Skip to main content

Interpreting model results

Open a trained model version and use Overview > Model Metrics to review training, validation, and testing results when available. Dataset Metrics describes the data distribution. Test Predictions provides the spatial predictions for a generated test run.

Keep validation and testing roles separate​

Training curves show whether optimization progressed, but a low training loss does not guarantee useful inference. Use Validation data to select transforms, hyperparameters, thresholds, checkpoints, and model versions. Use untouched Test data only after those choices are locked. Repeatedly choosing changes from test results biases the final performance estimate.

Neither bucket necessarily represents a new scene. An Automatic Pixel split can place correlated pixels from the same labels and images in Training and Validation. Interpret a result as scene-level generalization only when whole scenes or acquisitions were held out by design.

Loss and error values do not have a universal 0-to-1 range. Their scale depends on the model type, loss function, target values, and aggregation.

Classification​

Classification metrics include accuracy, precision, recall, F1, support, and confusion matrices.

  • Precision asks how many predictions for a class were correct.
  • Recall asks how many labeled examples of a class were found.
  • F1 balances precision and recall.
  • Support is the number of evaluated examples for a class.

Use per-class metrics when classes are imbalanced. In a confusion matrix, a strong diagonal means predictions often match labels; off-diagonal cells show which class pairs are confused. Check the axis labels before interpreting a particular cell.

Unmixing and Target Detection​

Class-label Unmixing and Target Detection results can include threshold-dependent precision, recall or true-positive rate, false-positive rate, F1, and accuracy. ROC and precision-recall curves show how the tradeoff changes with the threshold. AUROC and AUPRC summarize ranking across thresholds, but neither chooses an operating threshold for you.

Prefer precision-recall behavior when positive examples are rare. Inspect spatial predictions as well: a good aggregate score can hide structured false alarms.

Abundance-vector Unmixing instead reports overall and per-endmember RMSE and MAE in the target abundance units. ROC, precision-recall, and decision thresholds do not apply to that evaluation mode.

Regression​

Regression metrics are reported by label attribute:

  • RMSE emphasizes larger errors.
  • MAE is the average absolute error in target units.
  • R² compares the model with a constant baseline and can be negative.
  • SMAPE expresses a symmetric relative error and can be unstable near zero.

Use predicted-versus-observed plots to look for bias, saturation, outliers, and gaps in the target range.

Before inference​

Choose metrics that match the consequences of your application, inspect representative successes and failures, and test imagery from the sensors and conditions you expect in production. Then follow Inference results.