Label Less, Learn More: Resource-Efficient Active Semi-Supervised Learning for Onboard Satellite Image Annotation

Ahmed Abdelnaby; Mohamed Elmahallawy; Marius Bernahrndt; Tobias Hecking

cs.CVSep 25, 2026

Label Less, Learn More: Resource-Efficient Active Semi-Supervised Learning for Onboard Satellite Image Annotation

Authors: Ahmed Abdelnaby, Mohamed Elmahallawy, Marius Bernahrndt, Tobias Hecking

Organizations: School of Engineering and Applied Science, Washington State University, Richland, WA 99354, USA · Institute for Software Technology, German Aerospace Center (DLR), Porz, 51147 Cologne, Germany

Abstract

Large-scale pervasive sensing increasingly relies on high-resolution satellite imagery, yet task-specific onboard vision is constrained by costly annotation and limited computation, memory, energy, and communication resources. Existing approaches largely rely on either data-hungry supervised learning or large vision-language foundation models, limiting efficient adaptation and deployment under these constraints. We present SatLabel, a resource-aware learning framework that transforms limited satellite labels into progressively refined onboard models through adaptive sample acquisition and semi-supervised model adaptation. Rather than repeatedly training on uniformly sampled labels, SatLabel closes the loop between model uncertainty, class imbalance, and pseudo-label quality to selectively acquire informative samples while exploiting abundant unlabeled imagery. This enables a compact student to adapt to target sensing domains with reduced annotation and inference costs. We further introduce an optional Mixture-of-Experts (MoE) student with graph-based feature refinement to enhance representation capacity while retaining a lightweight footprint. We evaluate SatLabel on 11 remote-sensing datasets spanning core, extended, and unseen domains against RemoteCLIP zero-shot inference. SatLabel improves Macro-F1 on most core and extended datasets while maintaining strong cross-dataset transfer to unseen domains. More importantly, the Balanced student contains only 11.2 M parameters and occupies approximately 42.8 MB, compared with 151.3M parameters and 577 MB for RemoteCLIP, while requiring 3.65 versus 5.89 GFLOPs. Across four efficiency benchmarks, it achieves approximately 2x higher GPU-forward throughput and reduces energy per image on datasets.

Figures & tables

Fig. 1 : Overview of the proposed SatLabel methodology. The teacher–student framework integrates pseudo-labeling, consistency regularization, and uncertainty-guided active learning to progressively improve the student model.

Fig. 2 : An illustration of the adaptive student model architecture. The ResNet-18 backbone is augmented with a Mixture-of-Heads module, adaptive routing, and BatchGCN for efficient feature refinement, followed by a classification head for final prediction.

Dataset	RemoteCLIP	RS-M-CLIP	GeoRSCLIP	SkyCLIP	SatLabel
AID	86.83	88.90	76.33	70.89	94.41 (MoE+GCN)
EuroSAT	32.46	25.85	67.47	71.70	95.37 (Balanced)
UC Merced	77.95	91.50	74.50	74.80	91.00 (MoE+GCN)
RESISC45	68.38	92.62	73.83	70.94	86.63 (Balanced)
PatternNet	57.52	50.95	74.80	80.88	95.90 (Balanced)
RSI-CB256	50.38	47.50	61.00	50.09	87.01 (Balanced)

TABLE I : Accuracy comparison of SatLabel against baselines on the core and extended datasets. Best results are shown in bold and second-best results are underlined . Values are reported in percent. The student variant used by SatLabel is indicated in parentheses.

Fig. 3 : RemoteCLIP zero-shot Macro-F1 compared with the best balanced student model on the core and extended datasets. The student outperforms RemoteCLIP on most datasets, while RemoteCLIP remains strongest on WHU-RS19.

Model	Params (M) $\downarrow$	Size (MB) $\downarrow$	FLOPs (G) $\downarrow$
RemoteCLIP ViT-B/32	151.28	577.21	5.89
RS-M-CLIP	366.12	1396.81	5.89
GeoRSCLIP ViT-B/32	151.28	577.21	5.89
SkyCLIP ViT-B/32	151.28	577.21	5.89
SatLabel (Balanced)	11.21	42.83	3.65
SatLabel (MoE+GCN)	20.40	77.93	3.67

TABLE II : Comparison of SatLabel with baselines in terms of parameters, model size and computational complexity.

Model	Images/s $\uparrow$	Energy (J/img) $\downarrow$	Power (W) $\downarrow$	Peak GPU (MB) $\downarrow$
RemoteCLIP ViT-B/32	12128.4	0.128	$143.19\pm 62.97$	741.9
RS-M-CLIP	12211.2	0.124	$140.00\pm 59.80$	1560.4
GeoRSCLIP ViT-B/32	11692.5	0.127	$142.58\pm 65.30$	760.7
SkyCLIP ViT-B/32	12246.8	0.130	$136.24\pm 59.23$	741.9
SatLabel (Balanced)	24353.8	0.106	$\mathbf{110.55\pm 30.50}$	381.1
SatLabel (MoE+GCN)	12623.0	0.123	$132.34\pm 47.20$	575.4

TABLE III : Comparison of SatLabel with baselines in terms of runtime and resource efficiency on AID (10,000 images).

Model	Images/s $\uparrow$	Energy (J/img) $\downarrow$	Power (W) $\downarrow$	Peak GPU (MB) $\downarrow$
RemoteCLIP ViT-B/32	11966.9	0.055	$291.88\pm 46.42$	741.9
RS-M-CLIP	12028.5	0.055	$295.95\pm 31.00$	1560.4
GeoRSCLIP ViT-B/32	11712.4	0.056	$299.99\pm 46.62$	760.6
SkyCLIP ViT-B/32	12111.9	0.055	$297.14\pm 40.29$	741.9
SatLabel (Balanced)	24143.0	0.036	$\mathbf{185.11\pm 27.21}$	381.1
SatLabel (MoE+GCN)	12607.4	0.052	$268.27\pm 27.63$	575.4

TABLE IV : Comparison of SatLabel with baselines in terms of runtime and resource efficiency on EuroSAT (27,000 images).

Model	Images/s $\uparrow$	Energy (J/img) $\downarrow$	Power (W) $\downarrow$	Peak GPU (MB) $\downarrow$
RemoteCLIP ViT-B/32	11992.8	0.064	$232.96\pm 50.63$	741.9
RS-M-CLIP	11852.3	0.068	$222.07\pm 46.29$	1560.4
GeoRSCLIP ViT-B/32	11711.4	0.065	$245.00\pm 53.79$	760.7
SkyCLIP ViT-B/32	12012.8	0.064	$241.46\pm 47.66$	741.9
SatLabel (Balanced)	24282.9	0.044	$\mathbf{160.21\pm 31.45}$	381.2
SatLabel (MoE+GCN)	12598.1	0.059	$226.55\pm 35.38$	575.5

TABLE V : Comparison of SatLabel with baselines in terms of runtime and resource efficiency on PatternNet (30,400 images).

Model	Images/s $\uparrow$	Energy (J/img) $\downarrow$	Power (W) $\downarrow$	Peak GPU (MB) $\downarrow$
RemoteCLIP ViT-B/32	12029.0	0.097	$394.36\pm 5.35$	741.9
RS-M-CLIP	11891.9	0.103	$408.85\pm 2.68$	1560.4
GeoRSCLIP ViT-B/32	11585.5	0.107	$418.21\pm 16.80$	760.7
SkyCLIP ViT-B/32	12185.9	0.105	$402.34\pm 15.96$	741.9
SatLabel (Balanced)	24292.2	0.100	$\mathbf{254.30\pm 24.24}$	381.3
SatLabel (MoE+GCN)	12677.1	0.129	$351.10\pm 10.88$	575.6

TABLE VI : Comparison of SatLabel with baselines in terms of runtime and resource efficiency on OPTIMAL31 (1,860 images).

Dataset	Cov.	RC F1	Stud. Acc.	Stud. F1	Best setup
OPTIMAL31	100.00	75.98	92.63	93.56	06 Mix, Balanced
RSC11	90.91	60.42	56.97	71.97	06 Mix, MoE+GCN
RS2800	100.00	62.99	48.82	54.70	04 WHU, MoE+GCN
SIRI-WHU	100.00	59.07	47.33	51.14	01 Base, MoE+GCN

TABLE VII : Best student results on unseen datasets using canonical class mapping. Values are reported in percent.

	RemoteCLIP		Student ResNet-18		Student ResNet-50
Dataset	Acc.	F1	Acc.	F1	Acc.	F1
AID	86.83	86.04	94.41	94.76	94.53	95.02
EuroSAT	32.46	29.87	95.37	95.95	94.87	96.29
UC Merced	77.95	74.63	91.00	94.01	80.52	86.09
RESISC45	68.38	66.53	86.63	88.24	88.38	90.34
PatternNet	57.52	52.15	95.90	96.26	93.24	93.41

TABLE VIII : RemoteCLIP zero-shot performance compared with knowledge-distilled ResNet-18 and ResNet-50 student models. Best student results are shown in bold . Values are reported in percent.

Configuration	(Accuracy, %)	$\Delta$
(1) ResNet-18 supervised	89.4	—
(2) +KD from RemoteCLIP	91.2	+1.8
(3) +Consistency (FixMatch)	92.3	+1.1
(4) +Active learning (random) †	92.6	+0.3
(5) +Active learning (uncertainty)	93.1	+0.5
(6) +MoE (softmax gating)	93.3	+0.2

TABLE IX : Cumulative component analysis on AID (30 % labeled).

Panel of Figure 13 — Fig. 4 : t-SNE projections of student features on unseen datasets.

Panel of Figure 14 — Fig. 5 : Normalized confusion matrices on all benchmarks.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Group	Dataset	Classes	Images	Min	Max	Mean
Core	AID	30	10,000	220	420	333
Core	EuroSAT	10	27,000	2000	3000	2700
Core	UC Merced	21	2,100	100	100	100
Extended	RESISC45	45	31,500	700	700	700
Extended	PatternNet	38	30,400	800	800	800
Extended	RSI-CB256	7	24,747	884	6258	3535

Appendix

TABLE X : Dataset statistics. Support denotes images per class (min / max / mean).

Experiment	Training domains
01 Baseline	AID, UC Merced, EuroSAT
02 + RESISC45	AID, UC Merced, EuroSAT, RESISC45
03 + PatternNet	AID, UC Merced, EuroSAT, PatternNet
04 + WHU-RS19	AID, UC Merced, EuroSAT, WHU-RS19
05 + RSI-CB256	AID, UC Merced, EuroSAT, RSI-CB256
06 All Eval Mix	AID, UC Merced, EuroSAT, RESISC45, PatternNet, WHU-RS19, RSI-CB256

Appendix

TABLE XI : Student training configurations used in the ablation study.

Group	Dataset	RC Acc.	Stud. Acc.	Stud. F1	$\Delta$ Acc.
Core	AID	86.83	94.41	94.76	+7.58
Core	EuroSAT	32.46	95.37	95.95	+62.91
Core	UC Merced	77.95	91.00	94.01	+13.05
Extended	RESISC45	68.38	86.63	88.24	+18.25
Extended	PatternNet	57.52	95.90	96.26	+38.38
Extended	RSI-CB256	50.38	87.01	91.36	+36.63

Appendix

TABLE XII : ResNet-18 MoE+GCN student – best checkpoint results on all eleven datasets. Acc: accuracy. F1: Macro-F1.

\Delta

: student minus RemoteCLIP (Accuracy).

Dataset	RemoteCLIP	Student (R18)	$\Delta$ Acc.
AID	86.83 / 86.06	94.41 / 94.76	+7.58
EuroSAT	32.46 / 29.87	95.37 / 95.95	+62.91
UC Merced	77.95 / 74.63	91.00 / 94.01	+13.05
RESISC45	68.38 / 66.53	86.63 / 88.24	+18.25
PatternNet	57.52 / 52.15	95.90 / 96.26	+38.38
RSI-CB256	50.38 / 42.60	87.01 / 91.36	+36.63

Appendix

TABLE XIII : RemoteCLIP zero-shot vs. best ResNet-18 MoE+GCN student (Accuracy / Macro-F1, %).

\Delta

: student minus RemoteCLIP (Accuracy). Bold : best per dataset.

Class	Support	Precision	Recall	F1-Score
Forest	250	99.6	99.6	100.0
Beach	400	99.0	99.2	99.7
Meadow	280	99.3	98.9	99.6
Parking	390	98.5	99.2	99.4
Mountain	340	98.8	98.5	99.2
Viaduct	420	97.2	100.0	99.1

Appendix

TABLE XIV : AID – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Class	Support	Precision	Recall	F1-Score
Residential	3000	99.5	99.0	98.5
Industrial	2500	98.9	99.4	98.5
Sea Or Lake	3000	98.8	99.0	98.2
Forest	3000	98.2	98.9	97.9
Highway	2500	98.4	94.6	95.8
River	2500	93.4	98.8	95.3

Appendix

TABLE XV : EuroSAT – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Class	Support	Precision	Recall	F1-Score
Beach	100	100.0	100.0	99.4
Harbor	100	99.0	100.0	98.9
Parkinglot	100	100.0	99.0	98.9
Tenniscourt	100	100.0	99.0	98.9
Agricultural	100	99.0	99.0	98.4
Baseballdiamond	100	100.0	98.0	98.4

Appendix

TABLE XVI : UC Merced – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Class	Support	Precision	Recall	F1-Score
Sea Ice	700	99.7	98.3	97.0
Airplane	700	98.1	98.3	96.2
Beach	700	96.7	97.7	95.3
Circular Farmland	700	96.6	97.6	95.1
Golf Course	700	98.8	94.7	94.8
Harbor	700	93.7	99.6	94.6

Appendix

TABLE XVII : RESISC45 – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Class	Support	Precision	Recall	F1-Score
Forest	800	100.0	100.0	100.0
Oil Well	800	100.0	100.0	100.0
Freeway	800	100.0	99.7	100.0
Airplane	800	99.8	99.9	100.0
Runway	800	99.9	99.7	100.0
Beach	800	99.5	100.0	100.0

Appendix

TABLE XVIII : PatternNet – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Class	Support	Precision	Recall	F1-Score
Woodland	6258	99.8	99.8	92.0
Construction Land	3791	99.8	99.4	91.9
Water Area	4104	98.8	99.1	91.4
Other Objects	884	99.7	98.1	91.3
Other Land	3593	99.1	98.5	91.2
Cultivated Land	2817	98.1	98.8	90.9

Appendix

TABLE XIX : RSI-CB256 – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Class	Support	Precision	Recall	F1-Score
Football Field	50	100.0	100.0	87.6
Meadow	61	100.0	98.4	86.8
River	56	100.0	98.2	86.8
Pond	54	98.2	100.0	86.8
Port	53	100.0	98.1	86.7
Mountain	50	98.0	100.0	86.7

Appendix

TABLE XX : WHU-RS19 – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Class	Support	Precision	Recall	F1-Score
Dense Forest	69	100.0	98.6	99.3
Harbor	44	97.7	97.7	97.7
Sparse Forest	56	100.0	89.3	94.3
Storage Tanks	53	98.0	90.6	94.1
Grassland	51	90.7	96.1	93.3
Residential Area	70	86.8	94.3	90.4

Appendix

TABLE XXI : RSC11 – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Class	Support	Precision	Recall	F1-Score
Field	200	86.8	85.5	86.0
Resident	200	82.4	82.0	82.2
Forest	200	75.7	85.5	80.3
River Or Lake	200	83.9	75.5	79.5
Parking	200	72.2	67.5	69.8
Grass	200	71.8	67.5	69.6

Appendix

TABLE XXII : RS2800 – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Class	Support	Precision	Recall	F1-Score
Water	100	99.0	100.0	99.4
Residential	100	91.4	96.0	93.7
Harbor	100	92.9	92.0	92.5
Overpass	100	83.8	98.0	90.3
Commercial	100	86.5	90.0	88.2
Industrial	100	83.2	84.0	83.6

Appendix

TABLE XXIII : SIRI-WHU – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Class	Support	Precision	Recall	F1-Score
Airplane	30	100.0	100.0	100.0
Beach	30	96.7	96.7	96.7
Chaparral	30	96.6	93.3	94.9
Island	30	100.0	90.0	94.7
Harbor	30	93.1	90.0	91.5
Parking Lot	30	93.1	90.0	91.5

Appendix

TABLE XXIV : OPTIMAL31 – ResNet-18 student per-class results (%). Sorted by F1-Score descending.

Panel of Figure 30 — Fig. 6 : Dataset overview. The $30\times$ size imbalance (1 005 to 31 500 images) motivates dataset-balanced sampling and stratified active queries.

Fig. 7 : WHU-RS19 Macro-F1 comparison. RemoteCLIP remains strongest on this small dataset, while the balanced student substantially narrows the gap and MoE+GCN further improves the best checkpoint.

Panel of Figure 32 — Fig. 8 : t-SNE projections of RemoteCLIP features on core datasets. Classes are generally well-separated, confirming the quality of the teacher’s zero-shot representations.

Panel of Figure 33 — Fig. 9 : t-SNE projections of student features on core datasets. The student achieves comparable cluster separation to RemoteCLIP, confirming successful knowledge distillation.

Panel of Figure 34 — Fig. 10 : Normalized confusion matrices – extended datasets.

Panel of Figure 35 — Fig. 11 : Normalized confusion matrices – unseen datasets (prototype mode).

Panel of Figure 36 — Fig. 12 : Lowest-accuracy classes per dataset. Hard categories typically share visual patterns with neighboring classes, limiting discriminability under limited labeled supervision.

Panel of Figure 37 — Fig. 13 : Initial labeled pool class distributions. WHU-RS19 (config 04) introduces the largest imbalance: some classes receive as few as 15 initial labeled samples vs. over 200 for PatternNet or RESISC45 classes, motivating the stratified active-query strategy.

Data-Efficient Learning

Data-efficient learning develops methods that achieve accurate, robust models with fewer labeled examples, training samples, or interactions.

See all 236 papers

Aug 26, 2026

Semi-Supervised Adaptation of Vision-Language Models for Image Classification

Vision-language models like CLIP have shown sig- nificant potential in handling natural images, yet their perfor- mance is often limited by the distinct characteristics of satellite imagery. While parameter-efficient adaptation techniques exist, their efficacy is frequently limited by the scarcity of annotated samples. In this letter, we propose Self-Evolutionary CLIP (SE- CLIP), a semi-supervised framework designed for recursive label mining in scene classification. The approach follows a dual-phase pipeline, where an initial warm-up on a few annotated seeds is followed by a recursive discovery phase that iteratively identifies high-confidence samples from unlabeled pools. To maintain the integrity of the evolving support set, we employ a class-balanced selection strategy that prevents the model from being dominated by easily learned categories. Results on the UCM and NWPU benchmarks indicate that SE-CLIP significantly outperforms existing semi-supervised approaches. The framework provides a viable solution for adapting VLMs to the remote sensing domain with minimal human intervention.

Mohamed L. Mekhalfi, Mohamad M. Al Rahhal, Yakoub Bazi +4

Fondazione Bruno Kessler, Via Sommarive 18, 38123, Povo TN, Italy · Applied Computer Science Department, College of Applied Computer Science, King Saud University, 11543, Riyadh, Saudi Arabia · Computer Engineering Department, College of Computer and Information Sciences, King Saud University, 11543, Riyadh, Saudi Arabia +2

Remote Sensing Scene Classification VLM Adaptation Semi-Supervised Learning Remote Sensing Image Understanding

Ask PDF

Mar 12, 2026

OSMDA: OpenStreetMap-based Domain Adaptation for Remote Sensing VLMs

Vision-Language Models (VLMs) adapted to remote sensing rely heavily on domain-specific image-text supervision, yet high-quality annotations for satellite and aerial imagery remain scarce and expensive to produce. Prevailing pseudo-labeling pipelines address this gap by distilling knowledge from large frontier models, but this dependence on large teachers is costly, limits scalability, and caps achievable performance at the ceiling of the teacher. We propose OSMDA: a self-contained domain adaptation framework that eliminates this dependency. Our key insight is that a capable base VLM can serve as its own annotation engine: by pairing aerial images with rendered OpenStreetMap (OSM) tiles, we leverage optical character recognition and chart comprehension capabilities of the model to generate captions enriched by OSM's vast auxiliary metadata. The model is then fine-tuned on the resulting corpus with satellite imagery alone, yielding OSMDA-VLM, a domain-adapted VLM that requires no manual labeling and no stronger external VLM teacher. We conduct exhaustive evaluations spanning six zero-shot and five in-distribution benchmarks across vision-language tasks, where OSMDA leads to substantial improvement. We further compare against nine competitive baselines, demonstrating that our method achieves superior overall performance, while being substantially cheaper to train than teacher-dependent alternatives. These results suggest that, given a strong foundation model, alignment with crowd-sourced geographic data is a practical and scalable path towards remote sensing domain adaptation. Dataset and model weights will be made publicly available upon acceptance.

Stefan Maria Ailuro, Mario Markov, Mohammad Mahdi +3

INSAIT, Sofia University "St. Kliment Ohridski" · INSAIT, Sofia University “St. Kliment Ohridski”

Remote Sensing VLMs VLM Adaptation Vision-Language Models Unsupervised Domain Adaptation

Ask PDF

Apr 10, 2026

Low-Data Supervised Adaptation Outperforms Prompting for Cloud Segmentation Under Domain Shift

Adapting vision-language models to remote sensing imagery presents a fundamental challenge: both the visual and linguistic distributions of satellite data lie far outside natural image pretraining corpora. Despite this, prompting remains the dominant deployment paradigm, driven by the assumption that domain-specific language can guide frozen model representations toward specialized tasks. We test this assumption directly on a domain where the mismatch is prominent: cloud segmentation for satellite imagery. Using CLIPSeg on the CloudSEN12+ cloud segmentation benchmark, we evaluate 60 prompt variants spanning simple labels, domain terminology, appearance descriptors, and contextual cues, finding that every variant underperforms the zero-shot baseline (0.255 mIoU), with engineered prompts scoring as low as 0.07 mIoU. No amount of linguistic refinement bridges the gap between CLIP's natural image representations and satellite spectral imagery. In contrast, supervised fine-tuning with just 0.1% labeled data (~8 images) surpasses zero-shot performance overall, and 5-10% data recovers ~85% of maximum achievable mIoU. Full fine-tuning consistently outperforms low-rank adaptation by 0.03-0.09 mIoU, with the largest gaps for spectrally ambiguous classes, and at 0.5 to 1% labeled data, fine-tuning temporarily degrades performance on these classes before recovering, a supervision dip that aggregate mIoU can mask. For practitioners adapting vision-language models to specialized imagery, our results deliver a clear message: labeled data is not the expensive alternative to prompting; it is the worthwhile path.

Harshith Kethavath, Weiming Hu

University of Georgia, USA

VLM Adaptation Remote Sensing VLMs Fine-Tuning Image Segmentation Remote Sensing Image Segmentation

Ask PDF

Abstract

Data-Efficient Learning

Explore similar work

Semi-Supervised Adaptation of Vision-Language Models for Image Classification

OSMDA: OpenStreetMap-based Domain Adaptation for Remote Sensing VLMs

Low-Data Supervised Adaptation Outperforms Prompting for Cloud Segmentation Under Domain Shift