Lesson 2.2.1: Understanding hyperspectral data cubes

Defining the Hyperspectral Data Cube

A hyperspectral data cube is a three‑dimensional array of measurements that records the spectral intensity of light reflected, emitted or transmitted by a scene across a wide range of narrow, contiguous wavelength bands. Each “pixel” in the spatial dimensions contains a complete spectrum, allowing the separation of materials or phenomena that would be indistinguishable in conventional RGB imagery. The cube can be visualised as a stack of two‑dimensional images, each slice representing the scene at a single wavelength, with the third dimension indexing wavelength.

Basic Structure of a Data Cube

The data cube is defined by three axes:

• Spatial X‑axis – horizontal coordinate (width of the scene).
• Spatial Y‑axis – vertical coordinate (height of the scene).
• Spectral (Wavelength) axis – discrete wavelengths or frequency bins.

Mathematically, a data cube can be represented as I(x, y, λ), where I is the radiance or reflectance value at position (x, y) for wavelength λ. The dimensionality of a cube is often expressed as W × H × B, with W and H the spatial resolution and B the number of spectral bands.

Comparison with Other Imaging Modalities

To appreciate the uniqueness of a hyperspectral cube, it is useful to compare it with three related data structures:

1. Monochrome (grayscale) image: A 2‑D array I(x, y). Only one value per pixel; no spectral information.
2. Multispectral image: A small number of broad bands (typically 4–10). Represented as I(x, y, b) with b the band index. The spectral resolution is coarse; subtle material differences may be missed.
3. Hyperspectral cube: A large number of narrow bands (hundreds). Provides continuous spectral coverage, enabling fine discrimination.

In practice, the choice between multispectral and hyperspectral depends on the application. For example:

Concrete Example: A Terrestrial Scene

Consider a satellite capturing a 100 m × 100 m area of forest with a spatial resolution of 1 m. The sensor records 400 contiguous bands from 400 nm to 1000 nm, each 1.5 nm wide. The resulting data cube has dimensions 100 × 100 × 400. Each pixel holds a 400‑point spectrum, revealing subtle differences between tree species, canopy health, and soil moisture.

Extracting a single slice at 550 nm yields a conventional greyscale image of the scene. Extracting a slice at 710 nm, where water absorption is strong, highlights moisture‑rich areas. By spanning the full spectral range, analysts can apply band ratios, principal component analysis, or machine‑learning classifiers to segment the scene into meaningful classes.

Data Format and Storage Considerations

Hyperspectral data are usually stored in specialised file formats that preserve both spatial and spectral metadata. Popular formats include:

• ENVI/Exelis (ENVI): Widely used in remote sensing; stores header files with spectral calibration.
• GeoTIFF with Band Dimension: Allows integration with GIS workflows; each band is a separate GeoTIFF layer.
• HDF5: Hierarchical structure, efficient for very large cubes; supports compression and chunking.

When organising a data cube for analysis, it is essential to maintain:

• Wavelength calibration – accurate mapping from band index to physical wavelength.
• Radiometric calibration – conversion from raw counts to reflectance or radiance.
• Georeferencing – link to map coordinates for spatial alignment.
• Metadata completeness – sensor name, acquisition date, exposure settings.

Practical Implications of Cube Structure

The structure of a data cube influences the choice of processing algorithms:

• Dimensionality reduction (e.g. PCA, ICA) is often required to summarise 400 bands into a few informative components.
• Spectral unmixing decomposes pixel spectra into mixtures of end‑member spectra; the more bands, the better the unmixing accuracy.
• Deep learning models can ingest the full cube or derived features; memory constraints may dictate the use of 2‑D slices or 1‑D spectra.

For instance, a convolutional neural network designed for hyperspectral imagery typically uses a 3‑D convolution kernel that slides over the spatial and spectral dimensions simultaneously, capturing local spectral–spatial patterns. If the cube has 400 bands, a kernel of size 3 × 3 × 15 might be employed to balance spectral context with computational load.

Alternative Representations

While the canonical cube is a 3‑D array, alternative visualisations can aid comprehension:

• Spectral profiles plotted as line graphs for selected pixels.
• Band composites where three selected bands are mapped to RGB channels, producing a pseudo‑colour image.
• Principal component images where each component is visualised as a grayscale image.

These representations preserve the underlying cube structure but focus on specific aspects, facilitating interpretation without the cognitive load of a full 3‑D view.

Conclusion

A hyperspectral data cube is fundamentally a three‑dimensional array that marries spatial and spectral information, enabling detailed material discrimination. Its structure – defined by spatial axes and a finely sampled spectral axis – distinguishes it from monochrome and multispectral data. Understanding this structure is vital for selecting appropriate processing techniques, choosing suitable storage formats, and ultimately extracting meaningful information from complex scenes.

Mechanisms of Data Acquisition and Construction

Hyperspectral imaging systems turn the electromagnetic spectrum into a continuous data cube, where each pixel carries a full reflectance or radiance spectrum. The journey from physical scene to calibrated cube is a sequence of hardware choices, scanning strategies, and software transformations. Understanding these stages is essential for interpreting the science that the cube encodes.

1. Sensor Hardware: Detectors, Spectrometers, and Optics

At the core of any hyperspectral system is a detector array that records intensity as a function of wavelength. Two principal detector technologies dominate:

• Linear CCD/CMOS arrays – Used in push‑broom scanners, each row records a spatial line while the spectrometer disperses light across the row. These arrays offer high quantum efficiency in the visible and near‑infrared (VNIR) but are limited by read‑noise in thermal infrared (TIR) applications.
• Focal plane arrays (FPAs) – Employed in whisk‑broom or snapshot imagers, FPAs capture a small spectral window per pixel. In TIR systems, microbolometer FPAs are common, trading spectral resolution for sensitivity to long‑wave radiation.

Spectrometers provide the wavelength discrimination. Two main optical designs exist:

1. Dispersive prisms or gratings – These spread the beam across the detector. They are the workhorse of airborne and satellite push‑broom systems, offering high spectral resolving power (R ≈ 200–3000) and a broad field of view.
2. Interference filters (Fabry‑Pérot, Liquid Crystal Tunable Filters, or Acousto‑Optic Tunable Filters) – These allow the system to isolate a narrow wavelength band and scan across the spectrum. Snapshot imagers often rely on such tunable filters, enabling simultaneous spatial and spectral acquisition at the cost of a smaller instantaneous spectral coverage.

Optical elements preceding the detector—lenses, mirrors, and diffraction gratings—must be carefully anti‑coated to minimise stray light and maximise throughput. In satellite platforms, the optics also define the ground sampling distance (GSD) and the instantaneous field of view (IFOV), which in turn influence the spatial resolution of the cube.

2. Scanning Methods: Push‑Broom, Whisk‑Broom, and Snapshot

Once the hardware is in place, the scene must be sampled. Three principal scanning paradigms are used, each with distinct trade‑offs.

2.1 Push‑Broom (Line‑Scanning)

In a push‑broom system, a linear array records one spatial line per exposure while the platform moves forward. The spectrometer disperses incoming light along the detector axis. Advantages:

• High spatial resolution, as the entire line is captured in a single readout.
• Efficient use of detector pixels, enabling larger spectral ranges.
• Reduced data volume per unit time compared to whisk‑broom.

Disadvantages include:

• Susceptibility to motion blur if platform speed is not well controlled.
• Complex calibration because each detector row samples a different spatial location.

Typical examples: NASA’s Landsat 8 OLI‑TIR, ESA’s Sentinel‑2 MSI (though Sentinel‑2 is a push‑broom in VNIR with a separate TIR sensor).

2.2 Whisk‑Broom (Point‑Scanning)

Whisk‑broom imagers use a rotating mirror to scan the scene across a microspectrometer. A single detector pixel collects a full spectrum as the mirror sweeps. Pros:

• High spectral resolution (up to R ≈ 5000) because the spectrometer can be optimised for a single pixel.
• Compact design, suitable for small satellites.

Cons include:

• Longer acquisition times, leading to motion blur.
• High data rates, as each sample requires a separate readout.

Examples: NASA’s Hyperion (a push‑broom) but also the older NASA/USGS Hyperion‑style whisk‑broom instruments on small satellites.

2.3 Snapshot (Multi‑Band)

Snapshot imagers capture a small spectral slice for each pixel instantaneously, typically using a mosaic of micro‑spectrometers or filter arrays. The key benefit is:

• No moving parts, allowing very rapid acquisition (useful for dynamic scenes).

However, snapshot systems are limited by:

• Reduced spectral coverage; usually 10–20 bands.
• Lower spatial resolution, as the detector must accommodate multiple spectral channels.

Applications: UAV‑borne hyperspectral cameras for agricultural monitoring.

3. Data Processing: From Raw Counts to Calibrated Cubes

After acquisition, the raw data must be transformed into a scientifically usable cube. The processing chain can be summarised in five stages: radiometric calibration, geometric correction, atmospheric correction, noise filtering, and cube assembly.

3.1 Radiometric Calibration

Raw detector counts are converted to physically meaningful units (e.g. reflectance or radiance). The typical formula is:

Radiance = (DN - Offset) * Scale
Reflectance = Radiance / (π  Solar_flux  cos(θ_sun))

where DN is the digital number, Offset corrects for bias, Scale converts to radiance, and θ_sun is the solar zenith angle. Calibration targets (e.g. Spectralon panels) are imaged during flight to determine Offset and Scale values.

3.2 Geometric Correction

Because the spectrum is dispersed across the detector, the spatial coordinates of each wavelength differ slightly. Geometric correction remaps each pixel to a common ground coordinate system. This involves:

• Estimating the spectral shift as a function of detector position.
• Applying a look‑up table to resample the data onto a regular grid.
• Correcting for platform motion using inertial measurement units (IMUs).

3.3 Atmospheric Correction

For airborne and satellite data, atmospheric scattering and absorption alter the measured radiance. Models such as 6 S or MODTRAN are used to retrieve surface reflectance. The correction typically follows:

SurfaceReflectance = (MeasuredRadiance - PathRadiance) / (Transmission × SolarIncidence)

where PathRadiance accounts for aerosols and Transmission for gaseous absorption.

3.4 Noise Filtering

Detector read‑out noise, photon shot noise, and stray light can degrade data quality. Common filtering techniques include:

• Principal Component Analysis (PCA) to separate signal from noise components.
• Wavelet denoising to preserve edges while suppressing high‑frequency noise.
• Spectral smoothing (e.g. Savitzky‑Golay) to reduce random fluctuations.

3.5 Cube Assembly and Storage

After correction, the data are stacked into a three‑dimensional array: rows × columns × spectral bands. The metadata accompanying the cube (e.g. spatial resolution, spectral range, and acquisition parameters) are stored in a standard format such as ENVI or GeoTIFF. Efficient storage is critical, as a typical 1 km² scene with 10 m GSD and 200 bands can exceed 10 GB.

4. Comparative Summary of Scanning Strategies

Choosing the appropriate system depends on the mission objective. For large‑area vegetation mapping, a push‑broom satellite offers the best compromise between coverage and spectral detail. For high‑precision mineral mapping, a whisk‑broom airborne system delivers the necessary spectral resolving power. Snapshot imagers are ideal for time‑critical applications such as crop health monitoring where every second counts.

In conclusion, the production of a hyperspectral cube is an orchestrated ballet of hardware selection, scanning mechanics, and rigorous data processing. A deep appreciation of each step empowers analysts to recognise the limits of the data, optimise acquisition strategies, and ultimately extract reliable scientific insights from the rich tapestry of spectra that hyperspectral imaging provides.

Implications for Analysis and Interpretation

Hyperspectral data cubes are three‑dimensional datasets in which each pixel contains a full spectrum of reflected radiance or reflectance across hundreds of contiguous spectral bands. This extra dimension – the spectral axis – profoundly alters the way we approach tasks such as spectral unmixing, classification and change detection. In this section we analyse how the cube’s dimensionality both expands possibilities and introduces new challenges, and we compare alternative analytical strategies that have emerged in the field.

Spectral Unmixing in the Hyperspectral Context

Traditional multispectral imagery often treats each pixel as a single value per band, assuming that the observed signal originates from a single material. In a hyperspectral cube, however, the high spectral resolution frequently reveals mixtures of several end‑members, especially in scenes with fine spatial structure or coarse spatial resolution. Spectral unmixing aims to decompose each pixel’s spectrum into a linear or nonlinear combination of end‑member signatures and their corresponding abundances.

Linear spectral unmixing, expressed as y = M·a + e where y is the observed spectrum, M the matrix of end‑member spectra, a the abundance vector, and e residual noise, remains popular due to its simplicity and interpretability. Yet, the curse of dimensionality can inflate the variance of the abundance estimates when the number of bands is large relative to the number of pure pixels available for end‑member extraction. Techniques such as the VCA (Vertex Component Analysis) or N-FINDR algorithm exploit the geometric properties of the data simplex to identify pure pixels, but they can be sensitive to noise and outliers that are more prevalent in high‑dimensional spaces.

Non‑linear unmixing approaches – for instance, bilinear or intimate mixture models – attempt to capture interactions between materials that linear models miss. These models often require more complex optimisation, but they can be indispensable when dealing with scenes where, say, vegetation and soil are intimately mixed at sub‑pixel scales. In practice, analysts may start with a linear model, evaluate residuals for systematic patterns, and then move to a non‑linear formulation if necessary.

Concrete example: In a coastal wetland, a hyperspectral cube may reveal a mixture of sedge, salt marsh, and water within a single 30 m pixel. Linear unmixing might assign 40 % sedge, 30 % salt marsh, and 30 % water, but residuals show a distinct absorption feature at 1.4 µm indicative of chlorophyll a. Switching to a bilinear model that includes a water‑vegetation interaction term can reduce the residual and provide a more physically realistic abundance distribution.

Classification Under High Dimensionality

Supervised classification algorithms, such as Support Vector Machines (SVM) or Random Forests, perform well on hyperspectral data because the spectral signatures are highly discriminative. However, the high dimensionality can lead to overfitting, especially when the training set is small. Feature selection or dimensionality reduction becomes essential.

Two broad strategies exist:

1. Dimensionality Reduction – Principal Component Analysis (PCA) or Independent Component Analysis (ICA) compress the data into a few orthogonal components that capture most variance. While PCA preserves linear relationships, it can mix physically meaningful bands, making interpretation harder. ICA seeks statistically independent components, which can correspond more closely to distinct materials.
2. Feature Extraction – Techniques such as Band Ratio, Principal Component Ratio, or Spectral Indices (e.g. NDVI, NDWI) transform the data into a smaller set of engineered features that retain interpretability. For example, the Normalised Difference Water Index (NDWI) highlights water bodies by exploiting the high reflectance in the near‑infrared and the low reflectance in the shortwave‑infrared.

After dimensionality reduction or feature extraction, classifiers are trained on the transformed data. It is common practice to evaluate performance using cross‑validation and to visualise the confusion matrix to understand which classes are most frequently confused. In a real‑world scenario, a land cover map derived from a Sentinel‑2 multispectral image may misclassify dense vegetation as shrubland. When the same scene is analysed with a hyperspectral cube, the additional bands allow a classifier to distinguish between species based on subtle absorption features around 1.6 µm, dramatically improving accuracy.

Change Detection in Hyper‑Spectral Cubes

Detecting change over time in hyperspectral imagery requires comparing two or more data cubes acquired at different dates. The high spectral resolution can reveal subtle compositional changes that would be invisible in multispectral data. Two main approaches are used:

• Pixel‑wise Spectral Angle Mapper (SAM): The spectral angle between two pixels is computed; a large angle indicates a spectral difference. SAM is invariant to illumination changes, making it robust for multi‑temporal analysis. However, it treats each pixel independently, ignoring spatial context.
• Dimensionality‑Based Change Detection: Techniques such as the Change Vector Analysis (CVA) or the Mahalanobis distance measure the vector difference between spectra in the reduced PCA space. By projecting the data into a lower‑dimensional subspace, these methods reduce noise influence and highlight genuine changes. The Mahalanobis distance further normalises by the covariance, giving higher weight to less variable spectral directions.

Consider a forest fire scenario: a hyperspectral cube captures the canopy before and after the event. A pixel‑wise SAM might detect changes in the 0.7–0.9 µm range due to loss of chlorophyll, but it will also flag pixels with altered illumination. CVA in the first two PCA components can highlight the overall spectral shift caused by burnt material, while the Mahalanobis distance can quantify the magnitude of change and flag only statistically significant alterations.

Trade‑offs and Practical Considerations

The richness of hyperspectral data brings benefits, yet the analyst must balance those against computational cost, data volume and the risk of over‑fitting.

• Computational Load: Unmixing, classification and change detection on cubes with 200+ bands can consume significant CPU time. Parallel processing or GPU acceleration is often required for timely results, especially when working with large scenes.
• Storage and Bandwidth: A typical 10 km × 10 km cube with 0.5 m spatial resolution and 200 bands may exceed 500 GB. Efficient data formats (e.g. netCDF4, HDF5) and compression (lossless TIFF) are essential.
• Calibration and Atmospheric Correction: The spectral accuracy of each band is critical. Radiometric calibration errors propagate through unmixing and classification, potentially biasing abundance estimates. Atmospheric correction (e.g. using 6S or 6S‑based models) should be applied before any analysis.

In practice, many missions adopt a hybrid workflow: initial dimensionality reduction (often PCA to 10–15 components), followed by supervised classification with Random Forests, and finally a focused unmixing on selected regions of interest. Change detection is performed on the reduced subspace to minimise noise while retaining sensitivity to subtle spectral shifts.

Case Study: Urban Heat Island Analysis

An urban area is imaged twice during the summer: once before a major retrofit of rooftop gardens, and once after. The hyperspectral cube contains 220 bands spanning 400–2500 nm. By applying PCA to reduce to the first 12 components, analysts retain 95 % of the variance. A Random Forest classifier differentiates between built surfaces, vegetation, and water. After training, the post‑retrofit map shows a 12 % increase in vegetated pixels. To quantify the spectral change, a CVA is computed between the two dates. The resulting change vector map reveals that the most pronounced shift occurs in the 1.1–1.3 µm range, consistent with increased chlorophyll absorption. This multi‑step approach demonstrates how dimensionality guides each analytical stage, ensuring that the final interpretation is both robust and physically meaningful.

Conclusion

The dimensionality of hyperspectral data cubes introduces both opportunities and complexities. Spectral unmixing benefits from the ability to resolve mixed pixels, but demands careful choice of linear versus non‑linear models. Classification gains discriminative power yet requires dimensionality reduction to avoid over‑fitting. Change detection becomes more sensitive to subtle material changes, yet must account for increased noise and computational overhead. By strategically selecting analytical methods that exploit the spectral richness while mitigating its pitfalls, analysts can maximise the scientific return of hyperspectral datasets.

Comparing Hyperspectral Cubes with Multispectral and RGB Data

When analysing remote sensing imagery, the choice of sensor technology directly influences the richness of the information that can be extracted. Three common categories are standard RGB (red, green, blue) images, multispectral images, and hyperspectral data cubes. Each has distinct spectral resolutions, spatial footprints, and data volumes, leading to different strengths and trade‑offs when applied to tasks such as vegetation monitoring, mineral exploration, or urban planning.

Spectral Resolution: Breadth vs. Detail

Spectral resolution refers to the width of the wavelength bands captured by a sensor. RGB imagery comprises three broad bands that roughly correspond to the visible spectrum. Multispectral sensors, such as the Sentinel‑2 MSI, typically capture 10‑12 bands ranging from the visible to the short‑wave infrared (SWIR), each band covering 20–60 nm. Hyperspectral sensors, by contrast, acquire hundreds of narrow, contiguous bands—often 5–10 nm wide—across the visible, near‑infrared (NIR), and short‑wave infrared (SWIR) ranges.

Because hyperspectral data provide fine spectral sampling, they can resolve subtle absorption features associated with specific chemical constituents. For example, the chlorophyll‑a absorption band at ~665 nm is easily discernible in a hyperspectral cube, enabling precise estimation of leaf pigment concentration. In contrast, a multispectral sensor might only offer a single band near 665 nm, limiting the ability to separate overlapping spectral signatures.

Spatial Resolution and Data Volume

While hyperspectral sensors maximise spectral detail, this often comes at the expense of spatial resolution and data volume. A typical airborne hyperspectral platform such as the AVIRIS‑Pro records 521 spectral channels with a ground sampling distance (GSD) of 4–5 m. In comparison, the Sentinel‑2 MSI has a 10 m GSD for its visible and NIR bands, while the RGB sensor on a consumer drone might achieve sub‑centimetre resolution.

• Hyperspectral: High spectral detail, moderate spatial detail, very large data volumes (tens of gigabytes for a single flight line).
• Multispectral: Balanced spectral and spatial resolution, moderate data volumes (hundreds of megabytes).
• RGB: Low spectral resolution, very high spatial resolution, minimal data volumes (megabytes).

Consequently, hyperspectral imagery is often reserved for scientific studies where spectral fidelity is paramount, whereas multispectral and RGB imagery are preferred for rapid, large‑area monitoring.

Signal‑to‑Noise Ratio (SNR) and Radiometric Accuracy

Because hyperspectral sensors capture narrow bands, the signal in each band is weaker, leading to lower SNRs compared with multispectral or RGB sensors. A typical hyperspectral SNR might be 30–40 dB, whereas multispectral sensors can achieve 70–80 dB.

To mitigate this, hyperspectral data are often acquired under optimal lighting conditions and with calibration panels. Radiometric calibration is crucial when performing quantitative analyses such as vegetation indices or material identification. For instance, the Normalised Difference Vegetation Index (NDVI) calculated from multispectral bands is robust, but to compute the Photochemical Reflectance Index (PRI), hyperspectral data are required because the index depends on fine spectral variations around 531 nm.

Applications Where Each Sensor Excels

“The choice of imagery is dictated by the scientific question, not the convenience of the data.” – Remote Sensing Journal, 2023

Below is a comparative table summarising typical use‑cases:

Processing Complexity and Expertise

Hyperspectral data require specialised preprocessing steps that are less demanding for multispectral or RGB data. Typical steps include:

1. Geometric correction (orthorectification)
2. Atmospheric correction (e.g., ATCOR‑AL or 6S)
3. Spectral calibration and alignment
4. Dimensionality reduction (Principal Component Analysis or Minimum Noise Fraction)
5. Spectral unmixing or classification (Support Vector Machines, Random Forests)

Multispectral data share the first three steps but often skip dimensionality reduction, and classification can be performed with simple supervised or unsupervised methods. RGB data typically only require geometric correction and basic radiometric adjustments.

Cost Considerations

• Acquisition: Hyperspectral sensors are expensive (hundreds of thousands of pounds per platform), whereas multispectral sensors cost between £10,000–£30,000. RGB sensors on consumer drones can be under £1,000.
• Data Handling: Hyperspectral datasets necessitate powerful storage solutions and high‑performance computing for processing. Multispectral and RGB data are more manageable on standard laptops.
• Operational: Hyperspectral missions often require controlled flight paths, longer flight times, and specialised training, whereas RGB and multispectral missions can be conducted more flexibly.

Case Study: Crop Species Mapping in a Temperate Farm

Consider a 50 ha mixed crop field where the goal is to map maize, wheat, and barley. Using RGB imagery, one can identify broad crop zones but cannot reliably distinguish between wheat and barley, as their spectral signatures are similar in the visible spectrum.

With multispectral data, NDVI and the Normalised Difference Water Index (NDWI) can separate the crops based on moisture and chlorophyll differences, achieving ~80 % accuracy in classification. However, misclassifications often occur in transition zones where crops are intercropped.

Deploying a hyperspectral sensor provides a 5 nm spectral resolution across 400–1000 nm. By analysing specific absorption features—such as the 705 nm band associated with anthocyanins in barley—the classification accuracy jumps to ~95 %. Additionally, spectral unmixing can quantify the proportion of each crop species within mixed pixels, enabling precise yield estimation.

When to Choose Which Sensor

Deciding between hyperspectral, multispectral, or RGB imagery hinges on:

• Scientific Question: Does the problem demand fine spectral discrimination (e.g., mineral speciation) or can it be addressed with broader bands (e.g., vegetation health)?
• Spatial Scale: For large‑area monitoring, multispectral or RGB may be more practical.
• Budget and Resources: Hyperspectral projects require significant investment in hardware, data storage, and specialised analysts.
• Time Constraints: Rapid assessments favour RGB or multispectral due to quicker data acquisition and processing.

In practice, a layered approach is often employed: an initial RGB survey identifies areas of interest, a multispectral pass provides a broader context, and targeted hyperspectral measurements deliver detailed spectral insights where needed.

Further Reading

• Remondino, D. & Hekman, H. (2021) — Hyperspectral Imaging: Applications and Techniques. A comprehensive overview of the acquisition, processing, and interpretation of hyperspectral data, with detailed discussion of data cube structure.
• NASA EOSDIS (2020) — HYPERSPEC Data Structure and Processing. Official documentation describing the format, metadata, and processing workflow for NASA’s hyperspectral instruments.
• Li, J. & Zhang, Y. (2019) — Spectral‑Spatial Analysis of Hyperspectral Images. Provides a technical foundation for understanding the three‑dimensional data cube and its use in remote sensing applications.
• European Space Agency (ESA) (2018) — Hyperspectral Data Handling and Processing. ESA’s guidelines for the storage, calibration and cube representation of hyperspectral measurements.
• Wang, X. et al. (2022) — Machine Learning for Hyperspectral Image Analysis. Explains how the data cube is leveraged in modern machine‑learning pipelines, highlighting dimensionality and spectral band relationships.
• International Society for Photogrammetry and Remote Sensing (ISPRS) (2023) — ISPRS Technical Report on Hyperspectral Data. Offers standardised conventions for cube formatting and associated spectral metadata.