Research map / Label-efficient and robust learning / Recognition and long-tail
Learning De-Biased Representations for Remote-Sensing Imagery
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.
Proposes debLoRA, an unsupervised de-biasing add-on for LoRA variants that clusters features, calibrates tail-class features toward weighted cluster centers, and trains a LoRA module to map biased features to de-biased ones, improving tail-class accuracy for remote-sensing adaptation.
Problem
Data-efficient adaptation (LoRA) of foundation models to long-tailed, data-scarce remote sensing imagery yields feature spaces biased toward head classes, hurting tail-class performance.
Approach
Cluster features regardless of labels to obtain shared visual attributes, form a de-biased center per tail class as a class-count-weighted average of cluster centers, and train a LoRA-like mapping that pulls tail features toward these calibrated targets, with the calibration strength set by the class imbalance ratio.
Key results
- On SatMAE to FUSRS, debLoRA improves tail-class macro F1 by 4.7 points over LoRA (81.0 vs 76.3 on SatMAE→FUSRS tail); on SD→DOTA tail F1 rises from 91.8 to 94.5 with LoRA rank 8.
- On oriented object detection with DOTA, debLoRA raises tail mAP from 66.5 (LoRA) to 73.2 and average mAP from 76.7 to 80.4.
- Mean tail-class F1 across the three adaptation settings is 90.3 for debLoRA on LoRA variant versus 86.6 for LoRA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add debLoRA to LoRA | LoRA · SD → DOTA | macro F1, tail (%): 91.8 -> 95.1 | Helped |
| Add debLoRA to cLoRA | cLoRA · SD → DOTA | macro F1, tail (%): 89.3 -> 93.5 | Helped |
| Add debLoRA to cLoRA | cLoRA · OpenCLIP → DOTA | macro F1, tail (%): 92.2 -> 95.0 | Helped |
| Add debLoRA to LoRA | LoRA · ORS → FUSRS | macro F1, tail (%): 76.3 -> 81.0 | Helped |
| Add debLoRA to LoRA | LoRA · SD → DOTA oriented object detection | mAP, tail (%): 66.5 -> 73.2 | Helped |
| Add debLoRA to LoRA | LoRA · Places365-LT | top-1 accuracy, tail (%): 44.9 -> 49.2 | Helped |
| Add debLoRA to LoRA | LoRA · iNaturalist 2018 | top-1 accuracy, tail (%): 71.2 -> 78.4 | Helped |
| Add debLoRA to LoRA | LoRA · SD → fMoW-S2 | accuracy, tail (%): 38.1 -> 41.2 | Helped |
| Reduce LoRA rank from 64 to 8 (ablation) | LoRA · SD → DOTA | macro F1, tail (%): 2.2 percentage points | Hurt |
| Add debLoRA to rank 8 LoRA (ablation) | LoRA rank 8 · SD → DOTA | macro F1, tail (%): 91.8 -> 94.5 | Helped |
| Add debLoRA to rank 64 LoRA (ablation) | LoRA rank 64 · SD → DOTA | macro F1, tail (%): 94.0 -> 96.2 | Helped |
| Increase debLoRA cluster count from 16 to 32 (ablation) | debLoRA · SD → DOTA | macro F1, tail (%): 90.4 -> 95.1 | Helped |
Surprising findings
- cLoRA performs notably better under OpenCLIP→DOTA than under SD→DOTA, slightly surpassing LoRA, which the authors attribute to alignment between OpenCLIP's feature space and class-specific design.
Limitations
- Assumes visual attributes are shared across classes; may be less effective for highly class-specific attributes or large intra-class variation.
- Performance is sensitive to the number of clusters K, which must be tuned per dataset and setting.
- SAR evaluation relies on a customized dataset built from FUSAR-Ship and SRSDD due to scarce SAR benchmarks.
Experimental setup
Datasets: DOTA FUSAR-Ship SRSDD Places365-LT iNaturalist 2018 fMoW-S2
Benchmarks: DOTA FUSRS Places365-LT iNaturalist 2018 fMoW-S2
Metrics: macro F1 mAP top-1 accuracy
Baselines: Fine-Tune LoRA cLoRA ResLT SADE Reweight ECM Zero-Shot
Builds on: LoRA cLoRA FCOS
Closest papers by meaning
- Galileo: Learning Global & Local Features of Many Remote Sensing Modalities (ICML 2025 · similarity 0.85)
- Rethinking Transformers Pre-training for Multi-Spectral Satellite Imagery (CVPR 2024 · similarity 0.85)
- Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning (CVPR 2025 · similarity 0.84)
- RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark (CVPR 2025 · similarity 0.84)
- S2MAE: A Spatial-Spectral Pretraining Foundation Model for Spectral Remote Sensing Data (CVPR 2024 · similarity 0.84)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.83)
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.