Pretrained foundation models (FMs) have achieved remarkable success in computer vision, yet their high fine-tuning cost limits practical deployment. Parameter-efficient fine-tuning (PEFT) methods such as Low-Rank Adaptation (LoRA) improve efficiency by constraining updates to a predefined low-rank subspace. However, when applied to remote sensing tasks with substantial domain shifts, the fixed subspace is constructed without observing the downstream activation distribution and can therefore provide a poor coordinate system for adaptation, a phenomenon herein termed subspace mismatch. To address this issue, a unified framework is introduced, termed Domain-aware Relaxed Orthogonal Subspace adaptation (DROS), which reformulates low-rank adaptation as data-conditioned subspace learning and flexible subspace adaptation. Specifically, the weight decomposition is conditioned on second-order activation statistics estimated from the downstream training distribution, so that the initialization reflects the feature geometry actually induced by the remote-sensing data, followed by flexible geometric transformations enabled by a relaxed orthogonal parameterization. Furthermore, the framework is extended to multimodal settings (MM-DROS) by sharing transformation structures across modality-specific subspaces, facilitating efficient cross-modal interaction. Extensive experiments on multiple remote sensing benchmarks demonstrate that DROS achieves state-of-the-art performance, even surpassing full fine-tuning, without additional inference overhead.
Figures & tables
Fig. 1: Conceptual illustration of subspace mismatch in parameter-efficient fine-tuning (PEFT) under domain shifts. (a) LoRA [ 17 ] constructs a low-rank update within a randomly initialized subspace without using downstream activation statistics. (b) PiSSA [ 18 ] derives the adaptation subspace from pretrained weights, improving optimization efficiency but remaining independent of the downstream activation distribution. (c) DROS conditions the subspace initialization on second-order activation statistics from representative training data and further enables structured transformations within this data-conditioned subspace.
Fig. 2: Overview of DROS. Panel a presents the overall framework, where a trainable DROS side branch is attached to each of the n adapted linear layers in a frozen foundation model. Panel b details the construction for layer k , where k∈{1,…,n} . The frozen backbone collects Hk,i from each of N representative training samples, and averaging the products Hk,iHk,i⊤ gives Ck . The index k is omitted after a single layer is fixed. Covariance-conditioned SVD of Wrw=WpreC selects the bottom- r components to initialize B0 and A0 . The pretrained weight is split as Wpre=Wres+B0A0 , and the trainable side block Wside=BTA yields Wtuned=Wres+Wside .
Fig. 3: Overview of MM-DROS at one adapted linear layer. The layer index k is suppressed. The optical and SAR branches retain modality-specific residual weights and bases Bopt,Aopt and Bsar,Asar . The relaxed operator Tshared=diag(α)Odiag(β) is shared across the two branches. Each transformed low-rank branch is added to its corresponding residual weight.
Task Type
Dataset
Source (Region, Reported Dataset Year)
Resolution
Classes
Train Size
Test Size
Total Images
Classification
EuroSAT
Sentinel-2, Europe (2017)
10m
10
16,200
10,800
27,000
RESISC45
Google Earth, Global (2015)
0.3m
45
18,900
6,300
25,200
UCM
Aerial, California, US (2010)
0.3m
21
1,260
420
1,680
Semantic Segmentation
Cashew
Sentinel-2, Benin (GEO-Bench)
10m
7
1,350
400
1,750
Chesap.
RGBN, Chesapeake Bay, US (GEO-Bench)
1.0m
7
3,000
1,000
4,000
Cattle
RGB, New Zealand (GEO-Bench)
0.1m
2
524
66
590
TABLE I: Detailed statistics of the experimental datasets. For classification and GEO-Bench tasks, the official data splits provided by TorchGeo and GEO-Bench are adopted. For DGSS tasks, the number of cropped 512×512 patches used for training and evaluation is reported. The data source/sensor, covered region, and reported dataset year of each dataset are also given.
Method
#Params
MARS
#Params
DINOv2-L
EuroSAT
UCM
RESISC45
Avg
EuroSAT
UCM
RESISC45
Avg
FT
89.00M
96.47 ± 0.31
98.57 ± 0.00
96.85 ± 0.17
97.30
304.20M
96.63 ± 0.12
98.41 ± 0.69
95.25 ± 0.18
96.76
FZ
0.00M
86.83 ± 0.17
93.17 ± 0.11
87.84 ± 0.04
89.28
0.00M
86.13 ± 0.49
90.00 ± 1.04
84.23 ± 0.12
86.79
LoRA
0.28M
95.77 ± 0.29
98.73 ± 0.36
96.49 ± 0.10
97.00
0.59M
96.83 ± 0.31
98.10 ± 0.41
95.33 ± 0.20
96.75
Earth-Adapter
1.55M
96.50 ± 0.10
98.69 ± 0.11
96.75 ± 0.05
97.31
9.59M
96.37 ± 0.12
98.41 ± 1.10
95.62 ± 0.19
96.80
VPT
0.24M
86.97 ± 0.06
92.54 ± 0.14
87.80 ± 0.06
89.10
0.49M
96.43 ± 0.25
97.86 ± 0.24
95.12 ± 0.06
96.47
TABLE II: Comparative results on classification tasks using MARS and DINOv2-L backbones. Red and blue indicate the best and second-best results, respectively.
Fig. 4: Qualitative comparison of Top-5 classification predictions on RESISC45 using DINOv2-L. Red bars identify the ground-truth category, and Ours denotes DROS.
Method
#Params
MARS
#Params
DINOv2-L
Cashew
Chesap.
Cattle
Avg
Cashew
Chesap.
Cattle
Avg
FT
89.00M
78.30 ± 0.20
75.53 ± 0.75
83.69 ± 0.19
79.17
304.20M
62.65 ± 1.69
74.59 ± 0.18
83.18 ± 0.54
73.47
FZ
0.00M
69.05 ± 1.06
74.15 ± 0.37
84.02 ± 0.48
75.74
0.00M
55.68 ± 4.26
73.35 ± 0.16
82.55 ± 0.14
70.53
LoRA
3.03M
74.82 ± 1.13
75.25 ± 0.17
83.82 ± 0.11
77.96
6.21M
63.53 ± 1.62
75.13 ± 0.18
83.10 ± 0.31
73.92
Earth-Adapter
1.55M
70.29 ± 0.80
74.56 ± 0.41
83.57 ± 0.10
76.14
9.59M
58.98 ± 1.80
74.04 ± 0.16
82.89 ± 0.31
71.97
VPT
3.08M
69.83 ± 0.30
74.59 ± 0.24
84.03 ± 0.56
76.15
6.29M
50.51 ± 1.46
74.22 ± 0.26
82.94 ± 0.55
69.22
TABLE III: Comparative results on segmentation tasks using MARS and DINOv2-L backbones. For clarity, dataset names are abbreviated: Cashew : m-cashew-plant; Chesap. : m-chesapeake; Cattle : m-nz-cattle. Red and blue indicate the best and second-best performance, respectively.
Fig. 5: Qualitative comparison on Chesapeake semantic segmentation using DINOv2-L.
Method
#Params
MARS
#Params
DINOv2-L
U2R
R2U
P2V
V2P
Avg
U2R
R2U
P2V
V2P
Avg
FT
89.00M
38.82 ± 0.76
52.90 ± 0.54
41.22 ± 0.56
40.78 ± 0.94
43.43
304.20M
45.55 ± 0.82
58.38 ± 0.39
60.72 ± 1.31
55.57 ± 0.88
55.05
FZ
0.00M
36.69 ± 0.62
50.87 ± 1.16
37.32 ± 0.79
32.79 ± 1.15
39.42
0.00M
45.23 ± 1.40
57.37 ± 0.04
55.74 ± 1.14
49.93 ± 1.28
52.07
LoRA
3.03M
39.67 ± 0.60
53.84 ± 0.08
41.32 ± 1.63
39.39 ± 0.70
43.56
6.21M
45.27 ± 1.11
58.02 ± 0.30
60.76 ± 0.65
54.14 ± 0.86
54.55
Earth-Adapter
1.55M
40.33 ± 1.25
52.34 ± 0.70
39.31 ± 0.93
40.59 ± 0.53
43.14
9.59M
45.99 ± 0.94
57.72 ± 0.35
59.49 ± 1.00
55.66 ± 1.51
54.72
VPT
0.77M
36.78 ± 1.69
51.47 ± 1.17
37.51 ± 0.28
35.51 ± 0.19
40.32
1.57M
43.51 ± 1.99
57.46 ± 0.10
58.27 ± 0.49
55.75 ± 0.93
53.75
TABLE IV: Domain generalization results (mIoU %) on MARS and DINOv2-L backbones. Tasks include cross-city ( P2V , V2P ) and cross-scene ( U2R , R2U ) transfers. Red and blue indicate the best and second-best results, respectively.
Fig. 6: Qualitative comparison on the LoveDA U2R setting using DINOv2-L.
Method
#Params
MARS
#Params
DINOv2-L
BRIGHT
WHU
Avg
BRIGHT
WHU
Avg
FT
178.00M
63.78 ± 0.36
58.59 ± 0.59
61.19
608.40M
65.28 ± 0.62
60.36 ± 0.05
62.82
FZ
0.00M
55.22 ± 1.35
59.22 ± 0.08
57.22
0.00M
62.08 ± 1.69
59.80 ± 0.07
60.94
LoRA
6.06M
63.29 ± 0.61
59.89 ± 0.27
61.59
12.42M
64.88 ± 1.48
60.32 ± 0.33
62.60
Earth-Adapter
3.11M
62.62 ± 1.61
59.79 ± 0.06
61.20
19.18M
65.08 ± 0.87
60.32 ± 0.15
62.70
VPT
6.16M
54.65 ± 1.06
59.14 ± 0.13
56.89
12.58M
64.11 ± 1.20
60.25 ± 0.19
62.18
TABLE V: Multimodal segmentation results (mIoU %) on MARS and DINOv2-L backbones. For clarity, dataset names are abbreviated: WHU : WHU-OPT-SAR. Red and blue indicate the best and second-best results, respectively.
Fig. 7: Qualitative comparison on BRIGHT multimodal segmentation using DINOv2-L.
Component
RESISC45
Chesap.
U2R
Avg
Dominant (Top- r )
25.21 ± 0.32
70.32 ± 0.59
40.97 ± 1.19
45.50
Minor (Bot- r )
95.37 ± 0.10
75.58 ± 0.13
45.77 ± 0.61
72.24
TABLE VI: Mechanism verification of spectral component selection on DINOv2-L. Adapting the top- r principal components collapses classification accuracy and degrades dense tasks, whereas adapting the bottom- r components remains stable.
Subset Ratio
RESISC45
Chesap.
U2R
Avg
1%
94.80 ± 0.77
74.89 ± 0.45
45.48 ± 0.80
71.72
10%
95.37 ± 0.10
75.58 ± 0.13
45.77 ± 0.61
72.24
50%
94.74 ± 0.27
74.91 ± 0.20
45.59 ± 0.53
71.75
100%
94.63 ± 0.59
74.96 ± 0.08
45.08 ± 0.89
71.56
TABLE VII: Ablation of training subset ratio for domain-aware covariance estimation on DINOv2-L.
Fig. 8: Comparison of different subspace tuning principles: LoRA performs fine-tuning by randomly assigning subspaces, PiSSA restricts subspaces to the principal components of weights. DROS achieves optimal matching.
Method
RESISC45
Chesap.
U2R
Avg
FT
88.32 ± 0.71
69.72 ± 0.26
42.62 ± 0.59
66.89
FZ
70.00 ± 0.06
68.37 ± 0.32
42.23 ± 0.57
60.20
LoRA
87.19 ± 0.65
69.69 ± 0.79
42.60 ± 0.20
66.49
Earth-Adapter
88.68 ± 0.27
69.44 ± 0.63
42.20 ± 0.69
66.77
VPT
82.79 ± 2.08
68.45 ± 0.73
41.88 ± 0.14
64.37
PiSSA
86.21 ± 1.62
68.91 ± 1.13
42.49 ± 0.85
65.87
TABLE VIII: Comparative results under 10% training data on DINOv2-L.
Operator
RESISC45
Chesap.
U2R
Avg
None (w/o operator)
95.18 ± 0.10
74.80 ± 0.19
44.52 ± 0.23
71.50
O
95.28 ± 0.10
75.07 ± 0.10
44.78 ± 0.61
71.71
O + α
95.30 ± 0.05
74.91 ± 0.19
44.12 ± 0.98
71.44
O + β
95.33 ± 0.23
75.17 ± 0.11
45.61 ± 0.35
72.04
O + α + β
95.37 ± 0.10
75.58 ± 0.13
45.77 ± 0.61
72.24
TABLE IX: Ablation study of the relaxed orthogonal operator components on DINOv2-L.
Shared Operator
BRIGHT
None (non-shared)
65.41 ± 0.84
O
65.46 ± 0.87
O + α
65.47 ± 0.47
O + β
65.50 ± 0.71
O + α + β
65.60 ± 0.84
TABLE X: Ablation study of the shared subspace in MM-DROS on BRIGHT.
Fig. 9: Qualitative case study of MM-DROS on the unseen Los Angeles wildfire scenario.
Fig. 10: Comparison of loss curves and accuracy curves for different PEFT methods.
Method
RESISC45
Chesap.
U2R
#Params
Mem. (GB)
#Params
Mem. (GB)
#Params
Mem. (GB)
FT
304.20M
5.21
304.20M
6.41
304.20M
13.71
FZ
0.00M
1.24
0.00M
2.42
0.00M
4.65
LoRA
0.59M
2.75
6.21M
4.02
6.21M
10.77
Earth-Adapter
9.59M
3.18
9.59M
4.35
9.59M
11.24
VPT
0.49M
2.68
6.29M
3.95
1.57M
10.58
TABLE XI: Efficiency comparison on DINOv2-L across different methods. The number of learnable parameters (#Params) and the peak GPU memory in GB are reported for each task.
Vision foundation model adaptation tailors pretrained image and video models to new tasks and domains while balancing accuracy, efficiency, and generalization.
Remote-sensing foundation models (RSFMs) benefit from pretraining on imagery from multiple sensors and ground sampling distances (GSDs), but such exposure alone does not resolve scale mismatch during downstream adaptation. A fixed token-grid offset can correspond to different ground distances across sensors, making grid-based positional priors physically inconsistent. Meanwhile, heterogeneous spatial granularity means that compact urban regions and homogeneous landscapes may require different positional sensitivities even under the same GSD. Therefore, we propose {GeoRoPE}, a ground-aware, RoPE-compatible, and parameter-efficient spatial adaptation method for RSFMs. GeoRoPE recalibrates token-level positional interactions from two complementary aspects. First, Geo-Coordinate Calibration (GCC) rescales raw token-grid offsets according to the ground distance represented by one token-grid step, producing geo-calibrated relative coordinates across GSDs. Second, Geo-Frequency Calibration (GFC) adjusts the native RoPE frequency with a relation-specific factor, enabling position sensitive adaptation to scene-dependent spatial granularity. GeoRoPE is injected into pretrained RSFMs through a lightweight adapter, preserving the frozen spatial prior while adding geo-aware positional corrections. Experiments across multiple RSFMs, sensors, resolutions, and downstream tasks demonstrate that GeoRoPE improves cross-resolution robustness and scale-sensitive representation learning.
Yu Luo, Kun Hu, Mengwei He +7
The University of Sydney · Edith Cowan University · Adelaide University +3
Foundation models (FMs) have transformed machine learning from isolated task-specific model development toward general-purpose models pretrained on broad data and adapted to multiple downstream tasks. Earth observation (EO) is an important domain for this paradigm because satellite and airborne archives are large, high-revisit, and increasingly multimodal, while reliable field labels are often sparse. Remote sensing foundation models (RSFMs) cannot be transferred reliably/optimally without domain-specific adaptation. This is because EO data are governed by measurement physics and operational decision constraints. This chapter reviews the design principles arising from these domain-specific constraints. It first defines the FMs paradigm in remote sensing (RS), then synthesizes the current model landscape, pretraining objectives, architecture designs, downstream adaptation and trustworthiness requirements. The chapter also incorporates recent benchmark evidence showing that no single geospatial foundation model is universally best and that inconsistent evaluation remains a major issue to fair comparison and reliable deployment. In addition, two brief environmental monitoring case studies; physics-informed spectral targeted masking for harmful algal bloom prediction and reinforcement learning for adaptive environmental monitoring station selection to illustrate the FMs domain-guided principles in practice. This chapter posits that next-generation RSFMs should be evaluated not only by benchmark accuracy, but also by modality-aware transfer and physically plausible representations for trustworthy EO decisions.
Syed Usama Imtiaz, Mitra Nasr Azadani, Nasrin Alamdari
Department of Civil and Environmental Engineering, Florida State University
Remote sensing imagery suffers from clouds, haze, noise, resolution limits, and sensor heterogeneity. Existing restoration and fusion approaches train separate models per degradation type. In this work, we present Language-conditioned Large-scale Remote Sensing restoration model (LLaRS), the first unified foundation model for multi-modal and multi-task remote sensing low-level vision. LLaRS employs Sinkhorn-Knopp optimal transport to align heterogeneous bands into semantically matched slots, routes features through three complementary mixture-of-experts layers (convolutional experts for spatial patterns, channel-mixing experts for spectral fidelity, and attention experts with low-rank adapters for global context), and stabilizes joint training via step-level dynamic weight adjustment. To train LLaRS, we construct LLaRS1M, a million-scale multi-task dataset spanning eleven restoration and enhancement tasks, integrating real paired observations and controlled synthetic degradations with diverse natural language prompts. Experiments show LLaRS consistently outperforms seven competitive models, and parameter-efficient finetuning experiments demonstrate strong transfer capability and adaptation efficiency on unseen data. Repo: https://github.com/yc-cui/LLaRS
Yongchuan Cui, Peng Liu
Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China · School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing, China