Spectral processing

Spectral transforms produce another cube. They are distinct from feature extraction, which reduces spectra to maps. Keep a copy of the physically interpreted input and record every transform applied to it.

Band selection by wavelength

red = hp.band_at(ds, 660)
nir = hp.band_at(ds, 860)

The default nearest-band tolerance is 20 nm. This is a selection tolerance, not a claim that any band within 20 nm is adequate for a narrow feature. Inspect the actual selected wavelength and use tighter tolerances where scientifically needed.

Smoothing without bridging unusable regions

smoothed = hp.smooth_spectra(ds, window=7, order=2,
                             method="savgol", good_only=True)

Savitzky–Golay and moving-window options operate within usable-band runs. NaNs can be temporarily interpolated for filtering and then restored. The procedure alters spectral shape and is optional, not a substitute for atmospheric correction or wavelength calibration.

Spline gap filling

filled = hp.spline_gapfill(ds, df=60, threshold=0.018, despike=True)

This route intentionally fills selected spectral gaps after artifact masking and spline fitting, then masks designated water-absorption regions. spline_filled records interpolated values. The default exclusion windows are PRISMA-oriented; inspect their suitability before applying them to another sensor. An interpolated absorption is not an observation.

Continuum removal

continuum_removed = hp.continuum_removal(ds, window=(2000, 2300))

The spectrum is normalized by an upper convex hull within the selected interval and usable-band runs. The interval determines the continuum shoulders. Results depend on missing bands, noise, and the feature window; do not compare differently defined windows as if they were the same metric.

Derivatives

first_derivative = hp.spectral_derivative(ds, order=1, window=7, poly=2)

The implementation uses a Savitzky–Golay fit and the representative wavelength spacing of each run. Irregular spacing needs attention. Derivative values carry wavelength-dependent units and should not be described as surface reflectance despite sharing the source variable name.

Uncertainty and feature preservation

Several transforms discard uncertainty rather than propagate it. Check output variables and attributes. Compare narrow-feature analyses against unsmoothed spectra and report sensitivity to smoothing choices. See spectral API.