Skip to main content

Advanced Dataset Settings

Advanced settings control dimensionality reduction, reference-spectrum preprocessing, and sample selection. They apply when a draft is finalized.

Principal Component Analysis​

Set Perform PCA to Yes to create a PCA transform and apply it to samples during dataset generation. Set Number of latent variables to the number of components to retain; the current editor accepts 2 through 100.

PCA analysis fits on training image samples and training reference spectra. The generator stores PCA scores separately from the spectral values in TFRecords. Select the component count for the analysis you intend to use.

PCA adds finalization time. When PCA analysis results are available, the visualization uses the leading components.

Spectral resampling​

Spectral resampling applies to image pixels and original reference spectra before PCA analysis and dataset generation. Trained models retain the target response definition in their artifacts and use it during batch inference.

Open the existing Target wavelengths dialog and choose the MSI image as the target. Use Resampling to choose Interpolation (with its method) or Band integration, then set the band response:

  • Target image bandwidths (Gaussian): use the chosen target image's centers and full widths at half maximum (FWHM). This supports calibrated industrial, airborne and satellite sensors without assuming that all images with the same sensor name have the same bands.
  • Custom bandwidths: one positive target FWHM in nanometres per existing target band.
  • Sensor target preset: measured response curves for Sentinel-2A/B/C MSI or Landsat 8/9 OLI; documented Gaussian approximations for PlanetScope Dove-R and SuperDove. Choosing a preset replaces the target grid; select its required bands before applying. Presets are available when the target grid is editable.
  • Import response profile: supply a versioned JSON profile with wavelength samples and nonnegative response values for each target band. Response curves must begin and end at zero. See profile format.

Interpolation aligns band centers using the selected interpolation method.

Use Filter target bands to open the band-index and wavelength filters. They save automatically and also select the corresponding response bands. Calibration details reveals the saved source and version. A saved response profile owns the target grid; switch to interpolation before choosing a different image target. Locked targets accept response data only for the existing bands.

Convolution requires bandwidths or response curves. Source data must cover each selected band's response and provide sufficiently fine valid measurements. Matching centers alone does not establish matching spectral responses. Unknown units, malformed bandwidths, missing measurements and unsupported bandpasses require correction rather than fabricated values. Spectral gap estimation is not combined with response matching.

A new target axis cannot change the wavelengths shared by finalized dataset versions. Clone into a new dataset to select different bands. Existing finalized datasets and models without a response profile retain their previous behavior.

Reference spectra​

Reference spectra normalization applies only to signatures added on the Reference spectra tab: None, Area, Max, Length or Sum. Their spectral resampling uses the shared target definition above.

Sampling​

The current editor offers two sampling methods:

  • Random selects samples without replacement.
  • Interleaved selects evenly spaced positions through each class's sample sequence.

Class split type defines the sample unit:

  • Pixel keeps individual labeled pixels as samples.
  • Label selects labels and averages the pixels in each selected label into one spectrum.

For Automatic data assignment, this setting also determines whether the training/validation split operates on pixels or whole labels.

An Automatic Pixel split creates disjoint pixel samples, but pixels from the same label and image can enter both Training and Validation. Spatially adjacent hyperspectral pixels are often correlated, so this is not an independent test of scene-level generalization. A Label split keeps each label region together, but Training and Validation can still share an image. To evaluate generalization across scenes, sites, sensors, dates, or acquisition conditions, use Manual assignment and hold out whole images or acquisitions.

If a class has only one eligible image sample or reference signature, the current Automatic split can place the same sample in both Training and Validation. Add independent samples instead of interpreting that validation metric as held out.

Optional per-class caps limit Training, Validation, and Test samples. Min threshold per pixel and Max threshold per pixel exclude values outside the configured bounds during generation.