The Big Picture
Using an iterative multi-agent feedback loop increases detection of news-linked events in satellite imagery by nearly 5× and enabled curation of about 5,000 multi-temporal labeled sequences.
ON THIS PAGE
The Evidence
An agent-driven pipeline that extracts candidate locations from news, fetches matching satellite images, and has a verifier agent confirm visibility finds many more visible events than simple rules. The system requests one candidate location at a time, retries with new candidates when geocoding or verification fails, and uses the verifier's reasoning to guide searches. In experiments on 1,000 news articles using PlanetScope imagery, the approach produced a large boost in true positive event detections versus weighted-centroid and polygon-based baselines. Applying the pipeline to the GDELT news feed produced roughly 5,000 curated multi-temporal image sequences with human-verified captions and dates. verifier's reasoning helps steer the search, illustrating an agent-driven pipeline that benefits from structured decision-making and feedback loops.
Not sure where to start?Get personalized recommendations
Data Highlights
1Nearly 5× increase in event detections compared with a weighted-centroid baseline
21,000 news articles used in the experimental comparison with PlanetScope (3 m) imagery
3Approximately 5,000 multi-temporal labeled sequences curated from GDELT (PlanetScope and Sentinel-2 versions)
What This Means
Engineers building datasets or monitoring pipelines from public news sources will benefit because the method automates finding where events are visible in imagery and reduces manual search. Technical leaders evaluating agent-driven tooling should note this as a practical pattern for agent orchestration and improving end-to-end yield. Researchers in remote sensing and dataset curation can reuse the curated multi-temporal sequences for training and evaluation. The concept aligns with established patterns like Orchestrator-Worker Pattern and relates to broader research in Research Agents.
Key Figures
Fig 1: Figure 1: SkyScraper iterative feedback pipeline.
Fig 2: Figure 2: PlanetScope imagery and caption for a geocoded tornado event identified by SkyScraper.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreKeep in Mind
Not all news events are visually visible in the available imagery; the system can only detect events that leave an observable trace at the sensor resolution and revisit cadence. The approach depends on the quality of geocoding and the multimodal language models used for extraction and verification, which can produce incorrect candidates or reasoning. Human verification was still required for final labels, so costs and biases in news sampling and manual annotation remain important caveats. For guardrails and verification rigor, see the Mutual Verification Pattern.
Methodology & More
SkyScraper turns news articles into findable, captioned multi-temporal satellite image sequences by chaining several specialized agents in a loop: one extracts candidate place names and event dates from the article, a geocoding step converts names to coordinates, imagery is fetched for those coordinates over the event timeline, a multimodal verifier checks whether the event is visible in the images and explains failures, and a captioning agent writes a change caption using the article for context. Rather than extracting all candidate locations at once, the pipeline requests one candidate at a time and retries when geocoding fails or the verifier does not find the event—each failure feeds back reasoning that guides the next attempt. That iterative, agentic feedback is the core innovation: the verifier’s judgments steer the search toward more promising coordinates instead of relying on name frequency or naive centroid rules. Sub-Agent Delegation Pattern and Consensus-Based Decision Pattern illustrate the governance of such iterative agent interactions. In experiments using 1,000 news articles and PlanetScope (3 m) imagery, the agentic loop produced nearly five times more true positive event detections than a weighted-centroid baseline and outperformed a polygon-stacking method. Applying the system to GDELT articles from 2022–2024 produced roughly 5,000 multi-temporal labeled sequences (PlanetScope and Sentinel-2 versions) that were manually verified for captions and dates. The main implication is practical: agent-to-agent feedback can substantially raise the yield when mining noisy text sources for visible events, making it easier to build larger, more diverse training and benchmark datasets. Still, visibility limits, model errors, and the need for human verification mean this is a strong augmentation to—but not a replacement for—careful dataset curation and human-in-the-loop review. For researchers in broader scientific contexts, this work aligns with Research Agents and related multi-agent patterns like the Orchestrator-Worker Pattern.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Company affiliation (Planet Labs) but low h-indices and no major academic venue or citations — emerging/limited signals.