Skip to content
This repository has been archived by the owner on Jun 12, 2020. It is now read-only.

radiantearth/training-data-stac-spec

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 

Repository files navigation

This repository is archived. For the latest specification of STAC training data please refer to label extension in radiantearth/stac-spec repository.

Training Data STAC Profile

The Training Data STAC Profile is an extension of the SpatioTemporal Asset Catalog core to handle 'training data' assets. Training data is earth observation imagery (and potentially other types of assets) with associated labels that describe what is in the imagery. Labels typically describe a set of geospatial 'things' that are contained in the image, be it forests, roads, ships or walmarts. The training data specification is agnostic as to what is actually labeled - the structure specified can be used to describe anything that can be identified on the earth. The primary use of training data is as input in to Machine Learning models, training them to automatically recognize or segment the same types of objects in new imagery. But the labels can be used for a variety of purposes.

Training Data STAC Item

The core of the Training Data profile is a STAC Training Data Item that extends the core STAC definition. Some of the meanings have been tweaked and refined. There are two required assets for a Training Data Item - it should have a source asset that the labels are created from, and

element type info name description
id string Training Data Item ID The ID of the Item assigned by the Labels Creator
geometry geojson Geometry A polygon of the area of the image that labels are valid for, in lat/long (EPSG 4326),
datetime date and time Date and Time The searchable date/time of the source asset for the labels, in UTC (Formatted in RFC 3339)
links array Resource Links Dict of link objects to resources and related URLs (self required)
assets array Assets Dict of asset objects that can be be download: 'source-asset' and 'labels' required, with thumbnail strongly recommended)
td:contributor string Contributor The name of the contributor who created the Training Data Item
td:method string Method The method of gathering the training data - from an image, from ground truth, etc (TODO: pick the possible values)
td:task_type string (enum) Task One of 'Tile Classification', 'Object Detection', 'Segmentation', or 'Other'
provider string Provider (optional) The provider of the labels (image provider is in the input-asset
license string Data License (optional) Item's license name based on SPDX License List or following guidelines for non-SPDX licenses

Releases

No releases published

Packages

No packages published