ganghyunnnn/GSD-Sensitivity-Taxonomy-Labels

GSD-Sensitivity Taxonomy: Task Labels for Remote Sensing VQA

Per-task D / M1 / M2 taxonomy labels, inter-annotator agreement (IAA) data, and evaluation traces for four public RS-VQA benchmarks.

Companion to *G. Park and D.-H. Lee, "Identifying the Measurement Gap in Remote Sensing VQA with a GSD-Sensitive Taxonomy," IEEE Geosci. Remote Sens. Lett., 2026* — accepted, DOI to follow. Code: github.com/ganghyunnnn/GSD-Sensitivity-Taxonomy

⚠️ This dataset contains annotations and evaluation artifacts only. The underlying benchmark images and questions are not redistributed — download them from the original sources and join on task_id.

Taxonomy

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

Boundary rules. Proximity queries are M1 with a numeric distance threshold, D without. Bounding-box drawing is D (pixel-coordinate output). Comparisons inherit their operation: counting-based → M2, GSD-based spatial → M1.

Headline results. Across 293,607 questions, M-type prevalence ranges 2.9–70.4%. Measurement tasks fail 19–31 pp more often than descriptive tasks across two agent baselines and three VLM backbones, robust to Benjamini–Hochberg correction. IAA: Cohen's κ = 0.95. Rule-based classifier: 95.2% agreement.

Files

File	Contents
thinkgeo_taxonomy_labels.json	Per-task D/M1/M2 labels (multi-label) + verbatim question text and image filename
review_436.csv	Full 436-task human review with verbatim question text
thinkgeo_taxonomy_summary.json	Distribution summary over ThinkGeo
iaa_sample.csv, iaa_sample_annotator2.csv, iaa_annotator2.json, iaa_guideline.md, iaa_sample_annotator2_rationale_ko.md	88-task stratified IAA sample: both annotators, guideline, per-task rationale (Korean)
router_eval_3type.json	Rule-based classifier metrics on the 189-task split
backbone_*.json, rsvqa_*.json, floodnet_*.json	VLM evaluation traces (ThinkGeo / RSVQA-LR / FloodNet Track-2)
task_level_*.json	ThinkGeo agent baselines (Vanilla ReAct, Direct Prompting)
gsd_ablation.json, m2ab_*.json, m2search_*.json, routed_eval_*.json	Prompt-level interventions: GSD injection, M2 direct-vs-counting decomposition, counting-prompt search, taxonomy-routed prompting
failure_analysis_by_type.json, bootstrap_sensitivity.json	Failure rates by D/M type; bootstrap CI sensitivity

thinkgeo_taxonomy_labels.json is keyed by task_id (integer index into ThinkGeoBench); annotation.types holds the multi-label D/M1/M2 list. IAA CSVs use task_id, image, query, type_annotator, notes. Evaluation traces are per-task {task_id, type, prompt, prediction, reference, correct, ...}.

Loading

import json, urllib.request
from datasets import load_dataset

url = "https://huggingface.co/datasets/ganghyunnnn/GSD-Sensitivity-Taxonomy-Labels/resolve/main/thinkgeo_taxonomy_labels.json"
labels = json.loads(urllib.request.urlopen(url).read())

ds = load_dataset("ganghyunnnn/GSD-Sensitivity-Taxonomy-Labels", name="iaa_sample", split="train")

To reproduce the paper, download the source benchmarks and run python src/eval/run_all_experiments.py from the GitHub repository.

Source Benchmarks

ThinkGeo (Apache-2.0) · RSVQA-LR (CC BY 4.0) · FloodNet (MIT) · EarthVQA (academic-only, RSIDEA / Wuhan University). EarthVQA and FloodNet appear in cross-benchmark distribution counts only; no content from either is redistributed here.

Annotation Process

Annotator 1 (lead author) labeled all ThinkGeoBench tasks; Annotator 2 independently labeled an 88-task stratified sample from question text and a written guideline alone. A task may carry multiple tags when the answer needs more than one capability (e.g. D+M2 = identify + count). IAA is reported as Cohen's κ and macro-F1 per label.

Limitations

• Multi-label annotation introduces label-set ambiguity; rationale notes document marginal cases.
• M1 is under-represented in the evaluation split; the paper mitigates this with bootstrap analysis.
• The RSVQA-LR replication covers D/M2 only — no M1 questions exist at 10 m/px Sentinel-2 resolution.
• The GSD-injection ablation (N=161) is powered only for effects ≥12 pp.
• ThinkGeo task_id indexing must match the upstream JSON release used at annotation time.

License

Portions authored by this project — the D/M1/M2 labels, IAA rationale, evaluation outputs, guideline and this card — are released under CC BY 4.0.

thinkgeo_taxonomy_labels.json, review_436.csv, iaa_sample.csv, iaa_sample_annotator2.csv and iaa_sample_annotator2_rationale_ko.md additionally embed verbatim ThinkGeoBench question text and/or image filenames, which remain under Apache-2.0 (Shabbir et al., MBZUAI Oryx Lab); redistribution must preserve that attribution. rsvqa_*.json retains only upstream q_id integers. Full details: NOTICE.

Benchmark images are not redistributed and remain under their original licenses.

Citation

@article{park2026gsdtaxonomy,
  title   = {Identifying the Measurement Gap in Remote Sensing {VQA} with a {GSD}-Sensitive Taxonomy},
  author  = {Park, Ganghyun and Lee, Dong-Ho},
  journal = {IEEE Geoscience and Remote Sensing Letters},
  year    = {2026},
  note    = {Accepted for publication}
}