ganghyunnnn/rs-taxonomy-labels

RS-Taxonomy: GSD-Sensitive Task Labels for Remote Sensing VQA

This dataset accompanies the paper "Identifying the Measurement Gap in Remote Sensing VQA with a GSD-Sensitive Taxonomy" (IEEE GRSL, under review). It provides per-task D / M1 / M2 taxonomy labels, inter-annotator agreement (IAA) data, and reproducibility artifacts for four public RS-VQA benchmarks.

The taxonomy partitions tasks via a single counterfactual: if the GSD were doubled, would the answer value change (M1) or would the question become physically unanswerable while the value stays the same (M2)?

Type	Name	Definition
D	Descriptive	GSD-invariant; visual–semantic interpretation only. E.g., "What is the land use type?", "Is there an airport?"
M1	Spatial Metric	Output value scales with GSD (real-world distance / area). E.g., "Distance between the two hangars? (GSD = 0.3 m/px)"
M2	Cardinality	Counting tasks; values are GSD-invariant but answerability is resolution-conditioned (counting feasible only when GSD ≤ d/s for target size d, resolvability threshold s ≈ 10–15 px). E.g., "How many vehicles are in the parking lot?"

Boundary rules. (1) Proximity queries with a numeric distance threshold are M1, otherwise D. (2) Bounding-box drawing is always D (output is pixel coordinates). (3) Comparisons inherit the underlying operation: counting-based comparisons → M2; GSD-based spatial comparisons → M1.

Headline results from the paper. Across 293,607 questions in four benchmarks, M-type prevalence ranges from 2.9% to 70.4%. Measurement tasks fail 19–31 pp more often than descriptive tasks across two agent baselines and three VLM backbones, robust to BH multiple-comparison correction. A GSD prompt-injection ablation on 161 M1 tasks shows no significant improvement after correction. Inter-annotator agreement: Cohen's κ = 0.95. Rule-based classifier accuracy: 95.2%.

Important: This dataset contains only the taxonomy annotations and evaluation artifacts. The underlying RS-VQA benchmark images and questions are not redistributed here — download them from the original sources and join on task_id.

Dataset Summary

File	Type	Description
thinkgeo_taxonomy_labels.json	labels	ThinkGeoBench task → list of D/M1/M2 labels (multi-label)
thinkgeo_taxonomy_summary.json	aggregate	Distribution summary over ThinkGeo
iaa_sample.csv	IAA seed	50 stratified ThinkGeo tasks (annotator 1)
iaa_sample_annotator2.csv	IAA	Same 50 tasks labeled by annotator 2
iaa_annotator2.json	IAA	Annotator 2 labels in JSON form
iaa_sample_annotator2_rationale_ko.md	IAA notes	Per-task rationale (Korean)
iaa_guideline.md	docs	Annotation guideline
router_eval*.json	results	Rule/LLM/hybrid router evaluation
llm_router_preds*.json	results	Per-task LLM router predictions
backbone_*.json	results	Backbone VLM evaluation traces
rsvqa_*.json	results	RSVQA-LR evaluation traces
gsd_ablation.json	results	GSD-prompt ablation
bootstrap_sensitivity.json	results	Bootstrap CI sensitivity
failure_analysis_by_type.json	results	Failure rates by D/M type
task_level_*.json	results	Task-level prompting comparisons

Source Benchmarks

Benchmark	License	Where to download
ThinkGeo	Apache-2.0	https://github.com/mbzuai-oryx/ThinkGeo
RSVQA-LR	CC BY 4.0	https://zenodo.org/records/6344334
FloodNet	MIT	https://github.com/BinaLab/FloodNet-Supervised_v1.0
EarthVQA	Academic only*	https://github.com/Junjue-Wang/EarthVQA

* EarthVQA images are restricted to academic use (RSIDEA, Wuhan University). This dataset does not redistribute EarthVQA content; cross-benchmark distribution figures were computed locally from the academic release.

Citations for Source Benchmarks

@misc{thinkgeo,
  author    = {Shabbir, Akashah and Munir, Muhammad Akhtar and Dudhane, Akshay and
               Sheikh, Muhammad Umer and Khan, Muhammad Haris and Fraccaro, Paolo and
               Moreno, Juan Bernabe and Khan, Fahad Shahbaz and Khan, Salman},
  title     = {{ThinkGeo}: Evaluating Tool-Augmented Agents for Remote Sensing Tasks},
  year      = {2025},
  eprint    = {2505.23752},
  archivePrefix = {arXiv}
}

@article{rsvqa,
  author  = {Lobry, Sylvain and Marcos, Diego and Murray, Jesse and Tuia, Devis},
  title   = {{RSVQA}: Visual Question Answering for Remote Sensing Data},
  journal = {IEEE Trans. Geosci. Remote Sens.},
  year    = {2020}, volume = {58}, number = {12}, pages = {8555--8566},
  doi     = {10.1109/TGRS.2020.2988782}
}

@article{floodnet,
  author  = {Rahnemoonfar, Maryam and Chowdhury, Tashnim and Sarkar, Argho and
             Varshney, Debvrat and Yari, Masoud and Murphy, Robin R.},
  title   = {{FloodNet}: A High Resolution Aerial Imagery Dataset for Post Flood Scene Understanding},
  journal = {IEEE Access}, year = {2021}, volume = {9}, pages = {89644--89654},
  doi     = {10.1109/ACCESS.2021.3090981}
}

@inproceedings{earthvqa,
  author    = {Wang, Junjue and Zheng, Zhuo and Chen, Zihang and Ma, Ailong and Zhong, Yanfei},
  title     = {{EarthVQA}: Towards Queryable Earth via Relational Reasoning-Based Remote Sensing Visual Question Answering},
  booktitle = {Proc. AAAI Conf. Artificial Intelligence},
  year      = {2024}, volume = {38}, number = {6}, pages = {5481--5489},
  doi       = {10.1609/aaai.v38i6.28357}
}

File Provenance

File pattern	Provenance	Upstream license
thinkgeo_taxonomy_labels.json, thinkgeo_taxonomy_summary.json	Our labels keyed by upstream task_id (no question text)	n/a (our work)
iaa_sample.csv, iaa_sample_annotator2.csv	Our labels + verbatim ThinkGeoBench question text and image filenames for the 50-task IAA sample	ThinkGeoBench, Apache-2.0
iaa_annotator2.json	Our labels keyed by task_id only	n/a (our work)
iaa_sample_annotator2_rationale_ko.md	Our rationale + Korean translations/quotations of selected questions	quoted text: ThinkGeoBench, Apache-2.0
iaa_guideline.md	Our annotation guideline	n/a (our work)
review_436.csv	Our labels + verbatim ThinkGeoBench question text	ThinkGeoBench, Apache-2.0
router_eval*.json, llm_router_preds*.json, router_llm_qwen35_9b.json	Router accuracy and predictions	n/a (our work)
backbone_*.json, gsd_ablation.json, task_level_*.json	Model predictions on ThinkGeoBench tasks (model outputs are ours; questions referenced by task_id)	n/a (our work)
rsvqa_*.json	Model predictions on RSVQA-LR (only upstream q_id integers retained)	n/a (our work)
bootstrap_sensitivity.json, failure_analysis_by_type.json	Aggregate statistics	n/a (our work)

Schema

thinkgeo_taxonomy_labels.json

{ "<task_id>": ["D"], "<task_id>": ["M1", "M2"], ... }

Multi-label list per task. task_id is the integer index into ThinkGeoBench.

iaa_sample.csv / iaa_sample_annotator2.csv

Columns: task_id, image, query, type_annotator, notes where type_annotator ∈ {D, M1, M2, D+M1, D+M2, M1+M2, D+M1+M2, ...}.

router_eval*.json

Aggregated router accuracy / per-class precision-recall.

backbone_*.json, rsvqa_*.json

Per-task records: {task_id, type, prompt, prediction, reference, correct, ...}.

Loading

from datasets import load_dataset
import json, urllib.request

# Single file
url = "https://huggingface.co/datasets/ganghyunnnn/rs-taxonomy-labels/resolve/main/thinkgeo_taxonomy_labels.json"
labels = json.loads(urllib.request.urlopen(url).read())

# IAA CSV via datasets
ds = load_dataset(
    "ganghyunnnn/rs-taxonomy-labels",
    name="iaa_sample",
    split="train",
)

Reproducing Paper Results

Code lives in the companion GitHub repository: https://github.com/ganghyunnnn/rs-taxonomy

After downloading the source benchmarks, run:

python src/eval/run_all_experiments.py

Annotation Process

• Annotator 1 (lead, paper author): labeled all ThinkGeoBench tasks.
• Annotator 2: independently labeled a 50-task stratified sample for IAA.
• Multi-label scheme: a task may carry multiple D/M tags when the answer requires more than one capability (e.g., D+M2 = identify + count).
• IAA computed as Cohen's κ and macro-F1 per label.

Limitations

• Multi-label annotation introduces label-set ambiguity; rationale notes document marginal cases.
• The eval split's M1 under-representation is mitigated by bootstrap analysis in the paper.
• The RSVQA-LR replication is confined to D/M2 (no M1 questions exist in that benchmark at 10 m/px Sentinel-2 resolution).
• The GSD injection ablation (N=161) is powered only for effects ≥12 pp; smaller improvements would require a larger sample.
• ThinkGeo task_id indexing must match the upstream JSON release used at the time of annotation; see data/README.md in the GitHub repo.

License

Portions authored by this project — per-task D / M1 / M2 labels, IAA rationale, evaluation outputs, this dataset card, and the annotation guideline — are released under Creative Commons Attribution 4.0 International (CC BY 4.0).

A subset of the published files additionally redistributes verbatim question text and/or image filenames from upstream RS-VQA benchmarks (see File Provenance above). For those embedded portions the upstream license takes precedence:

• ThinkGeoBench question text and image filenames in iaa_sample.csv, iaa_sample_annotator2.csv, review_436.csv, and iaa_sample_annotator2_rationale_ko.md remain under Apache License 2.0 (Shabbir et al., MBZUAI Oryx Lab). Downstream redistribution must preserve the upstream attribution; see the NOTICE file in the companion GitHub repository for the full text.

The underlying benchmark images are not redistributed and remain under their original licenses.

Citation

@article{park2026rstaxonomy,
  title   = {Identifying the Measurement Gap in Remote Sensing VQA with a GSD-Sensitive Taxonomy},
  author  = {Park, Ganghyun and Lee, Dongho},
  journal = {IEEE Geoscience and Remote Sensing Letters},
  year    = {2026},
  note    = {Under review}
}