What challenges did Vexcel face before AI implementation?

Turn billions of pixels across multi-view aerial imagery into a natural-language-searchable knowledge base without manual inspection or bespoke retraining for each query. Evaluate which embedding model, fusion strategy, captioning approach, and search method work best for geospatial semantic search.

What was the impact of the AI solution?

Caption integration improved F1 scores by 11% for pools and 13% for roads. Amazon Nova Multimodal Embeddings achieved the highest average F1 scores across benchmark queries. The concepts evolved into a preview searchable product spanning Vexcel's global imagery library.

Source

ExploringEvidence: Medium50/100

Embed the world: Multimodal AI for searchable aerial imagery at scale

Use case typeSemantic searchUpdated Jun 22, 2026

Vexcel worked with AWS to evaluate multimodal embeddings, captioning, fusion strategies, and vector search for turning multi-view aerial imagery into a natural-language-searchable knowledge base. The system used Amazon Bedrock, Amazon OpenSearch Serverless, Amazon S3, AWS Secrets Manager, and automated evaluation against OpenStreetMap ground truth across about 100 configurations. The solution evolved into a preview product for searchable vector embeddings across Vexcel's global imagery library spanning 45+ countries.

Organization: Vexcel
Industry: Tech & Comms
Location: United States
Published: June 2026

Reported outcomes

+11%

F1 improvement for pools with captionsOther quantified impact

+13%F1 improvement for roads with captions

Strategic outcomes

Other strategic outcomeTurned aerial imagery into a searchable knowledge baseNew product / capabilityEvolved into a preview searchable product

Why do we believe this?Outcome claims, sources, and evidence checks

Normalized claim

F1 improvement for pools with captions: 11% increase

AWS Machine Learning BlogJun 22, 2026Blog postInferred claimMedium evidence strength

an 11% F1 score improvement for pools

Normalized claim

F1 improvement for roads with captions: 13% increase

AWS Machine Learning BlogJun 22, 2026Blog postInferred claimMedium evidence strength

13% for roads

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: Vexcel
Provider: AWS
Maturity: Exploring
Linked source: AWS Machine Learning Blog

The system used Amazon Bedrock, Amazon OpenSearch Serverless, Amazon S3, AWS Secrets Manager, and automated evaluation against OpenStreetMap ground truth across about 100 configurations

Customer identity supportedSource describes one deploymentMaturity supported

Primary read

Use case focus

Showing 2 of 2

1Semantic search
2Multimodal analytics

Turn billions of pixels across multi-view aerial imagery into a natural-language-searchable knowledge base without manual inspection or bespoke retraining for each query.
Evaluate which embedding model, fusion strategy, captioning approach, and search method work best for geospatial semantic search.

Built a five-stage modular pipeline for AOI selection, imagery ingestion, embedding and indexing, search, and evaluation.
Tested Amazon Nova Multimodal Embeddings, Amazon Titan Multimodal Embeddings, and Cohere embeddings with caption integration and metadata filtering.
Used OpenStreetMap ground truth and a configurable evaluation harness to compare search quality across many configurations.

Technologies

Amazon Bedrock Amazon OpenSearch Serverless Amazon S3 AWS Secrets Manager Amazon Nova Multimodal Embeddings Amazon Titan Multimodal Embeddings Amazon Nova Lite Cohere Embed v4 OpenStreetMap

Caption integration improved F1 scores by 11% for pools and 13% for roads.
Amazon Nova Multimodal Embeddings achieved the highest average F1 scores across benchmark queries.
The concepts evolved into a preview searchable product spanning Vexcel's global imagery library.

Architecture

A five-stage modular pipeline: AOI selection persisted to Amazon S3; imagery ingestion from Vexcel's API with Amazon S3 caching and AWS Secrets Manager for credentials; embedding and caption generation using Amazon Bedrock models; indexing in Amazon OpenSearch Serverless or Amazon S3 Vectors; natural-language search with multi-view fusion, image-caption fusion, and metadata filtering; and automated evaluation against OpenStreetMap ground truth.

Sources & evidence1

Evidence: Medium50/100Evidence strength