Spectral Harmonization

1. Scope & General Multi-Sensor Contract

Spectral Harmonization converts multiple Surface Reflectance rasters from different sensors into one common reference spectral domain. The tool is intentionally sensor-agnostic: built-in sensor profiles are conveniences for metadata and response definitions, not a whitelist of allowed sensors.

2. Reference Raster & Target Spectral Definition

2.1 Reference Raster

The Reference Raster is the normal center of the workflow. It can come from a built-in or unknown sensor and is used to define the target/reference domain whenever Target Spectral Definition = Auto — From Reference Raster.

Input Rasters
├─ WorldView-3 Surface Reflectance
├─ Pléiades Surface Reflectance
└─ Beijing-3A Surface Reflectance

Reference Raster
└─ Beijing-3A Surface Reflectance

Target Spectral Definition
└─ Auto — From Reference Raster

Result
├─ WorldView-3 → Beijing-like raster
├─ Pléiades   → Beijing-like raster
└─ Beijing reference → passthrough/reference-domain raster

2.2 Target Spectral Definition

3. General Sensor Support

The engine resolves spectral information in a general way. Existing raster metadata wins; known profiles only fill missing metadata when the match is defensible.

3.1 Current built-in shortcut profiles

Additional sensors can be supported without redesigning the tool by adding a profile, supplying exact RSR/SRF, defining manual bands, providing user SBAF coefficients, or using an overlapping reference raster for empirical regression.

3.2 Unknown/custom sensors

4. Method 1 — RSR / SRF Spectral Convolution

RSR/SRF convolution is strongest when the source has dense spectral sampling, such as EMIT or another hyperspectral raster, and the target spectral response is available.

The target response can come from exact Custom RSR/SRF curves or from center wavelength + FWHM definitions. Center/FWHM profiles are converted to a Gaussian approximation:

Under the default Common Comparable Target Bands Only policy, the engine keeps only target bands supported by every selected input. It does not invent missing spectral coverage.

5. Method 2 — SBAF / Spectral Band Adjustment Factor

SBAF applies a validated or user-supplied linear transformation between comparable source and target bands.

5.1 General coefficient model

SBAF is not inherently a Sentinel-2-only method. In the general architecture, a coefficient set belongs to a specific source → target relationship and should only be used for that validated pair.

5.2 Sentinel-2 in the current implementation

Sentinel-2A/B/C remains supported internally for the validated HLS-style MSI→OLI profile. Platform detection is automatic per input and is not a global user-facing requirement.

Sentinel-2A/B/C Surface Reflectance
        ↓
Auto detects platform
        ↓
Validated MSI → OLI SBAF available
        ↓
Use matching HLS-style coefficient table

6. Method 3 — Empirical Cross-Sensor Regression

Empirical regression learns the mapping between a source raster and an actual overlapping Reference Raster. It is particularly useful for arbitrary commercial or custom multispectral sensors when no validated SBAF exists.

6.1 Models

Without explicit band mapping, the engine tries wavelength matching, then semantic roles such as Blue/Green/Red/NIR/SWIR, and finally same-position fallback when dimensions match.

7. Auto Routing — Resolved Independently per Input

Auto is now resolved per source raster. A mixed collection can therefore use different methods while all outputs converge to the same reference domain.

7.1 Mixed commercial example

Inputs:
WorldView-3
Pléiades
EMIT
Beijing-3A

Reference:
Beijing-3A

Method = Auto

Possible routing:
WorldView-3 → Empirical Regression → Beijing domain
Pléiades    → Empirical Regression → Beijing domain
EMIT        → RSR Convolution      → Beijing domain
Beijing     → Reference Passthrough

8. Target Bands & Spatial Grid

8.1 Target Band Policy

8.2 Spatial Grid

Spatial resampling options are nearest, bilinear (default), cubic and average.

9. Application Parameters

10. Multi-Output Collection & Provenance

Every selected input produces one child raster output. Collection metadata records both the requested method and the actual method used per input.

{
  "collection": {
    "collection_type": "spectral_harmonization",
    "raster_count": 4,
    "requested_method": "auto",
    "methods_used": [
      "empirical_regression",
      "rsr_convolution",
      "reference_passthrough"
    ],
    "method_by_input": {
      "worldview3_scene.tif": "empirical_regression",
      "pleiades_scene.tif": "empirical_regression",
      "emit_scene.tif": "rsr_convolution",
      "beijing_reference.tif": "reference_passthrough"
    },
    "target_definition": "reference_metadata",
    "target_sensor_profile": "beijing3a",
    "target_label": "Beijing-3A ...",
    "reference_raster": "beijing_reference.tif"
  },
  "outputs": [
    {"source":"worldview3_scene.tif", "method":"empirical_regression"},
    {"source":"pleiades_scene.tif", "method":"empirical_regression"},
    {"source":"emit_scene.tif", "method":"rsr_convolution"},
    {"source":"beijing_reference.tif", "method":"reference_passthrough"}
  ]
}

Each child output also carries source name, detected source sensor/profile, processing level, target definition, target profile/label, band policy, spatial grid and method-specific analysis metadata.

11. Recommended Workflows

WorldView-3 SR
Pléiades SR
+ Beijing Reference SR
↓
Auto / Empirical
↓
Beijing-like common bands
per input

EMIT L2A SR
+ target/reference RSR
↓
RSR Convolution
↓
Target-like bands
per scene

Sentinel-2 L2A
+ Landsat OLI target
↓
Validated HLS-style SBAF
(platform auto-detected)
↓
OLI-like common bands

12. Validation & Quality Control

• Use decoded physical Surface Reflectance. Auto validation rejects known DN/radiance/TOA inputs and implausible physical ranges.
• For RSR convolution, reliable source wavelengths and a defensible target spectral response are essential.
• Prefer exact official RSR/SRF curves over nominal center/FWHM Gaussian approximations.
• For SBAF, coefficients must belong to the correct source/target relationship; do not reuse Sentinel→Landsat coefficients for WorldView→Beijing or other unrelated pairs.
• For empirical regression, use close acquisition dates, good registration, cloud/shadow masking and stable surfaces.
• Check semantic band roles and wavelength matching, especially for NIR, red-edge and SWIR bands that may differ substantially between sensors.
• Use Common Comparable Target Bands Only when building cross-sensor time series unless every target band is scientifically supported.
• Record method_by_input when Auto is used, because different sources in the same collection may have different harmonization methods.
• Do not interpret preserved extra source bands as harmonized target bands.

13. Scientific & Technical References

1. NASA Harmonized Landsat Sentinel-2 (HLS). Algorithms — Bandpass Adjustment. Describes OLI-referenced MSI bandpass adjustment derived from hyperspectral spectra convolved with OLI/MSI RSRs. NASA HLS.
2. Claverie, M. et al. (2018). The Harmonized Landsat and Sentinel-2 surface reflectance data set. Remote Sensing of Environment 219, 145–161. DOI: 10.1016/j.rse.2018.09.002.
3. Roy, D. P. et al. (2016). Characterization of Landsat-7 to Landsat-8 reflective wavelength and normalized difference vegetation index continuity. Remote Sensing of Environment 185, 57–70. DOI: 10.1016/j.rse.2015.12.024.
4. U.S. Geological Survey. Landsat Collection 2 Surface Reflectance documentation. USGS.