Research map / Architectures and efficiency / Pruning and efficient tuning
Parameter-Efficient Adaptation of Geospatial Foundation Models through Embedding Deflection
Romain Thoreau, Valerio Marsocci, Dawa Derksen
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 |
|---|---|---|---|
| adapt only the UPerNet layers instead of all layers with uAtt (ablation) | DEFLECT · MADOS, Scale-MAE | mIoU: 38.7 -> 50.6 | Improved |
| adapt UPerNet layers instead of only the first layer (ablation) | DEFLECT · MADOS, Scale-MAE | mIoU: 47.1 -> 50.6 | Improved |
| adapt UPerNet layers instead of only the first layer (ablation) | DEFLECT · MADOS, Cross-Scale MAE | mIoU: 39.2 -> 38.2 | Hurt |
| remove low-rank reparameterization of uAtt parameters (ablation) | DEFLECT · MADOS, DINO-MC | mIoU: 51.6 -> 52.9 | Improved |
| increase LoRA rank from 16 to 32 (competitor) (ablation) | LoRA · MADOS, Scale-MAE | mIoU: 45.0 -> 30.2 | Hurt |
| add linear projection layer to spectral embeddings (ablation) | DEFLECT · MADOS, DINO-MC | mIoU: 48.5 -> 51.6 | Improved |
| sample 50% of pixels instead of 10% in pixel-set encoding (ablation) | DEFLECT · MADOS / BurnScars | mIoU / IoU: 50.6 -> 51.8 (Scale-MAE MADOS) | No clear effect |
| replace LoRA/SLR with DEFLECT on DINO-MC | LoRA / SLR · Average performance (5 datasets, DINO-MC) | Avg. performance: 52.7 -> 64.8 (vs SLR) | Improved |
| use DEFLECT instead of full finetuning | Cross-Scale MAE · ForestNet (mF1) | mF1: 37.1 -> 46.3 | Improved |
| use DEFLECT instead of full finetuning | Scale-MAE · MADOS (mIoU) | mIoU: 47.0 -> 50.6 | Improved |
| use RGB+random initialization instead of Repeat for new bands (ablation) | PEFT methods · BurnScars / MADOS | IoU / mIoU: 70.5 -> 73.2 (SLR, DINO-MC, BurnScars) | Mixed |
Closest papers by meaning
- Parameter Efficient Self-Supervised Geospatial Domain Adaptation (CVPR 2024 · similarity 0.87)
- AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modalities (CVPR 2025 · similarity 0.85)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.84)
- No One Knows the State-of-the-Art in Geospatial Foundation Models (NeurIPS 2026 · similarity 0.84)
- Learning De-Biased Representations for Remote-Sensing Imagery (NeurIPS 2024 · similarity 0.84)
- SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery (CVPR 2024 · similarity 0.84)
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.
