HSI reflectance calibration: white references, dark frames and illumination

Define what the calibrated spectrum represents

A hyperspectral camera initially records signals shaped by the sample, illumination, optics and detector response. To compare material spectra, a reflectance workflow uses reference measurements to account for some of those influences. The aim is a defined, repeatable measurement rather than a cube whose values merely look plausible. Record the wavelength axis, acquisition settings and reference data alongside the sample.

Reflectance is a ratio, but the measurement geometry matters. A directional camera observation of a glossy leaf or curved surface is not a geometry-independent material constant, and it is not hemispherical albedo. Here, reflectance means an estimated reflectance factor under the stated illumination and viewing arrangement. That distinction matters when comparing spectra collected with different lighting or viewing angles. [5][7]

Sources: [1], [5], [7]

Capture a dark frame with incoming light blocked

A dark reference measures the detector with incoming light blocked, typically by a shutter or opaque lens cap. It estimates the background signal under the acquisition settings. Use the instrument’s documented procedure, and average several frames where the workflow supports it. Exposure, gain and detector conditions affect which dark measurement is appropriate; a convenient old frame may no longer represent the current acquisition. [1][2]

An illuminated black tile is not a dark frame. It can reflect some incident light: calibrated diffuse standards are sold with nominal reflectances as low as 2%, rather than zero. Such a target can help assess reflectance response or linearity, but it does not measure the detector without incident photons. The distinction follows directly from what each measurement records. [1][3]

Sources: [1], [2], [3]

Use the reference panel’s certified spectrum

The illuminated panel reference records the combined response of illumination, optics and detector, multiplied by the panel’s own reflectance. A suitable panel need not be perfectly white, but its certified reflectance must be known across the useful wavelengths. Assigning every band a value of 1 silently assumes a perfect reflector. Use band-specific panel values and keep its calibration information with the dataset. [3][5]

Place the panel at the sample’s measurement height and capture it under comparable lighting and collection geometry. A height mismatch can change the light reaching the reference relative to the sample and bias spectral levels. A flat panel also cannot remove every shadow or specular highlight on an irregular object. [2][6]

Sources: [2], [3], [5], [6]

Apply the ratio only within its assumptions

For linear, unsaturated measurements with matching exposure and gain, the band-wise model is ρsample(λ) = ρpanel(λ) × [S(λ) − D(λ)] / [W(λ) − D(λ)]. Here, S is the sample signal, W is the illuminated panel signal, D is the dark signal and ρpanel is the certified panel reflectance at that band. The ratio before panel scaling is relative to the panel. [4][5]

Evaluate the denominator before division. Reject saturated data and mask bands whose dark-subtracted reference is too weak for a reliable estimate. A ratio cannot restore clipped measurements or create useful signal where illumination is absent. Keep invalid-band masks visible in later analysis rather than interpreting unstable ratios as absorption features.

If sample and reference exposures differ, do not use their unadjusted counts in this same-exposure formula. A validated linear-response workflow can subtract an appropriate dark measurement at each exposure and normalize by exposure time. Verify that model for the instrument and settings being used. [4]

Sources: [2], [4], [5]

Check a synthetic 49.0% reflectance calculation

Consider one illustrative band at 750 nm. Suppose the sample signal is 5,100 digital numbers, the illuminated panel signal is 10,100 and the dark signal is 100. Let the panel’s certified reflectance at that band be 0.98. These are invented teaching values, with matching 10 ms exposures, fixed gain, stable detector conditions and comparable geometry; they are not measurements or a product specification.

The dark-subtracted signals are 5,000 and 10,000. Their ratio is 0.5, and panel scaling gives 0.98 × 0.5 = 0.49: a reflectance factor of 49.0%. Assuming a perfect panel would instead give 50.0%, a one-percentage-point difference in this example. The interactive reflectance calculator reproduces this arithmetic.

The interactive example changes synthetic signals and panel reflectance to expose the same relationship. Its calculated output explains reference normalization; it does not establish instrument accuracy or uncertainty.

Sources: [4], [5]

Match illumination to the bands and the sample

A source that looks bright can still provide weak signal in the wavelengths needed for analysis. Tungsten–halogen sources commonly provide broad illumination for visible, NIR and SWIR reflectance or transmission work, but blue output and sample heating can matter. The entire fixture matters too: filters, reflectors or light guides may remove infrared output. Inspect the resulting reference signal across the planned bands. [8]

A visible-only white LED cannot illuminate NIR or SWIR bands for which it supplies no useful light. That does not mean all LEDs are visible-only: a published laboratory HSI system used an engineered LED source spanning 400–1000 nm. Choose from the measured source spectrum and system response, not the lamp’s colour or technology label. [9]

Fluorescence needs a different plan. Its source excites emitted light, so broadband illumination across every detection band is not necessarily appropriate. Do not apply a reflectance-panel workflow to fluorescence without defining the intended measurement. [8]

Sources: [8], [9]

Make the acquisition repeatable and inspect failures

Let equipment stabilize according to its documented workflow. Set exposure using the full spatial and spectral frame so bright bands and highlights remain unsaturated. Keep lamp positions, working distance, viewing angle, gain and exposure consistent between references and sample. Repeat references after relevant changes, and verify sample heating and spectral stability over the intended acquisition time. [1][2][8]

Use an enclosure or baffling where practical to reduce ambient light and unwanted reflections. Dark subtraction does not measure illuminated instrument stray light: photons scattered between wavelengths or spatial positions require separate characterization. NIST describes that characterization as a distinct calibration problem. Ordinary dark and panel references also do not establish wavelength accuracy, spatial registration or freedom from glare. [10]

Sources: [1], [2], [8], [10]

Use a wider processing model for remote sensing

The close-range ratio is useful within its stated assumptions. UAV, airborne and orbital observations introduce varying solar illumination, viewing geometry and an atmospheric path. A calibrated radiance product and a surface-reflectance product represent different processing stages. EnMAP’s L2A processor, for example, applies atmospheric correction to top-of-atmosphere radiance; NEON documents further geometry-related corrections for its airborne reflectance products. [7][11][12]

Start with the product definition, quality masks and processing history when using remote imagery. For capture methods and measurement scales, continue with the hyperspectral acquisition guide. The satellite comparison explains mission sampling and access, while the Australian HSI resources connect research communities and funding information.

Sources: [7], [11], [12]

Frequently asked questions

References and further reading

1. Resonon: Spectronon basic data acquisition and reference capture
2. Specim: Optimizing laboratory hyperspectral imaging
3. Labsphere: Spectralon diffuse reflectance standards
4. Shaikh et al. (2021): Calibration of a Hyper-Spectral Imaging System Using a Low-Cost Reference
5. Pillay, Hardeberg and George (2019): Hyperspectral Calibration of Art: Acquisition and Calibration Workflows
6. Specim: The role of white reference in obtaining reflectance data
7. NSF NEON: Introduction to bidirectional hyperspectral reflectance data
8. Specim: Illumination sources
9. Stergar, Hren and Milanič (2022): Laboratory hyperspectral imaging using a broadband LED light source
10. Woodward et al. (2009), NIST: Hyperspectral Imager Characterization and Calibration
11. Langheinrich, de los Reyes and Bachmann (2023), DLR: The EnMAP L2A Processor
12. NSF NEON: Imaging spectrometer calibration and reflectance processing