Multimodal AI turns aerial imagery into searchable data
AWS and Vexcel, an aerial imagery provider operating across 45+ countries, developed a multimodal AI system that converts billions of aerial images into natural-language-searchable data without requiring per-feature model training. The system uses embedding models, LLM captioning, and vector search to index imagery once and query it with plain English. Amazon Nova Multimodal Embeddings delivered the highest F1 scores in their evaluation, and the work evolved into Vexcel Intelligence, a commercial searchable imagery product.
TL;DR
- Vexcel and AWS built a semantic search system for aerial imagery using multimodal embeddings and vector search on Amazon Bedrock and OpenSearch Serverless
- The system eliminates per-feature model training, allowing natural-language queries across millions of images without manual tile-by-tile inspection
- Amazon Nova Multimodal Embeddings achieved the highest F1 scores in their evaluation across benchmark queries
- The architecture handles multi-view oblique imagery from Vexcel's fleet operating across 45+ countries and territories
Why It Matters
Geospatial data is critical for insurance, real estate, government, infrastructure, and agriculture, but converting billions of pixels into actionable intelligence has required either manual inspection or training custom models for each new query. This work demonstrates a scalable alternative using commodity multimodal AI and vector databases, reducing the time from question to answer from weeks to seconds.
Business Impact
Organizations relying on aerial imagery can now answer ad-hoc geospatial questions without engineering overhead or labeled training data. The approach reduces operational friction for use cases like locating swimming pools, identifying road networks, counting solar panels, or detecting specific building features, making geospatial intelligence accessible to non-technical stakeholders.
Key Implications
- Multimodal embeddings plus vector search can replace bespoke computer vision pipelines for geospatial queries, lowering barriers to entry for imagery-based analysis
- LLM captioning of aerial imagery may add cost without proportional search quality gains, requiring careful evaluation of fusion strategies
- The architecture is generalizable across industries and geographies, as demonstrated by Vexcel's 45+ country coverage and the emergence of Vexcel Intelligence as a commercial product
- Index-once, query-many approaches reduce time-to-insight for exploratory geospatial analysis and enable rapid iteration on new use cases
What to Watch
Monitor adoption of similar multimodal embedding approaches in other geospatial and imagery-heavy domains. Track whether vector search becomes the default for large-scale imagery retrieval and whether LLM captioning proves cost-effective as embedding model quality improves. Watch for competitive offerings and whether this pattern extends to other sensor modalities beyond aerial imagery.
Subscribe to the newsletter
The latest stories and analysis, delivered to your inbox.
Free. No spam. Unsubscribe any time.
