Datasets
Build a reproducible training input from imagery, finalized labels, and reference spectra. A dataset version records those inputs and their preprocessing settings so you can train and compare models against a defined body of data.
Prepare a training dataset
- Create a draft and choose the classes, processing level, and target wavelengths for your task.
- Assign images and reference spectra to the appropriate data buckets. Review split settings to keep evaluation data independent of training.
- Resolve validation messages and finalize the version. Use the completed version to train a model.
Find and organize datasets
The Datasets table shows each dataset's latest-version status, classes, sensor and wavelength summary, project, and ownership metadata. Use the table to open, pin, filter, archive, restore, or delete datasets. When a project is active, the table is scoped to that project.
Dataset details
Open a dataset to use its three main tabs:
- Overview: version metadata, project, validation messages, general and advanced settings, finalization actions, and analysis results
- Images: available project files and the Training, Validation, and Test assignment groups
- Reference spectra: signatures assigned to dataset classes and to the Training or Validation bucket
The Sample review tab and the Analyze action are available only when Alpha features are enabled.
Settings and bucket assignments can be changed while a version is a draft. Finalization generates the dataset artifacts and locks that version. Use the version picker to inspect older versions or select New version; use Clone version (new dataset) when the result should be a separate dataset.