Research map / Label-efficient and robust learning / Self-supervised and contrastive
Parameter Efficient Self-Supervised Geospatial Domain Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adds Scaled Low-Rank (SLR) adapters to frozen ViT-L MAE-style remote sensing foundation models and self-supervisedly adapts them to new sensor modalities, then fine-tunes only 1-2% of parameters for downstream classification and segmentation.
Problem
Visual foundation models for remote sensing are pre-trained mainly on optical RGB/multispectral data, leading to distribution gaps for heterogeneous sensors (e.g., SAR); full fine-tuning is costly and fails on small labeled datasets.
Approach
Insert scaled low-rank adapters (input/output scaling vectors plus low-rank matrices) into the linear layers of frozen ViT MAE models, train them with masked autoencoding on unlabeled target-domain data, then use them for linear, scaling, or fine-tuning downstream tasks.
Key results
- SLR adapters raise average linear-evaluation accuracy of MAE from 63.46% to 69.83% (+6.37 points) across 8 remote sensing datasets.
- SLR fine-tuning outperforms full fine-tuning on average by +0.35% (MAE), +1.12% (SatMAE), and +1.49% (Scale-MAE) while training about 1-2% of parameters.
- With k=10 shots on EuroSAT-SAR, SLR Scale (0.5M trainable params) outperforms full fine-tuning by about 10 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| self-supervised SLR adapter training before linear evaluation | MAE · 8 remote sensing datasets (average) | accuracy: 63.46% to 69.83% (+6.37% absolute improvement) | Helped |
| self-supervised SLR adapter training before linear evaluation | SatMAE · 8 remote sensing datasets (average) | accuracy: +2.93% | Helped |
| self-supervised SLR adapter training before linear evaluation | Scale-MAE · 8 remote sensing datasets (average) | accuracy: +3.68% | Helped |
| use SLR adapter fine-tuning instead of full-model fine-tuning | MAE · 8 remote sensing datasets (average) | accuracy: +0.35% | Helped |
| use SLR adapter fine-tuning instead of full-model fine-tuning | SatMAE · 8 remote sensing datasets (average) | accuracy: +1.12% | Helped |
| use SLR adapter fine-tuning instead of full-model fine-tuning | Scale-MAE · 8 remote sensing datasets (average) | accuracy: +1.49% | Helped |
| use SLR Scale instead of full fine-tuning | SatMAE · EuroSAT, k=10 | accuracy: +5% | Helped |
| use SLR Scale instead of full fine-tuning | MAE · EuroSAT-SAR, k=10 | accuracy: +10% | Helped |
| use SLR adapter fine-tuning instead of full-model fine-tuning | SatMAE · BENGE-8k Sentinel-1 SAR segmentation | accuracy: +2.36% | Helped |
| replace SLR adapters with LoRA (ablation) | ImageNet MAE adapted to SAR · EuroSAT-SAR | accuracy: 85.86 ± 0.3 -> 87.14 ± 0.1 | Hurt |
| remove self-supervised adapter pre-training and train adapters from random initialization (ablation) | SLR adapters · BENGE-8k, k=10 | accuracy: +6% | Hurt |
| remove self-supervised adapter pre-training and train adapters from random initialization (ablation) | SLR adapters · EuroSAT, k=10 | accuracy: +7% | Hurt |
Surprising findings
- Full continual pre-training of all parameters on RESISC45 and EuroSAT-SAR gives only marginal gains over SLR adapters (SLR reaches >98% of its accuracy).
- On few-shot EuroSAT with SatMAE, linear evaluation of the frozen model is competitive with SLR linear at k=10, suggesting limited benefit there.
Limitations
- Performance may degrade when source and target distributions differ very widely (e.g., vision model to audio data).
- Method is limited by public availability of pre-trained foundation models.
- Evaluated only on ViT-L MAE-style models and classification/segmentation tasks on remote sensing datasets.
Experimental setup
Datasets: EuroSAT RESISC45 FireRisk TreeSatAI EuroSAT-SAR BENGE-8k UCMerced
Benchmarks: EuroSAT RESISC45 FireRisk TreeSatAI EuroSAT-SAR BENGE-8k UCMerced
Metrics: accuracy average accuracy
Baselines: Linear evaluation full fine-tuning LoRA (IA)3 BitFit Norm tuning random-init SLR adapters
Builds on: masked autoencoder LoRA (IA)3 BitFit ScaleMAE SatMAE
Closest papers by meaning
- RoMA: Scaling up Mamba-based Foundation Models for Remote Sensing (NeurIPS 2025 · similarity 0.88)
- Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning (CVPR 2025 · similarity 0.87)
- SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery (CVPR 2024 · similarity 0.87)
- ExPLoRA: Parameter-Efficient Extended Pre-Training to Adapt Vision Transformers under Domain Shifts (ICML 2025 · similarity 0.86)
- Galileo: Learning Global & Local Features of Many Remote Sensing Modalities (ICML 2025 · similarity 0.86)
- Bridging Remote Sensors with Multisensor Geospatial Foundation Models (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.