In brief: Multimodal AI embeddings with Amazon Nova enable semantic natural language search across aerial image archives without task-specific training.
AWS and aerial imagery provider Vexcel have developed a system that makes billions of aerial image pixels searchable through multimodal AI embeddings – without training specialized models for each new task. With Amazon Nova Multimodal Embeddings, users can query aerial imagery in natural language to locate pools, road networks, or solar panels, for example.
The traditional approach to analyzing aerial image archives requires either manual inspection of individual map tiles or training a custom computer vision model for each new question. Vexcel, which operates one of the world’s largest aerial image capture programs and collects high-resolution data in over 45 countries, conducted experiments with the AWS Generative AI Innovation Center: multimodal embeddings, LLM-driven image captioning, and vector search via Amazon Bedrock and Amazon OpenSearch Serverless.
Evaluation was based on OpenStreetMap data as ground truth. Four key experiments compared different embedding models, fusion strategies (how multiple viewing angles per location are combined), image-text integration, and search methods. Amazon Nova Multimodal Embeddings achieved the highest F1 scores in evaluation. The approach eliminates the need to train a specialized model for each new query – instead, the image library is indexed once and various questions are answered through natural language queries within seconds.
The core technical challenge of aerial image search differs fundamentally from search in consumer photo collections: aerial images show a location not from a single perspective, but as multi-view datasets – different viewing angles of the same location. The fusion of these perspectives, the decision on image captions (whether they actually improve search quality or only increase costs), and the choice of embedding model critically shape search accuracy.
Vexcel has incorporated the results into a product offering called Vexcel Intelligence, currently in preview phase. The system unlocks the aerial image library as a searchable, AI-queryable resource and addresses use cases in insurance, real estate, administration, infrastructure, and agriculture.
Source: aws.amazon.com · Published June 22, 2026
Lumi AI News — AI-assisted curation in accordance with Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.1.