Research map / Vision-language and multimodal / Multimodal LLMs
CHOICE: Benchmarking the Remote Sensing Capabilities of Large Vision-Language Models
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.
Introduces CHOICE, a hierarchical, leakage-free multiple-choice benchmark (10,507 problems, 23 tasks, 50 cities) that evaluates 24 general-domain, remote-sensing, and proprietary VLMs on remote sensing perception and reasoning.
Problem
Existing remote sensing VLM evaluations are fragmented, use coarse taxonomies, rely on exact-match metrics, and reuse public datasets (e.g., DOTA, DIOR) that may have leaked into VLM training, so they cannot objectively measure hierarchical remote sensing capabilities.
Approach
Builds a three-level capability taxonomy (2 L-1, 6 L-2, 23 L-3 tasks) using all-new imagery collected worldwide, and converts nearly all tasks into 2-4 option multiple-choice questions with definitive answers, plus rotated-box visual grounding and mask-based referring segmentation, to allow objective scoring.
Key results
- Qwen2-VL-72B achieves the highest accuracy in most perception tasks and outperforms GPT-4o by 7% in change detection and 15.3% in physical property, while open-source InternVL2-40B matches or beats proprietary models in several tasks.
- Most general-domain VLMs score below 70% on single-instance identification and cross-instance discernment, and most RSVLMs fall under 50% on these dimensions.
- Accuracy on attribute reasoning and assessment reasoning is near 50% for most models, with some RSVLMs below 30%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a remote-sensing-specific VLM (VHM) vs general-domain models | RSVLMs vs general VLMs · CHOICE L-2 | accuracy across six L-2 dimensions | Mixed |
| Scaling general VLM size (Qwen2-VL-7B to 72B) | Qwen2-VL · CHOICE L-2 | SII accuracy: 0.638 to 0.702 | Helped |
| Open-source Qwen2-VL-72B vs proprietary GPT-4o on Change Detection | Qwen2-VL-72B · CHOICE CD task | accuracy: outperforms GPT-4o by 7% | Helped |
| Image-level comprehension and common-sense reasoning are strong for most models | general and RS VLMs · CHOICE L-2 | ILC and CSR accuracy | No clear effect |
Surprising findings
- RSVLMs show no consistent superiority over general-domain VLMs and often underperform them, even on remote sensing tasks.
- Proprietary models such as GPT-4o often choose the "can't judge" option (70.2% of RA questions), inflating refusal-based behaviour.
Limitations
- Pixel-level referring expression segmentation is largely unsolved by evaluated VLMs, and the benchmark covers only optical RGB-type imagery in the main tasks.
- Coverage of tasks is incomplete (e.g., urban heat island and vegetation health are absent).
- Evaluation uses zero-shot default configurations and a single prompt strategy for each model.
Experimental setup
Datasets: Google Earth Engine imagery OpenStreetMap SWISSIMAGE Sentinel-1/2 Landsat-8 SDGSAT-1 WorldPop CNBH-10m
Benchmarks: CHOICE
Metrics: accuracy IoU>0.5 accuracy
Baselines: GPT-4o-2024-11-20 GPT-4o-mini Gemini-1.5-Pro GeoChat LHRS-Bot VHM RemoteCLIP GeoRSCLIP
Builds on: CLIP GPT-4 OpenStreetMap
Closest papers by meaning
- VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing (NeurIPS 2026 · similarity 0.92)
- Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.91)
- SkyScript: A Large and Semantically Diverse Vision-Language Dataset for Remote Sensing (AAAI 2024 · similarity 0.89)
- VRSBench: A Versatile Vision-Language Benchmark Dataset for Remote Sensing Image Understanding (NeurIPS 2024 · similarity 0.89)
- Quality-Driven Curation of Remote Sensing Vision-Language Data via Learned Scoring Models (NeurIPS 2025 · similarity 0.89)
- Landsat30-AU: A Vision-Language Dataset for Australian Landsat Imagery (AAAI 2026 · similarity 0.89)
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.