Skip to main content
Version: 0.13.0

Image regions tool

Introduced 0.13.0

The ImageRegionsTool, which an agent sees as FindImageRegions, reads a raster and returns the regions in it as labelled polygons — cropland, water, built-up, a burn scar — with areas. The work runs through _ingest/pipeline/_simulate; the labelled polygons are the tool's return value.

Two backends, one contract​

When you pass values (class ids to keep), the tool uses vectorize: it reads the labels from the pixels. score is 1.0 (the class is read from the pixel value).

Otherwise it uses image_segment: a segmentation model finds the regions. score is the model's confidence. Configure that model on the image_segment ingest processor. This tool takes no separate model_id.

Every answer names which backend produced it.

Parameters​

ParameterRequiredDescription
uriYesThe raster to read (https, s3, gs, azure).
bandNoWhich band holds the classes, for vectorize.
valuesNoComma-separated class values to keep. Their presence selects vectorize.
windowNoPixel window to read, as x,y,width,height.
max_regionsNoHow many regions to describe. Default is 20.

Register a flow agent that runs the ImageRegionsTool​

POST /_plugins/_agent/agents/_register
{
  "name": "Raster_Agent",
  "type": "flow",
  "description": "reads what is in a raster",
  "tools": [
    {
      "type": "ImageRegionsTool",
      "name": "DemoFindImageRegions"
    }
  ]
}
POST /_plugins/_ml/agents/<agent_id>/_execute
{
  "parameters": { "uri": "https://example.com/scene.tif" }
}