Making Satellite Data Searchable with Natural Language

Image

Opening Up Earth Observation to Everyone

Satellite imagery holds enormous potential for environmental monitoring, urban planning, agriculture, and more, but tapping into that potential often requires specialized knowledge and tooling. While SpatioTemporal Asset Catalog (STAC) APIs are powerful and flexible, they can be challenging to use for those unfamiliar with geospatial data formats and query structures.

We’re prototyping an approach that aims to simplify this process. What if you could just type “Show me Sentinel-2 images of Berlin from September 2023,” and get a result instantly?

That’s the idea behind our natural language interface prototype. By combining modern NLP techniques, serverless architecture, and geospatial standards like STAC, we’re exploring how satellite imagery could become searchable using plain English.

Core Concept: Translating Language into Data Queries

The system interprets free-form language and translates it into structured STAC queries behind the scenes. It’s built on a serverless stack powered by AWS Lambda, with a FastAPI backend exposed through a REST interface. AWS Bedrock provides the foundation for language understanding tasks, such as identifying locations and interpreting time expressions.

Next we’ll have a look at what happens under the hood when a user submits a query.

System Breakdown

The architecture is modular and composed of several key components:

Query Processor

This module breaks down the natural language input to identify core elements; sensor type (e.g., Sentinel-2 or Landsat), location, and timeframe.

Location Processor

If a place is mentioned e.g. “Berlin”, “north of Paris”, or “the Alps”, the location processor geocodes it into a bounding box that is compatible with the STAC API.

Time Parser

Phrases like “last month”, “from 2023”, or “September 1st to 30th” are translated into standard ISO 8601 date formats suitable for STAC.

Collection Finder

Using the extracted parameters (sensor, location, date), the collection finder determines which satellite image collections are most relevant, ensuring the search is precise and efficient.

An Example in Action

Let’s say a user types:

“Show me Sentinel-2 images of Berlin from September 2023.”

The flow would look like this:

  1. Query Processor: Identifies the satellite (Sentinel-2), the place (Berlin), and the time period (September 2023).
  2. Location Processor: Geocodes “Berlin” to a bounding box covering the city.
  3. Time Parser: Converts “September 2023” into 2023-09-01 to 2023-09-30.
  4. Collection Finder: Matches the parameters with available collections in the satellite data catalogs. It can scan across multiple STAC catalogues and match the parameters with available collections in the satellite data catalogs, identifying the most appropriate collection(s).

The system builds a STAC query with these parameters and uses AWS Lambda to send the request and fetch matching assets/imagery. All this happens behind the scenes, without the user needing to understand what STAC, ISO dates, or bounding boxes are.

Dis

Managing the infrastructure as code

The infrastructure is defined and deployed using the  AWS Cloud Development Kit (CDK) . This allows the whole pipeline, including the Lambda function, API Gateway, and associated resources, to be managed as code, enabling rapid iteration and reproducibility.

Demo Use Case: A Natural Interface on the Web

To showcase the system, we developed a demo web app using Loveable. The app lets users type natural language queries into a search bar. The backend handles the processing, and relevant satellite images are visualized on an interactive map. For testing, we used the UK EO Data Hub STAC catalog as a data source.

No geospatial expertise needed—just a browser and a question.

Why This Matters

By abstracting away the technical complexity, this project democratizes access to satellite data. It empowers users; whether they’re researchers, journalists, or curious citizens, to explore Earth observation data in an intuitive and accessible way.

What’s Next?

Future enhancements could include:

  • Support for more complex spatial queries (e.g., multi-region requests)
  • Processing of data found (e.g. calculating NDVI)
  • Chatbot-style interfaces
  • Using AI to recommend appropriate visualisations for different types of geospatial data. For example, determining suitable colour ramps or visual representation techniques based on the type of dataset, such as using thermal gradients for Land Surface Temperature (LST), specific palettes for Sea Surface Temperature (SST), or shaded relief for sea ice thickness. This would be based on how such datasets are typically visualised in scientific and operational contexts.

This project is just one example of how AI and cloud-native technologies can lower the barrier to complex data systems. We’re excited about the results and as we continue prototyping, we’re especially curious about how this could unlock more intuitive tools to enable more people to access satellite data.