Research map / Label-efficient and robust learning / Knowledge distillation
Brewing Stronger Features: Dual-Teacher Distillation for Multispectral Earth Observation
Filip Wolf, Blaž Rolih, Luka Čehovin Zajc
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DINOv3 class-token distillation to MS contrastive baseline (ablation) | DEO (Swin-Tiny) · segmentation, 3 optical + 3 MS | avg mIoU: 69.16 -> 70.94 | Improved |
| add separate optical pass and projection head (ablation) | DEO (Swin-Tiny) · segmentation, optical avg | avg mIoU: 79.07 -> 81.20 | Improved |
| add separate optical pass and projection head (ablation) | DEO (Swin-Tiny) · segmentation, MS avg | avg mIoU: 62.81 -> 62.69 | Hurt |
| add DINOv3 patch token distillation (ablation) | DEO (Swin-Tiny) · segmentation, 3 optical + 3 MS | avg mIoU: optical +0.53; MS 62.69 -> 62.46 | Mixed |
| add heavy optical augmentations (ablation) | DEO (Swin-Tiny) · segmentation, overall | avg mIoU: 72.10 -> 72.48 | Improved |
| replace low-res Sentinel-2 optical with high-res aerial imagery in pretraining (ablation) | DEO (Swin-Tiny) · segmentation, MS avg | avg mIoU: 63.02 -> 63.51 | Improved |
| use DINOv3 instead of DINOv2 as distillation teacher (ablation) | DEO · segmentation, overall | avg mIoU: 73.86 -> 73.90 | No clear effect |
| use RADIOv2.5 instead of DINOv3 as teacher (ablation) | DEO · segmentation, optical vs MS | avg mIoU: optical 85.03 -> 83.57; MS 62.77 -> 63.12 | Mixed |
| DEO dual-teacher vs strongest multispectral baseline | MS segmentation · multispectral segmentation avg | macro mIoU: 53.25 -> 57.45 | Improved |
| DEO vs Scale-MAE on optical change detection | optical change detection · LEVIR | binary F1: 92.1 -> 91.3 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 3 on most shared results.
- TerraFM — TerraFM: A Scalable Foundation Model for Unified Multisensor Earth Observation (ICLR 2026) · better on 2 of 2 shared results
- CROMA — CROMA: Remote Sensing Representations with Contrastive Radar-Optical Masked Autoencoders (NeurIPS 2023) · better on 2 of 2 shared results
- Scale-MAE — Scale-MAE: A Scale-Aware Masked Autoencoder for Multiscale Geospatial Representation Learning (ICCV 2023) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Fine-grained Image-to-LiDAR Contrastive Distillation with Visual Foundation Models (NeurIPS 2024 · similarity 0.87)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.86)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.85)
- DUNIA: Pixel-Sized Embeddings via Cross-Modal Alignment for Earth Observation Applications (ICML 2025 · similarity 0.85)
- Bridging Remote Sensors with Multisensor Geospatial Foundation Models (CVPR 2024 · similarity 0.85)
- Parameter Efficient Self-Supervised Geospatial Domain Adaptation (CVPR 2024 · similarity 0.85)
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.
