Read a spectral cube without turning a curve into an identity

ACTION

1. Name the class or continuous response.
2. Specify spatial and spectral support.
3. Define acquisition timing and field/lab reference protocol.
4. State what the model must not claim.

CHECK

• Target labels are independent of the image features.
• Reference samples represent the intended mapping domain.

OUTPUT

Target specification, sampling plan and non-claims.

ACTION

1. Confirm row, column and band axis order.
2. Read centre wavelength and bandwidth/FWHM for every band.
3. Record radiance/reflectance scale, units, fill value, CRS and transform.
4. Inventory quality layers, bad-band metadata and acquisition geometry.

CHECK

• Band count matches wavelength metadata.
• Wavelengths increase as expected and scaling reproduces plausible values.

OUTPUT

Cube contract and band inventory.

ACTION

1. Identify raw DN, radiance, apparent reflectance or surface-reflectance product.
2. Read atmospheric/radiometric processing provenance.
3. Keep radiance and reflectance products separately named.
4. Inspect calibration or validation targets when supplied.

CHECK

• The product level supports the proposed cross-pixel or cross-date comparison.
• Reflectance is not inferred from the file appearance.

OUTPUT

Measurement-level decision record.

ACTION

1. Apply fill, cloud/shadow and geometry masks relevant to the product.
2. Read sensor/product bad-band flags.
3. Review low-SNR, atmospheric-absorption and spectral-edge regions.
4. Keep masked bands in provenance even when excluded from analysis.

CHECK

• Masking follows metadata and observed QA, not a universal wavelength recipe.
• NaN/NoData are not converted to zero reflectance.

OUTPUT

Valid-pixel mask, valid-band mask and exclusion log.

ACTION

1. Overlay reference geometry on the cube.
2. Use plot/polygon summaries when a point does not represent the field footprint.
3. Record pixel count, statistic and variability.
4. Convolve high-resolution reference spectra to sensor response where comparison requires it.

CHECK

• Field and image timing/support are compatible.
• Spectral-library samples match material state and measurement conditions closely enough for the claim.

OUTPUT

Reference-linked spectral table.

ACTION

1. Plot class medians plus dispersion, not only one mean curve.
2. Inspect signal-to-noise and residual striping.
3. Compare spectra on the same units and wavelength support.
4. Mark absorption regions before calculating features.

CHECK

• Apparent separation exceeds within-class variation where claimed.
• Display smoothing has not silently replaced analysis values.

OUTPUT

Spectral QA plots and candidate-feature rationale.

ACTION

1. Choose raw bands, derivatives, ratios, continuum removal or dimensionality reduction from a physical hypothesis.
2. Fit scalers, PCA or feature selection on training folds only.
3. Preserve wavelength names and transform parameters.
4. Test sensitivity to preprocessing choices.

CHECK

• No validation or test pixels informed preprocessing.
• Feature count is defensible relative to independent sample count.

OUTPUT

Fold-safe feature pipeline and feature register.

ACTION

1. Group samples by site, flight line, plot or spatial block.
2. Keep nearby pixels from the same object in one fold.
3. Reserve a final test domain where feasible.
4. Record class balance and spectral coverage by fold.

CHECK

• No mixed-pixel neighborhood appears in both training and validation.
• Rare classes remain evaluable.

OUTPUT

Spatial fold registry and held-out test set.

ACTION

1. Fit a simple physically interpretable baseline.
2. Compare more flexible models under identical folds.
3. Tune hyperparameters inside training data only.
4. Retain class probabilities or prediction intervals.

CHECK

• Complexity improves held-out performance, not only training fit.
• Metrics include per-class or response-specific error.

OUTPUT

Validated model candidates and selection record.

ACTION

1. Apply the exact fitted preprocessing pipeline.
2. Mask invalid pixels and extrapolative spectra.
3. Map uncertainty, confidence or spectral distance.
4. Inspect seams, striping and impossible salt-and-pepper patterns.

CHECK

• Output grid and valid-data mask match the source contract.
• Low-confidence and outside-domain pixels remain distinguishable from target absence.

OUTPUT

Prediction, uncertainty and applicability rasters.

ACTION

1. Save source identifiers, band table and masks.
2. Export training/fold registry, fitted pipeline and model card.
3. Deliver prediction, uncertainty and applicability rasters.
4. Document supported claims, sensitivity and non-claims.

CHECK

• Every feature can be traced to wavelengths and preprocessing.
• A reviewer can reproduce the split and final map.

OUTPUT

Analysis-ready cube evidence, model and audit trail.