Research map / Image generation / Super-resolution
Scale-aware Recognition in Satellite Images under Resource Constraints
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 a budget-constrained scheme for satellite image retrieval that distills HR-trained models into LR models, samples HR tiles by LR-vs-KD model disagreement, and uses an LLM with in-context examples to decide which concepts need HR.
Problem
Satellite image concept recognition depends on spatial scale, but high-resolution imagery is costly and limited; the paper asks which resolution suits a concept and where to acquire HR images under a budget.
Approach
Combines three components: MSE-based distillation of HR model scores into an LR student (privileged information), ranking locations for HR acquisition by disagreement between the LR student and the LR model (no HR imagery needed), and in-context LLM prediction of whether a concept is better served by LR or HR imagery.
Key results
- Full system with GRAFT reaches 0.633 seen mAP@100 vs 0.501 for GRAFT HR, using 5,954 vs 25,163 HR sq. km. (about 23.6% of HR imagery).
- On unseen classes, GRAFT full reaches 0.502 mAP@100 vs 0.541 for GRAFT HR, while the GRAFT LR baseline reaches 0.379 mAP@100.
- LLM-based concept scale inference achieved 100% accuracy on seen and unseen concepts, versus 90% for OSM-area and always-HR baselines (on the 10 unseen classes).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full scale-aware system with LLM resolution selection, disagreement-based HR sampling, and KD fallback instead of HR-only inference | GRAFT · Sentinel-2/NAIP, seen classes | mAP^100: 0.501 -> 0.633 | Helped |
| Use the full scale-aware system instead of HR-only inference | GRAFT · Sentinel-2/NAIP, unseen classes | mAP^100: 0.541 -> 0.502 | Hurt |
| Use the full scale-aware system instead of HR-only inference | Supervised ResNet-50 · Sentinel-2/NAIP | mAP^100: 0.695 -> 0.736 | Helped |
| Distill HR-model scores into the LR model (ablation) | GRAFT on LR imagery · Sentinel-2/NAIP, seen classes | mAP^100: 0.482 -> 0.534 | Helped |
| Distill HR-model scores into the LR model (ablation) | Supervised ResNet-50 on LR imagery · Sentinel-2/NAIP | mAP^100: 0.451 -> 0.570 | Helped |
| Add LLM-based concept resolution selection to KD plus disagreement sampling (ablation) | GRAFT · Sentinel-2/NAIP, seen classes | mAP^100: 0.621 -> 0.633 | Helped |
| Add LLM-based concept resolution selection to KD plus disagreement sampling (ablation) | GRAFT · Sentinel-2/NAIP, unseen classes | mAP^100: 0.461 -> 0.502 | Helped |
| Use SR imagery rather than LR KD scores to calculate disagreement (ablation) | GRAFT full system · Sentinel-2/NAIP, unseen classes | mAP^100: 0.502 -> 0.556 | Helped |
| Use SR imagery rather than LR KD scores to calculate disagreement (ablation) | GRAFT full system · Sentinel-2/NAIP | inference time: 372 -> 12,884 | Hurt |
| Increase the HR acquisition budget from 100 to 1000 locations (ablation) | GRAFT (Ours full) · Sentinel-2/NAIP, unseen classes | mAP^100: 0.439 -> 0.502 | Helped |
| Select concepts' resolution with an LLM rather than OSM average area | Concept scale inference · Unseen concepts | accuracy: 90% -> 100% | Helped |
Surprising findings
- On unseen classes, the full system's mAP@100 (0.502) is below the GRAFT HR baseline (0.541), so the gains are not uniform across concept sets.
- Super-resolution imagery with the HR model for disagreement scoring gave some unseen-class gains but took about 34x longer inference than the KD-based approach.
Limitations
- Resolution is treated as discrete (LR vs HR), not continuous, and concept segmentation within images is not supported.
- Training and evaluation labels come from crowd-sourced OpenStreetMap data, which is noisy.
- The LLM concept-scale evaluation uses only 10 unseen concepts, a small set.
Experimental setup
Datasets: Sentinel-2 NAIP NICFI basemaps OpenStreetMap annotations
Benchmarks: Sentinel-2/NAIP benchmark Sentinel-2/NICFI benchmark
Metrics: mAP@100 mAP@20 HR images used (sq. km.) inference time (s) precision@K concept-scale accuracy
Baselines: GRAFT HR GRAFT LR RemoteCLIP CLIP-RSICD OpenCLIP Patchdrop SatMAE Cross-Scale MAE ScaleMAE IS-count night-lights sampling
Builds on: knowledge distillation learning with privileged information in-context learning Patchdrop IS-count
Closest papers by meaning
- Rethinking Transformers Pre-training for Multi-Spectral Satellite Imagery (CVPR 2024 · similarity 0.84)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.83)
- ZoomEarth: Active Perception for Ultra-High-Resolution Geospatial Vision-Language Tasks (CVPR 2026 · similarity 0.83)
- XLRS-Bench: Could Your Multimodal LLMs Understand Extremely Large Ultra-High-Resolution Remote Sensing Imagery? (CVPR 2025 · similarity 0.83)
- GeoLLaVA-8K: Scaling Remote-Sensing Multimodal Large Language Models to 8K Resolution (NeurIPS 2025 · similarity 0.82)
- GLH-Water: A Large-Scale Dataset for Global Surface Water Detection in Large-Size Very-High-Resolution Satellite Imagery (AAAI 2024 · similarity 0.82)
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.