🌍 GSTA: Geo-SpatioTemporal Atlas

A curated benchmark atlas for geospatial-temporal LLM & MLLM research

   

GSTA is a lightweight, metadata-only registry of benchmarks for geospatial, spatial, temporal, urban, mobility, map, navigation, remote-sensing, and Earth-observation research with language models.

Quick Start · At a Glance · Schema · Contribute

---

GSTA does not redistribute benchmark samples. Each row is a curated metadata entry pointing to the original paper, code, and data sources when stable public URLs were verified. Please always check the original benchmark license, access policy, and citation requirements before downloading or redistributing linked resources.

✨ Why GSTA?

Geospatial-temporal evaluation is scattered across urban computing, remote sensing, map reasoning, navigation, GIS, mobility, Earth observation, and multimodal scene understanding. GSTA provides a compact atlas for researchers who need to answer questions like:

• Which benchmarks evaluate geospatial reasoning in LLMs or MLLMs?
• Which datasets are text-only, multimodal, map-based, remote-sensing, or tool-oriented?
• What papers, code repositories, public data URLs, metrics, and caveats should I check before choosing an evaluation suite?
• How does benchmark coverage differ across domains, modalities, and years?

📦 Dataset at a Glance

benchmark entries	16 metadata columns	Paper-linked publication metadata	LLM + MLLM model interfaces
Maps navigation & routing	Urban mobility & city tasks	Remote Sensing Earth observation	Spatiotemporal reasoning & planning

Current interface coverage

Model interface	Typical inputs
MLLM	images, map snapshots, remote-sensing imagery, video, visual scenes
LLM	text, tables, coordinates, trajectories, structured spatiotemporal records
LLM; MLLM	mixed text/tool/visual settings

Included benchmark inventory

GeoBench · CityBench · UrBench · MapEval · GEOBench-VLM · XLRS-Bench · STARK · TurnBack · TripCraft · GeoBenchX · GeoAnalystBench · USTBench · FRIEDA · GTR-Bench · GeoMMBench · VIR-Bench · UrbanFeel · TimeSpot · ERGeoBench · EarthSpatialBench · CityLens · GeoAgentBench · GPSBench · UrbanGeoEval · Compositional GeoQA · TP-RAG

🚀 Quick Start

from datasets import load_dataset

# The dataset is a metadata table; each row describes one benchmark.
ds = load_dataset("zhangdw/GSTA", split="train")

print(ds.column_names)
print(ds[0])

Prefer pandas for fast inspection?

import pandas as pd

url = "https://huggingface.co/datasets/zhangdw/GSTA/resolve/main/gsta_benchmarks.csv"
df = pd.read_csv(url)

# Example: find multimodal map / remote-sensing benchmarks
view = df[df["model_type"].str.contains("MLLM", na=False)]
print(view[["benchmark", "domain", "task_type", "metrics"]].head())

🗂️ Dataset File

File	Description
gsta_benchmarks.csv	One benchmark per row. This is the canonical metadata table for the current public snapshot.

🧭 Example Discovery Workflows

Find benchmarks by model interface

llm_only = df[df["model_type"].eq("LLM")]
multimodal = df[df["model_type"].str.contains("MLLM", na=False)]

Find navigation, map, or mobility benchmarks

mask = df["domain"].str.contains("map|navigation|mobility|urban", case=False, na=False)
print(df.loc[mask, ["benchmark", "domain", "task_type", "paper_url"]])

Audit rows that need manual URL follow-up

needs_code = df["code_url"].isna() | df["code_url"].eq("")
needs_data = df["data_url"].isna() | df["data_url"].eq("")
print(df.loc[needs_code | needs_data, ["benchmark", "code_url", "data_url", "url_status", "notes"]])

🧱 Schema

Column	Description
benchmark	Benchmark name or acronym; should be unique within GSTA.
abstract	One-sentence summary of what the benchmark evaluates.
paper_title	Title of the paper that introduced or released the benchmark.
publication_year	Publication year of the benchmark paper.
venue	Venue and track if needed.
research_group	Main organization, lab, or research group behind the benchmark paper, based on public author affiliations and official project/repository ownership.
paper_url	Canonical paper URL.
code_url	Official code repository URL, if verified.
data_url	Official public data or access URL(s), if verified; semicolon-separated for multiple URLs.
url_status	URL verification, access, and release-status note, especially for rows with blank code_url cells or non-direct data access.
domain	Domain or scenario covered by the benchmark.
task_type	Semicolon-separated short task labels.
model_type	LLM, MLLM, or LLM; MLLM.
sample_count	Core benchmark scale.
metrics	Main evaluation metrics.
notes	Caveats, release status, count notes, or registry-specific comments.

🧩 Model Type Convention

Label	Meaning
LLM	Text-based language models, including standard LLMs and reasoning-oriented LLMs/LRMs. Coordinates, trajectories, tables, sensor values, and time series still count as LLM if serialized as text or structured text.
MLLM	Multimodal large language models with non-text inputs such as images, remote-sensing imagery, map images, videos, visual scenes, or raster data.
LLM; MLLM	Benchmarks that evaluate both text-based and multimodal model interfaces.

✅ Intended Use

GSTA is designed for:

• literature review and benchmark discovery;
• selecting evaluation suites for geospatial-temporal LLM/MLLM research;
• tracking benchmark coverage across domains, modalities, metrics, and venues;
• maintaining lab-internal or community-facing benchmark inventories;
• identifying gaps in current geospatial-temporal model evaluation.

🤝 Contributing a Benchmark

Community contributions are welcome. A high-quality new entry should include:

• a stable benchmark name;
• paper title, year, venue, and canonical paper URL;
• verified official code and data links when available;
• concise domain and task labels;
• model interface label: LLM, MLLM, or LLM; MLLM;
• sample count and primary metrics;
• notes for caveats, missing public releases, version differences, or count discrepancies.

If a benchmark has multiple released subsets or versions, prefer a conservative row-level summary and document the version-specific details in notes.

🛠️ Maintenance Notes

This public snapshot is maintained as a live CSV registry. Some rows intentionally leave code_url empty when a stable public code repository was not verified, and some data_url entries point to official access instructions rather than direct downloads. url_status records the current follow-up state for those cases. GSTA favors traceability over completeness: uncertain URLs should remain blank until verified against the original release channel.

📚 Citation

If you use GSTA, please cite this dataset and the original benchmark papers you rely on.

@misc{gsta2026,
  title        = {GSTA: Geo-SpatioTemporal Atlas},
  author       = {Dawei Zhang},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  url          = {https://huggingface.co/datasets/zhangdw/GSTA}
}

---

GSTA aims to make geospatial-temporal evaluation easier to discover, compare, and maintain.