Lesson 2.1.3: Difference between RGB, multispectral and hyperspectral imaging

Definition of RGB, Multispectral and Hyperspectral Imaging

Imaging technology is the backbone of modern remote sensing, medical diagnostics and industrial inspection. Its evolution has moved from simple visual capture to the acquisition of detailed spectral information across a wide range of wavelengths. In this section we unpack three key classes of imaging – RGB, multispectral and hyperspectral – and analyse the physical principles that differentiate them. By the end you will recognise how the choice of spectral sampling impacts data volume, analytical power and practical feasibility.

RGB Imaging – The Human Eye’s Analogue

RGB, or Red–Green–Blue imaging, is the most familiar form of digital photography. It mirrors the trichromatic vision of the human eye, where three types of cone cells (short, medium and long wavelength sensitive) provide the basis for colour perception. An RGB sensor comprises a regular array of pixels, each covered by a colour filter that transmits only one of the three primary bands. The sensor measures the integrated radiance within each band, producing three intensity values per pixel.

Key physical characteristics:

• Band count: Three discrete bands, typically centred around 620–750 nm (red), 520–580 nm (green) and 450–510 nm (blue).
• Spectral width: Each band covers a broad range (≈ 50–100 nm) to maximise photon capture and maintain high signal‑to‑noise ratio.
• Spatial resolution: Often very high, as the sensor can allocate many pixels to a single band. For example, a 24‑megapixel RGB camera can deliver sub‑millimetre detail on a ground target.
• Data volume: Low – a 10‑megapixel image occupies around 30 MB (RGB, 8‑bit per channel).
• Calibration: Straightforward; radiometric calibration maps pixel values to spectral radiance, but colour fidelity requires careful white‑balance and gamut mapping.

Concrete example: A consumer DSLR captures a landscape in RGB. The resulting image emphasises visual aesthetics: the sky appears blue, foliage green, and stone grey. The underlying spectral information is coarse; subtle differences in vegetation health or mineral composition are lost.

Multispectral Imaging – A Step Beyond Colour

Multispectral imaging extends the RGB paradigm by adding extra bands that cover portions of the visible, near‑infrared (NIR) and sometimes short‑wave infrared (SWIR) spectra. Each band is typically isolated by a narrowband filter, and the sensor may either be a single‑pixel array that scans sequentially (push‑broom) or a multi‑band focal plane array that captures all bands simultaneously.

Key physical characteristics:

Concrete example: The Landsat 8 satellite carries the Operational Land Imager (OLI) with 11 spectral bands, including a dedicated NIR band at 850 nm. A single pixel represents a 30 m square on the ground. By combining the red, green, blue and NIR bands, users can compute the Normalised Difference Vegetation Index (NDVI), a powerful indicator of vegetation vigor that is invisible to the naked eye.

Comparison with RGB:

“Where RGB is the visual snapshot, multispectral is the first analytical glance.”

Multispectral imaging balances spectral detail against data volume, making it suitable for large‑scale monitoring (agriculture, forestry, urban planning). However, the limited number of bands means that fine spectral signatures (e.g. distinguishing between two similar crop varieties) are beyond reach.

Hyperspectral Imaging – The Spectral Continuum

Hyperspectral imaging (HSI) pushes the envelope further by sampling the spectrum with a high density of narrow, contiguous bands – often hundreds of channels covering 400–2500 nm. The sensor architecture is usually a push‑broom or whisk‑broom scanner that records a full spectrum for every pixel (a “data cube” with two spatial and one spectral dimension).

Key physical characteristics:

• Band count: 100–200+ bands depending on the platform.
• Band width: Very narrow, typically 5–10 nm, allowing discrimination of subtle spectral features.
• Spectral resolution: High; can resolve fine absorption features of pigments, minerals or gases.
• Spatial resolution: Often lower than multispectral for a given sensor size due to the need to allocate pixels to many spectral channels (e.g. 1 m for airborne HSI).
• Data volume: Very high – a single 1‑km² scene can generate several gigabytes.
• Calibration: Demands precise radiometric and spectral calibration; often includes in‑flight calibration panels.

Concrete example: The Hyperion sensor aboard the EO‑1 satellite captures 220 spectral bands from 400 to 2500 nm at 30 m spatial resolution. A single pixel contains a 1‑dimensional spectrum that can be used to identify specific mineral assemblages in a rock outcrop or to detect subtle stress in crops caused by nutrient deficiency.

Comparison with multispectral:

“Multispectral is like a colour palette; hyperspectral is a full spectrum of every hue and shade.”

Because HSI records the continuous spectral response, it enables advanced analytical techniques such as spectral unmixing, chemometric modelling and machine‑learning classification. These methods can distinguish materials that appear identical in RGB or multispectral imagery. The trade‑off is the need for specialised hardware, larger data storage, and more complex processing pipelines.

Physical Principles Underpinning the Differences

All three imaging modalities rely on the interaction of electromagnetic radiation with matter. The key variables are:

1. Absorption and scattering: Different materials absorb and scatter light at characteristic wavelengths. For example, chlorophyll absorbs strongly at 430 nm and 675 nm but reflects in the near‑infrared, giving rise to the “red edge” feature that is exploited in NDVI.
2. Spectral reflectance curves: The measured reflectance as a function of wavelength defines a material’s spectral signature. Narrow‑band sensors (HSI) capture the fine structure of these curves; broadband sensors (RGB) only sample the integrated response.
3. Photon flux density: Narrower bands receive fewer photons, reducing signal‑to‑noise ratio. Sensors compensate by increasing exposure time, sensor area or by using higher‑gain electronics.
4. Optical design: Filters (interference, dichroic) and gratings or prisms disperse light into the desired spectral bands. The choice of optical element determines the spectral resolution and throughput.

In practice, the selection of imaging type is a compromise between:

• Spectral discriminability: Ability to differentiate materials.
• Spatial detail: Resolution required for the application.
• Data management: Storage, transmission and processing capacity.
• Cost and operational constraints: Platform availability, sensor expense, and mission duration.

For instance, a precision agriculture application may begin with RGB drones to monitor canopy colour, progress to multispectral imaging for NDVI monitoring, and finally employ HSI to detect early onset of fungal disease by identifying the subtle spectral shifts in the 700–900 nm range.

Choosing the Right Imaging Modality

A decision matrix can guide practitioners:

In summary, RGB imaging offers immediate visual context, multispectral imaging adds a layer of analytical depth suitable for broad land‑cover classification, and hyperspectral imaging delivers the richest spectral information for detailed material discrimination. Understanding these distinctions equips you to select the most appropriate sensor for any remote sensing or imaging task.

Spectral and Spatial Resolution Mechanisms

In remote sensing the two most critical performance metrics are spectral resolution – the ability to distinguish between different wavelengths – and spatial resolution – the size of the ground area represented by a single pixel. Achieving high performance in both dimensions simultaneously is a classic engineering dilemma because the mechanisms that improve one often degrade the other. This section dissects the physical and optical pathways that realise these resolutions, and details the practical trade‑offs that designers face when building RGB, multispectral and hyperspectral sensors.

1. Spectral Resolution: From Filters to Diffraction Gratings

Spectral resolution is defined by the width of the spectral band that a sensor can discriminate. In an RGB camera, three broad bandpass filters (red, green, blue) are laid over the detector array. Each filter passes a wide wavelength range – roughly 50–100 nm – and the sensor produces a single value per colour channel. This coarse sampling is sufficient for colour rendition but insufficient for detecting subtle spectral signatures such as chlorophyll concentration or oil spills.

Moving to multispectral imaging, the detector is still typically a conventional CCD or CMOS array, but the light is split into multiple narrow bands using one of several techniques:

• Colour filter arrays (CFAs) – e.g. the Bayer pattern, extended to more than three colours. Each pixel is covered by a narrow band filter, so spectral resolution is limited by the filter width and the number of distinct colours.
• Filter wheels – a mechanical disc with different interference filters. The sensor captures a scene repeatedly, each time with a different filter. This yields very high spectral purity but sacrifices temporal resolution.
• Band‑pass interference filters – thin‑film structures that transmit a narrow band (~10 nm). These can be stacked or combined with CFAs to produce more channels.

Hyperspectral sensors abandon the detector‑centric approach and employ dispersive optics. Two principal architectures are used:

1. Push‑broom spectrometers – a long slit collects light from a line of the ground; a diffraction grating or prism disperses the light onto a 2‑D detector. Each row of the detector captures a contiguous spectral range, providing hundreds of narrow bands (typically 5–10 nm wide).
2. Whiskbroom (whisk‑broom) scanners – a single pixel or small array is swept across the scene with a rotating mirror, sampling one narrow band at a time. Each scan builds a full spectrum for every pixel.

In both cases, the spectral resolution is governed by the grating density (lines mm^–1) and the optical geometry. A higher line density increases dispersion, allowing narrower bands but also spreading the light over a larger detector area, which reduces photon flux per pixel and increases noise.

2. Spatial Resolution: Pixel Size, Optics and Field‑of‑View

Spatial resolution is the ground sampling distance (GSD) – the size of the area on the ground that projects onto a single pixel. It is a function of three interrelated parameters:

For satellite platforms, the altitude is fixed by the orbit; therefore, improving spatial resolution relies on reducing pixel pitch or increasing focal length. However, these changes are limited by diffraction and the size of the telescope. The diffraction limit, given by d ≈ 1.22 λ / D (where λ is wavelength and D is aperture diameter), sets a floor to the achievable resolution. Larger apertures reduce d but are costly and bulky.

In airborne systems, the altitude can be managed to trade resolution for coverage. For example, a fixed‑wing aircraft flying at 1000 m can achieve a 0.5 m GSD with a 20 cm focal length and 10 µm pixels, whereas a drone at 50 m might use the same optics to reach 0.1 m GSD.

3. The Spectral–Spatial Trade‑off: Engineering Constraints

Increasing spectral resolution typically spreads the incoming light over more detector rows or columns, thereby reducing the photon count per spectral element. This leads to a higher radiometric noise floor. Conversely, concentrating light into fewer spectral bands allows more photons per band, improving signal‑to‑noise ratio. The same principle applies to spatial resolution: smaller pixels capture less light, again degrading SNR and requiring larger optics or more sensitive detectors.

These constraints manifest in several concrete design choices:

• Detector format – A hyperspectral push‑broom sensor might use a 1024 × 1024 pixel array to accommodate 200 spectral bands. If the mission requires finer spatial detail, the array must be larger or the optics magnified, both of which increase cost.
• Optical throughput – High‑resolution filters and gratings can be lossy. For instance, a 1 nm bandwidth interference filter may transmit only 50 % of the incident photons, whereas a 50 nm RGB filter can transmit 80 %. Designers must balance spectral purity against photon budget.
• Data volume – Hyperspectral images can generate several gigabytes per scene. Downlink bandwidth and onboard storage become limiting factors, especially for small satellites. Engineers often employ compression or selective band sampling to mitigate this.
• Radiometric calibration – Narrower bands are more susceptible to calibration drift. A hyperspectral sensor may need a dedicated calibration panel that is refreshed every few hours, whereas an RGB camera can use a static reference.

4. Practical Examples Illustrating the Trade‑off

Example A: The Sentinel‑2 MSI

Sentinel‑2 uses a multispectral instrument with 13 bands ranging from 443 nm to 2190 nm. The visible bands have a 10 m GSD, the near‑infrared (NIR) bands 20 m, and the short‑wave infrared (SWIR) bands 60 m. The choice to group NIR and SWIR bands at coarser GSD reflects the lower photon flux in the infrared and the need to keep the sensor size manageable. The MSI achieves this by using a push‑broom scanner with a 300 mm focal length and a 2048 × 2048 detector, trading some spatial detail for the ability to capture the full spectral range.

Example B: WorldView‑4 RGB Camera

WorldView‑4 employs a panchromatic sensor with a 0.31 m GSD and a multispectral sensor with a 1.24 m GSD. The panchromatic channel uses a single broadband filter (420–900 nm) to maximise photon capture, enabling very fine spatial detail. The multispectral channels have 6 nm narrow bands but at a coarser GSD because the same focal length is shared between the two sensors, and the detector array must accommodate both modalities.

Example C: Hyperion on EO‑1

Hyperion is a hyperspectral push‑broom sensor with 220 bands (10 nm width) covering 400–1000 nm. Its GSD is 30 m, matching the satellite’s 400 m revisit time. The design prioritises spectral fidelity over spatial detail, suitable for large‑scale vegetation monitoring. To keep the data volume within the satellite’s downlink capacity, Hyperion applies onboard compression and selective band transmission.

5. Emerging Solutions to Mitigate the Trade‑off

Recent advances aim to break the classical trade‑off boundary:

1. Compressed sensing – By exploiting the sparsity of many natural scenes in a spectral basis, fewer measurements can reconstruct full spectra, reducing data volume and potentially allowing higher spatial resolution.
2. Photonic integrated circuits (PICs) – On‑chip gratings and waveguides can miniaturise dispersive elements, freeing space for larger detector arrays and enabling more precise spectral tuning without bulky optics.
3. Hybrid detectors – Combining visible and infrared sensors on a single chip allows simultaneous capture of wide spectral ranges with reduced optical path complexity.
4. Adaptive optics – Real‑time wavefront correction can sharpen images, effectively improving spatial resolution without increasing detector pixel density.

These innovations promise that future missions could achieve both high spectral and spatial resolution, but each introduces its own engineering challenges, such as increased power consumption or complex calibration procedures.

6. Conclusion

Understanding how spectral and spatial resolution are realised and the interdependencies between them is essential for designing effective imaging systems. RGB devices maximise spatial detail at the cost of spectral breadth, multispectral sensors strike a balance by adding a modest number of narrow bands, and hyperspectral systems prioritise spectral richness, often sacrificing spatial granularity. The engineering trade‑offs revolve around detector size, optical throughput, data handling, and mission constraints such as altitude and revisit time. By analysing concrete examples from operational satellites and exploring emerging technologies, we gain insight into the practical limits and future possibilities of imaging science.

Comparative Summary Table

Comparative Analysis and Decision Criteria

When deciding between RGB, multispectral and hyperspectral imaging, the key considerations are the spectral resolution required, the spatial resolution that the application demands, the data volume that can be handled, and the cost and complexity of the system. The table below summarises the core distinctions and then the following sections analyse how these differences translate into real‑world decision points.

Further Reading

• Jensen, J.R. (2007) — Remote Sensing of the Environment: An Introduction. Provides foundational concepts of spectral imaging and distinguishes RGB, multispectral, and hyperspectral systems.
• Lu, Y., & Weng, Q. (2015) — A survey of image classification methods and applications. Discusses spectral signatures and the advantages of multispectral over RGB data.
• International Society for Photogrammetry and Remote Sensing (ISPRS) (2019) — ISPRS Guidelines for Hyperspectral Imaging. Outlines standards for hyperspectral data acquisition and processing.
• Wang, H. (2013) — Hyperspectral Imaging: Techniques for Spectral Analysis in Medicine. Explains the technical differences between RGB, multispectral, and hyperspectral imaging in a clinical context.
• Wang, G., & Liu, L. (2020) — Fundamentals of Remote Sensing Technology. Covers the physical principles behind spectral imaging and the practical applications of each imaging type.
• European Space Agency (ESA) (2018) — Sentinel-2 User Handbook. Provides real-world examples of multispectral imaging missions and compares their spectral characteristics to RGB and hyperspectral data.

Simple Python example: simulate RGB, multispectral and hyperspectral spectra

import numpy as np
import matplotlib.pyplot as plt

Define wavelengths (400-1000 nm) and a simple vegetation spectrum
wavelengths = np.arange(400, 1001, 1)  # nm
veg = 0.2 + 0.4 * np.exp(-(wavelengths-680)**2 / (2*30**2))  # red edge peak

RGB bands (centres and widths)
rgb_centres = np.array([460, 540, 640])  # nm
rgb_widths  = np.array([100, 100, 100])  # nm

Multispectral bands (additional NIR)
ms_centres = np.array([450, 550, 650, 850])  # nm
ms_widths  = np.array([30, 30, 30, 30])      # nm

Hyperspectral: every 5 nm
hs_centres = np.arange(400, 1001, 5)
hs_widths  = np.full_like(hs_centres, 5)

def band_response(wl, centre, width):
    ""Simple top‑hat band response.""
    return np.where((wl >= centre - width/2) & (wl <= centre + width/2), 1.0, 0.0)

Compute band‑averaged reflectance
def average_reflectance(spectrum, centres, widths):
    return np.array([np.sum(spectrum * band_response(wavelengths, c, w)) /
                     np.sum(band_response(wavelengths, c, w))
                     for c, w in zip(centres, widths)])

rgb_ref = average_reflectance(veg, rgb_centres, rgb_widths)
ms_ref  = average_reflectance(veg, ms_centres, ms_widths)
hs_ref  = average_reflectance(veg, hs_centres, hs_widths)

Plot
plt.figure(figsize=(10,4))
plt.plot(wavelengths, veg, label='Vegetation spectrum')
plt.scatter(ms_centres, ms_ref, colour='C1', label='Multispectral