Geospatial intelligence has historically required specialized analysts with GIS expertise to extract insight from satellite and earth observation data. AI is changing that by enabling natural language interfaces to complex spatial queries - and by making visual analysis of satellite imagery accessible without deep technical expertise in remote sensing.

The Problem with Traditional Geospatial Workflows

Satellite data is voluminous and technically demanding. Raw imagery arrives in formats (GeoTIFF, HDF5, NetCDF) that require specialized software to open. Spatial queries require SQL extensions (PostGIS) or dedicated GIS tools. Analysts need to understand coordinate reference systems, band mathematics, and temporal compositing before they can answer a basic question like “has the vegetation coverage in this region changed over the last year?”

This expertise barrier concentrates geospatial capability in specialist teams, making it slow and expensive to answer questions that cross into geospatial territory.

Multi-Agent Architecture for Geospatial AI

The multi-agent approach works well here because geospatial analysis naturally decomposes into distinct steps, each requiring different capabilities:

Query understanding agent - Receives a natural language question (“What is the extent of urban expansion near this logistics hub over the past 3 years?”) and decomposes it into a structured geospatial query: bounding box, time range, required data sources, analysis type.

Data retrieval agent - Queries public and licensed data sources. For open data, this includes Sentinel-2 (10m resolution optical imagery via Copernicus), Landsat (30m resolution, deep historical archive via USGS Earth Explorer), and OpenStreetMap for vector data. Data is retrieved from S3 (AWS hosts public Sentinel and Landsat archives in the Registry of Open Data on AWS) or from provider APIs.

Analysis agent - Processes retrieved imagery. For change detection, this typically involves computing a normalized difference index (NDVI for vegetation, NDWI for water, NDBI for built environment) for two time periods and comparing pixel-level values. SageMaker provides the compute for this; pre-trained models from SageMaker JumpStart handle common analysis types.

Synthesis agent - Combines analysis outputs with contextual data (industry reports, news feeds, public records) and generates a structured report. Bedrock with Claude handles this layer.

Public Data Sources

  • Sentinel-2 - 10m resolution multispectral imagery, 5-day revisit, global coverage, free via Copernicus
  • Landsat 8/9 - 30m resolution, continuous archive since 1972 (earlier with older sensors), USGS
  • SRTM/Copernicus DEM - Digital elevation models, free
  • OpenStreetMap - Road networks, buildings, land use, free
  • Sentinel-1 SAR - Synthetic aperture radar, cloud-penetrating, useful for monitoring where optical imagery is frequently cloud-covered

Implementation Notes

For most enterprise use cases, full multi-agent automation is the target state but not the starting point. A practical first version is a retrieval tool that fetches the right imagery and produces a visual output, with a human analyst doing the interpretation. The AI analysis layer is added incrementally as confidence in output quality builds.

Validation is non-trivial. Geospatial AI outputs should be checked against ground truth for a sample of the historical data before relying on them for decisions. The good news is that historical ground truth for many common analysis types (land cover change, flood extent) is available from academic and government sources.