Unmixing models
Unmixing produces a separate continuous response layer for each class. It is useful when multiple labeled materials can contribute to a pixel or when a class-by-class heatmap is more useful than a single categorical answer.
Requirements
- A finalized dataset version with at least two classes.
- At least two classes selected as model inputs.
- Representative training and validation examples.
The displayed values are model responses. Do not treat them as calibrated material percentages unless you have validated that interpretation for your data and configuration.
Evaluation
Use the metrics that match the configured evaluation target:
- Class labels: review threshold-dependent precision, recall, and F1 together with ROC and precision-recall behavior.
- Abundance vectors: review overall and per-endmember RMSE and MAE in the target abundance units. ROC, precision-recall, and decision thresholds do not apply to this mode.
Inspect the per-class output layers on appropriately separated evaluation data to check spatial and spectral failure modes.