Skip to main content

Models overview

Turn a finalized dataset into a model you can evaluate and apply to new imagery. The Models page keeps training settings, model versions, metrics, and test predictions together.

Choose the output you need​

Clarity supports four model types:

  • Classification assigns one class to each pixel.
  • Unmixing estimates a continuous response for each class in each pixel.
  • Target Detection scores pixels for one selected target class.
  • Regression predicts continuous values for the dataset's numeric label attributes.

See Model types for requirements and output differences.

Train, evaluate, and apply​

  1. Create a model using a finalized dataset version and the model type that matches your task.
  2. Train the version, then review its metrics and held-out Test Predictions. Keep Test data untouched during model selection.
  3. Run inference on compatible imagery, reviewing processing provenance and wavelength coverage before submission.
  4. Create a new model version when settings need to change, preserving the previous version for comparison.

Model details​

Select a model to open its details drawer. The Overview tab contains the model metadata, dataset and band settings, training controls, and metrics. Test Predictions contains predictions generated from held-out test data.

Some managed or administrator accounts also see Compare and Deployments. These tabs are not available to every user.

Each model can have multiple versions. A version can be:

  • a draft whose settings can still be edited;
  • training or generating test results;
  • trained and ready for compatible inference; or
  • deprecated and no longer offered for new inference.

Common tasks​