Skip to main content

Model training

Prerequisites​

Before creating a model, prepare:

  • a finalized dataset version;
  • training and validation data that satisfy the selected model type;
  • the intended processing level and wavelength set; and
  • representative labels for every output you need.

See Model types for type-specific requirements.

Create a model​

  1. Open Models and select New model.
  2. Enter a name and select a model type.
  3. Select the dataset and dataset version.
  4. Review the bands, processing level, and class or attribute settings.
  5. Create the model.

The new model opens as an editable draft.

Review settings​

On the Overview tab, review Settings before training. Expand Advanced Settings only when you need to change initialization, hyperparameters, thresholds, loss, preprocessing, inference, early stopping, or data augmentation.

The initial defaults include a Base CNN, batch size 512, 10 maximum epochs, the Adam optimizer, a fixed learning rate of 0.001, and no preprocessing or data augmentation. Defaults can change as the product evolves, so the values shown in the draft are authoritative.

Train and evaluate​

  1. Select Train.
  2. Leave the model drawer open or return later; training continues remotely.
  3. After training completes, inspect the training and validation results under Model Metrics, along with Dataset Metrics.
  4. Use Validation data to select preprocessing, hyperparameters, thresholds, checkpoints, and the model version.
  5. After the version is selected and locked, generate one test run if it does not yet have testing metrics.
  6. Open Test Predictions to inspect the untouched test predictions in their spatial context.

During training, End early requests an early stop. A stopped run may offer Continue training, Use saved model, or Restart from scratch, depending on the saved checkpoint state.

Do not judge a model from training loss alone. Validation supports model selection; untouched Test data estimates performance after selection. If a test result guides another iteration, it is no longer an independent final estimate.

Next steps​