t-SNE visualization of source and target feature distributions under the 1-shot FARAD-Ka → MiniSAR setting. AcroFOD shows limited source–target overlap, while AsyFOD produces a more concentrated target cluster but still leaves many source samples poorly matched.
Fig. 4: Examples of SAR images and corresponding optical scenes from the datasets used in our experiments. (a), (c), and (e) are SAR images from MiniSAR, FARAD-X, and FARAD-Ka, respectively. (b), (d), and (f) show the corresponding optical scenes on Google Earth.
Domain
FARAD-X
FARAD-Ka
MiniSAR
Band
X
Ka
Ku
Time
2015.10
2015.08
2005.05
Bandwidth
3 GHz
5 GHz
3 GHz
Center Freq.
9.6 GHz
35.6 GHz
16.8 GHz
Angle
26∘ – 34∘
26∘ – 34∘
26∘ – 29∘
Resolution
0.1 m
0.1 m
0.1 m
TABLE I: Imaging parameters and domain characteristics of the datasets.
Methods
(a) FARAD-X → MiniSAR
(b) FARAD-Ka → MiniSAR
1-shot
3-shot
5-shot
8-shot
10-shot
1-shot
3-shot
5-shot
8-shot
10-shot
Source-only [ 43 ]
22.46
25.77
FSDet [ 45 ]
19.38
30.42
36.84
41.01
47.69
27.72
37.33
44.07
53.36
45.26
SSDA-YOLO [ 42 ]
23.13
28.48
39.04
40.08
39.45
29.17
39.68
46.67
46.56
50.69
AcroFOD [ 46 ]
27.35
36.20
51.81
61.29
64.53
30.98
41.30
55.56
64.08
66.18
AsyFOD [ 17 ]
29.09
36.67
49.60
56.58
63.37
37.43
36.90
52.78
60.91
58.62
TABLE II: Comparison of different methods on bidirectional few-shot SAR domain adaptation tasks in terms of mAP 50 (%).
Fig. 5: Qualitative comparison on MiniSAR test images under the 5-shot FARAD-X → MiniSAR adaptation task. Each column represents one test scene, while the rows from top to bottom correspond to the ground truth, AcroFOD, AsyFOD † , and the proposed method. Green and red boxes denote ground-truth annotations and detection results, respectively.
Fig. 6: Visualization of P3 shared and expert activation responses on MiniSAR test images under the 5-shot FARAD-X → MiniSAR setting. From left to right: original image, shared path, and Experts 1–4.
Adaptation Task
Level
Expert 1
Expert 2
Expert 3
Expert 4
TVD (%)
Entropy
FARAD-X → MiniSAR
P3
23.01/22.78
28.47/29.02
24.44/24.23
24.08/23.97
0.55
0.9973/0.9966
P4
24.30/24.94
24.30/23.89
26.62/28.08
24.79/23.09
2.10
0.9993/0.9977
FARAD-Ka → MiniSAR
P3
22.97/22.76
26.73/26.87
27.12/27.35
23.18/23.02
0.37
0.9977/0.9973
P4
25.89/25.50
24.92/24.90
25.30/25.16
23.88/24.44
0.56
0.9996/0.9997
TABLE III: Domain-wise router distribution statistics on the source and target image sets. Each expert entry reports the source/target mean routing weights (%). TVD denotes the total variation distance between the source and target mean routing distributions, and entropy denotes the normalized source/target routing entropy.
Fig. 7: t-SNE visualization of source-domain and target-domain shared feature distributions under the 1-shot FARAD-Ka → MiniSAR adaptation task. Blue circles and orange triangles denote source-domain and target-domain samples, respectively. (a) AcroFOD. (b) AsyFOD † . (c) Ours. The proposed method shows a closer source–target distribution in the shared feature space, indicating more effective transferable representation learning for few-shot SAR domain adaptation.
Methods
FARAD-X → MiniSAR
FARAD-Ka → MiniSAR
1-shot
3-shot
5-shot
8-shot
10-shot
Avg.
1-shot
3-shot
5-shot
8-shot
10-shot
Avg.
AsyFOD † baseline
43.39
49.09
66.10
72.07
71.23
60.38
41.30
43.30
66.01
69.72
73.21
58.71
Fused-feature alignment
40.41
46.17
66.73
70.10
66.82
58.05
42.64
47.30
66.58
71.98
71.94
60.09
Shared only
46.16
50.43
58.57
74.19
71.51
60.17
41.67
46.83
61.69
74.15
74.73
59.81
w/o Lalignsh
28.37
40.61
55.59
63.79
59.31
49.53
36.50
42.85
54.19
63.74
64.92
52.44
w/o Lrd
46.32
50.17
63.81
73.92
72.71
61.39
41.05
51.67
68.01
69.55
73.10
60.68
TABLE IV: Ablation results of the proposed components in terms of mAP 50 (%).
Experts E
FARAD-X → MiniSAR
Avg.
1-shot
3-shot
5-shot
8-shot
10-shot
1
44.97
52.21
68.35
70.49
68.77
60.96
2
44.54
53.48
68.41
72.14
75.46
62.81
4
47.19
51.19
68.98
76.06
73.58
63.40
8
43.12
50.04
67.68
73.35
77.48
62.33
TABLE V: Effect of the number of scattering-specific experts on FARAD-X → MiniSAR in terms of mAP 50 (%).
Experts E
FARAD-Ka → MiniSAR
Avg.
1-shot
3-shot
5-shot
8-shot
10-shot
1
44.17
54.18
70.73
76.77
75.52
64.27
2
49.61
46.68
71.06
78.48
76.75
64.52
4
47.26
51.60
69.32
75.72
75.42
63.86
8
43.08
49.85
67.52
77.06
78.07
63.12
TABLE VI: Effect of the number of scattering-specific experts on FARAD-Ka → MiniSAR in terms of mAP 50 (%).