Inference results
Inference applies a trained model version to imagery that passes the model picker's eligibility checks and adds the generated result to a workspace.
Eligibility and wavelength coverage
Clarity offers model versions whose:
- status is trained rather than draft or deprecated;
- processing level matches the selected image layer;
- required wavelengths have usable support under the selected inference policy; and
- project and account access permit their use.
These checks establish product eligibility, not scientific comparability. A matching processing-level name does not prove that two layers have the same units, calibration, atmospheric correction, or acquisition conditions. Review that provenance before inference.
The model picker marks a version Exact wavelengths match or Will be resampled. An exact match preserves the model's wavelength axis. When the source can supply every model-band position, Clarity linearly resamples it to the model wavelengths.
Missing or unusable wavelengths
Inference requires usable spectral support for the model's required wavelengths. Missing bands are not silently filled with zeros. Review the per-image support report when a batch contains incompatible inputs.
In the model's compatibility popover, Estimate missing wavelengths offers explicit reconstruction choices:
- Linear interpolation or Shape-preserving (PCHIP) for internal gaps.
- Constant edge extension or Linear edge extension beyond usable coverage.
Enter a positive wavelength limit for each enabled method. Review the updated compatibility results and confirm before generating. Only eligible pending images run; incompatible images remain skipped. Existing results retain their provenance and are not silently regenerated when these settings change.
Estimated inputs are assumptions, not additional measurements. Validate the chosen policy against representative data before relying on predictions. Result properties report the estimated input count, with full details retained in the result's spectral provenance.
Generate from Files
Use the shared Run analysis dialog to generate inference results.
- Open Files.
- Select compatible images and choose Run analysis, or select Run analysis… in an image's context menu.
- Choose Deep learning.
- Review the selected image layers.
- Select an eligible model and version, then review wavelength coverage.
- Select Run deep learning and follow the job status.
- When generation completes, select Add to workspace if the result is not already open.
Generate from a workspace
- Open the image in a workspace and select Explore.
- Expand the image source and its Analysis layers subsection.
- Select the Generate analysis layers plus button, choose Run analysis…, then Deep learning. The source context menu also offers Run analysis….
- Choose the input layer and an eligible model version, then review its wavelength coverage.
- Select Run deep learning and wait for the completed result.
The output appears as an inference layer in the workspace. Its rendering depends on the model type: a categorical layer for classification, class response layers for unmixing, a target score layer for Target Detection, or continuous attribute layers for regression.
Test predictions are different
The model drawer's Test Predictions tab contains predictions generated from the designated Test bucket. Those predictions are an independent held-out measure only if the Test data remained untouched during model selection. Use Run analysis > Deep learning for operational imagery that is not part of the model test run.
If no model is offered, check the version status, processing level, wavelength metadata and overlap, and project access.