SkyScraper: A Multi-Agent Feedback System for Automatic Detection of News Events from Satellite Imagery

This article introduces the SkyScraper system, which uses an iterative multi-agent workflow to geocode news articles and match them with satellite image sequences, addressing the scarcity of multi-temporal remote sensing event description datasets. It successfully detects 5 times more events than traditional methods and constructs a multi-temporal remote sensing dataset containing 5000 sequences.

Satellite remote sensing image changes emerge gradually, but multi-temporal event description datasets (≥2 images) are scarce due to time-consuming search and annotation. Traditional methods rely on manual annotation or rule-based processing, only handling bi-temporal image pairs and focusing on land use and land cover changes; recent LLM methods still depend on pre-labeled datasets, with multi-temporal descriptions limited to the drone video domain.

SkyScraper is an iterative multi-agent workflow with five steps: 1. Extraction (LLM extracts geographic entities and timelines); 2. Geocoding (Mapbox API converts coordinates); 3. Image Acquisition (PlanetScope imagery); 4. Verification (multi-modal LLM cross-validates event visibility); 5. Description (generates change descriptions). The key innovation is iterative feedback: when geocoding or verification fails, it uses failure information to request new candidate locations and optimize the search.

The research team used 1000 news articles and compared the weighted centroid, GIPSY, and SkyScraper methods. The results show that SkyScraper detected nearly 5 times more events than traditional methods. Reasons for improvement: agent verification eliminates false positives, iterative learning optimizes search, and multi-modal fusion enhances accuracy.

Applying SkyScraper to 2022-2024 GDELT news articles and using PlanetScope imagery to build a multi-temporal description dataset, the team obtained the SkyScraper GDELT dataset containing approximately 5000 sequences after annotator verification. They also generated a Sentinel-2 version, demonstrating large-scale data curation capabilities.

Application value: Supports journalism (provides visual evidence), disaster response (detects disaster impacts), urban planning and environmental monitoring (tracks changes). Future outlook: With the development of satellite technology and AI agents, it will play a greater role in earth observation and other fields.

Technical insights: Agent feedback is superior to single-round reasoning, multi-modal verification improves reliability, and modular design facilitates expansion. Limitations: Dependence on news quality, satellite imagery availability, high computational cost, and possible errors in verification agents. Conclusion: SkyScraper is an important advancement in remote sensing analysis and provides a practical case for AI directions such as multi-agent collaboration.