Lesson 2.2.3: Reading and interpreting hyperspectral image data

Section 1: Foundations of Hyperspectral Data Structure

Hyperspectral imaging (HSI) captures light reflected or emitted from a scene across a large number of contiguous spectral bands, typically ranging from the visible through the short‑wave infrared. Unlike conventional RGB cameras that record only three broad bands, or multispectral sensors that might collect five to ten discrete bands, a hyperspectral sensor records hundreds of narrow, overlapping bands. This dense spectral sampling turns each pixel into a small spectrum, allowing the material composition of that pixel to be inferred with high precision.

To understand how this information is organised, it is helpful to think of an HSI scene as a three‑dimensional array, or data cube. Each dimension corresponds to a different attribute: the two spatial axes (often denoted X and Y) and the spectral axis (λ). The data cube therefore contains, for every (x, y) coordinate, a full spectrum spanning the sensor’s wavelength range. In mathematical terms, if a sensor records N spectral bands and covers a scene of M × P pixels, the data cube can be represented as a matrix of size M × P × N.

Spatial dimensions: X and Y

The X and Y axes are analogous to the horizontal and vertical axes on a conventional photograph. They describe the location of each pixel in the scene. The resolution of these axes is typically governed by the sensor’s optics and the platform’s altitude. For example, a satellite sensor operating at 30 m per pixel will produce a larger, coarser spatial grid than a UAV‑borne sensor at 0.5 m per pixel. Spatial resolution is often expressed as the ground sample distance (GSD). It determines the smallest feature that can be distinguished in the image.

Spectral dimension: λ

The spectral axis is unique to hyperspectral data. Each value along this axis corresponds to the radiance or reflectance measured at a specific wavelength. Because each band is narrow (typically 5–20 nm in width), the spectrum captured for a single pixel contains fine detail about the material’s absorption features. For instance, chlorophyll exhibits a pronounced absorption band near 680 nm, while water shows a distinct dip around 970 nm. By examining the shape of the spectrum, one can discriminate between vegetation, water, minerals, and even subtle variations within a single class.

Dimensionality: Why hyperspectral data is “more than an image”

When we consider dimensionality, we are comparing the number of independent variables that describe a data point. A conventional RGB image has three dimensions per pixel (red, green, blue). A multispectral image might have 5–10 dimensions. A hyperspectral image, by contrast, can have 200–300 dimensions or more. This high dimensionality is both a strength and a challenge:

• Strength: The richer spectral signature allows for more accurate classification and material identification.
• Challenge: The curse of dimensionality can lead to overfitting in machine learning models and increases computational load for storage and processing.

To illustrate, consider a simple example: a scene containing a patch of healthy grass, a patch of dry grass, and a patch of a water body. In an RGB image, these may appear as subtly different shades of green and blue, potentially confounding a naïve classification algorithm. In a hyperspectral image, each material has a distinct spectral fingerprint. The healthy grass shows a strong near‑infrared reflectance peak, the dry grass has a lower near‑infrared response, and water has a pronounced absorption around 970 nm. By examining the full spectrum, a classification algorithm can separate these classes with far greater reliability.

Representing the data cube: From raw numbers to visualisation

In practice, the data cube is stored as a multi‑band raster file, often in formats such as ENVI, GeoTIFF, or HDF5. Each band is a 2‑D array of pixel values, and the collection of bands forms the third dimension. Visualising the data cube can be done in several ways:

1. Band selection: Selecting two or three bands and mapping them to the RGB colour space. This is useful for creating pseudo‑colour images that highlight specific spectral features.
2. Principal Component Analysis (PCA): Reducing dimensionality while preserving as much variance as possible. The first few principal components often capture the most significant spectral variations across the scene.
3. Band ratioing: Calculating the ratio between two bands to normalise for illumination effects.

Each visualisation technique offers a different perspective on the data, helping analysts identify patterns that may not be apparent in raw spectral form.

Comparing hyperspectral to multispectral and conventional imaging

These differences are not merely academic; they influence every step of the workflow, from acquisition to analysis. For example, a satellite mission such as the European Space Agency’s Sentinel‑2 provides 13 multispectral bands and is designed for broad‑scale land monitoring. In contrast, airborne missions like NASA’s Hyperion or the European Space Agency’s PRISMA capture full spectra at sub‑kilometre resolution, enabling detailed mineral mapping or crop stress assessment.

Concrete example: spectral signatures from a field experiment

In a recent field trial, a UAV equipped with a commercial hyperspectral sensor flew over a mixed crop field. The sensor recorded 240 bands from 400 nm to 1000 nm at 5 nm spacing. The raw data were processed into a data cube with a spatial resolution of 0.3 m. Analysts extracted spectra from three representative pixels:

• Healthy wheat: strong near‑infrared peak (~800 nm), moderate chlorophyll absorption near 670 nm.
• Stressed wheat: lower near‑infrared peak, increased absorption in the visible due to chlorophyll degradation.
• Soil: flat, low reflectance across the spectrum, with a slight rise in the short‑wave infrared.

By feeding these spectra into a supervised classification algorithm (e.g. Support Vector Machine), the analysts achieved >95 % accuracy in distinguishing the three classes, a performance far superior to that obtained with the same data processed as a simple RGB image.

Key takeaways for the student

1. A hyperspectral data cube is a 3‑D array: X (horizontal), Y (vertical), and λ (spectral).

2. The spectral dimension contains hundreds of narrow, contiguous bands, enabling detailed material discrimination.

3. Dimensionality is the core differentiator between conventional, multispectral, and hyperspectral imagery.

4. Visualisation and dimensionality reduction techniques help manage the data volume and extract useful information.

5. Practical examples demonstrate the superior classification performance achievable with hyperspectral data.

Section 2: Practical Steps for Reading Hyperspectral Data

Hyperspectral images are not simply large stacks of RGB pixels; they are multi‑dimensional data cubes that encode spectral signatures across hundreds of contiguous bands. Importing these data into analysis platforms such as ENVI, QGIS, or Python requires a systematic approach that honours the file format, metadata, and the physical calibration that underpins any subsequent interpretation. This section walks through the typical workflow, compares common formats, and offers concrete examples of how to extract and utilise metadata during calibration.

2.1 Understanding the Common Data Formats

Before you can read a hyperspectral cube you must recognise the format. The most frequently encountered are:

1. ENVI (.hdr + .img) – a proprietary yet widely supported format that stores spectral data in a binary file with a header describing the cube.
2. HDF5 (.h5) – a hierarchical data format that can encapsulate multiple datasets and metadata groups, often used by NASA missions.
3. GeoTIFF (.tif) – a georeferenced raster format that can hold a band‑stack, though it is less common for hyperspectral data due to size constraints.
4. NetCDF (.nc) – similar to HDF5 but with a different organisational scheme, frequently used in climate science.

Each format has its own conventions for storing spectral calibration, spatial reference, and ancillary information such as illumination geometry. Familiarity with these conventions is essential to avoid mis‑interpretation.

2.2 Preparing the Workspace

Regardless of the format, the following preparatory steps apply:

• Confirm the software version (e.g. ENVI 4.6, QGIS 3.30, Python 3.11).
• Verify that the required drivers or plugins are installed (e.g. GDAL for GeoTIFF, h5py for HDF5).
• Check the file integrity – a quick checksum or a view of the header can reveal corruption.
• Store the data in a dedicated project folder to maintain organisational clarity.

2.3 Importing ENVI Data

ENVI is the most common platform for hyperspectral imagery. Importing an ENVI cube involves two files: the binary image (.img) and the header (.hdr). The header contains critical information such as the number of rows, columns, bands, and the byte order.

import envi
cube = envi.open('sample.hdr')
print(cube.shape)  # (rows, cols, bands)
print(cube.bands.centres)  # Wavelengths in nm

In ENVI’s graphical interface, the steps are:

1. Open the “Open” dialogue and select the .hdr file.
2. ENVI automatically reads the .img file and displays a cube preview.
3. Use the Band Selection tool to isolate spectral ranges of interest.

ENVI stores metadata in the header in a key‑value format. For calibration, look for entries such as radiance_mult and radiance_add. If missing, consult the mission’s calibration guide or contact the data provider.

2.4 Importing HDF5 Data

HDF5 files organise data in a hierarchical structure. A typical hyperspectral HDF5 might contain a dataset named Data and a group Metadata.

import h5py
with h5py.File('sample.h5', 'r') as f:
    data = f['/Data'][:]  # shape: (bands, rows, cols)
    meta = dict(f['/Metadata'].attrs)
print(meta['Wavelengths'])

When using QGIS, the Raster Layer import dialogue can read HDF5 directly if GDAL supports the driver. However, you may need to specify the dataset path using the syntax:

/path/to/file.h5:/Data

Metadata extraction can be performed via the Raster > Metadata menu. Calibration constants are often stored under Calibration attributes.

2.5 Importing GeoTIFF and NetCDF

GeoTIFF is straightforward: simply open the file in QGIS or ENVI. NetCDF requires specifying the variable name:

gdal_translate -of GTiff input.nc:variable output.tif

Both formats may embed georeferencing information. Always verify the PROJCS or grid_mapping metadata to ensure correct spatial alignment.

2.6 Metadata Extraction and Validation

Once the cube is loaded, validate the metadata against the mission’s documentation. Key checks include:

• Wavelength spacing: confirm that the wavelengths are contiguous and free of gaps.
• Spatial resolution: ensure pixel_size matches the sensor’s specifications.
• Radiometric calibration: verify the presence of radiance_mult and radiance_add (for radiance) or reflectance_mult and reflectance_add (for reflectance).
• Georeferencing: confirm the CRS matches the expected EPSG code.

When discrepancies arise, consult the instrument handbook or reach out to the data centre.

2.7 Radiometric Calibration Workflow

Calibration converts raw digital numbers (DN) to physical units. The general formula is:

Physical Unit = DN × Multiplier + Addend

For example, to obtain radiance from an ENVI cube:

radiance = cube * meta['radiance_mult'] + meta['radiance_add']

Similarly, converting radiance to surface reflectance typically requires dividing by the solar irradiance and accounting for the sun–sensor geometry:

solar_irradiance = meta['solar_irradiance']  # W/m^2/nm
theta_sun = meta['solar_zenith_angle']  # degrees
reflectance = radiance / (solar_irradiance * np.cos(np.radians(theta_sun)))

Always perform these calculations on a per‑band basis. Many libraries, such as spectral-python, provide ready‑made functions for common sensors.

2.8 Comparative Analysis of Alternative Software

Choosing the platform depends on the project’s budget, the need for automation, and the user’s comfort with programming.

2.9 Case Study: Importing a Sentinel‑2 MSI Scene

Although Sentinel‑2 is multispectral, the workflow mirrors hyperspectral processing. The scene comes as a zip of .jp2 files, each representing a band. Steps:

1. Unzip the archive.
2. Use rasterio to stack the bands into a single array:

import rasterio
bands = []
for i in range(1, 13):
    with rasterio.open(f'B{i:02}.jp2') as src:
        bands.append(src.read(1))
stack = np.stack(bands, axis=0)

with rasterio.open('B01.jp2') as src:
    meta = src.profile
print(meta['crs'])
print(meta['transform'])

Although simpler than true hyperspectral data, this example illustrates the same principles: careful file handling, metadata extraction, and calibration.

2.10 Troubleshooting Common Issues

• Incorrect band order – some formats store bands in reverse (e.g. top‑to‑bottom). Check the bands.order metadata and re‑index if necessary.
• Missing metadata – request a metadata.xml from the provider or reconstruct from the sensor’s documentation.
• Large file size – use cloud‑based tools (e.g. Google Earth Engine) or compress the cube with lossless methods.

Document every step, including the exact command line or menu path used, to ensure reproducibility and to aid peer review.

Section 3: From Spectra to Material Identification

In the previous sections we learned how to acquire hyperspectral image data and how to visualise the raw spectral cube. The final step in the workflow is to translate those spectra into material identities. This requires a disciplined approach to spectral extraction, normalisation, and library matching. The following section walks through each stage, discusses common pitfalls, and shows how different choices can influence the final classification.

3.1 Spectral Extraction: From Pixels to Signatures

Every pixel in a hyperspectral image contains a vector of radiance or reflectance values across the measured wavelength range. The first task is to decide which pixels to treat as representative of a material. Two common strategies are:

• Pixel‑by‑pixel extraction: Treat every pixel individually. This preserves fine spatial detail but inflates the number of spectra to be processed, potentially increasing noise and computational load.
• Region‑based extraction: Define a polygon or rectangular region of interest (ROI) that encloses a homogeneous material. The spectrum is then derived by averaging all pixels within the ROI.

For example, when analysing a vineyard canopy, you might average over a 5 × 5 pixel patch to obtain a single “leaf” spectrum, thereby reducing speckle noise. In contrast, for a patterned artwork where each pigment occupies a single pixel, pixel‑by‑pixel extraction is essential.

3.2 Normalisation to Reflectance

Raw detector counts are influenced by illumination geometry, sensor response, and atmospheric scattering. Converting to surface reflectance is therefore mandatory for meaningful library comparison. Two widely used normalisation methods are:

1. Dark‑current subtraction and white‑reference division: Subtract a dark frame (sensor counts with no illumination) from the raw spectrum and divide by the response of a calibrated white reference (e.g. Spectralon) measured under the same lighting conditions. The resulting vector approximates absolute reflectance.
2. Continuum removal: Fit a convex hull (or straight‑line continuum) to the spectrum and divide the original by this envelope. This emphasises absorption features and removes broad‑band illumination effects, producing a pseudo‑reflectance spectrum.

Each method has trade‑offs. Dark‑current subtraction is straightforward but relies on a stable dark reference. White‑reference division yields true reflectance but requires a perfectly flat white target and careful handling of sensor non‑linearity. Continuum removal is robust to illumination changes but alters the absolute scale, making it unsuitable for absolute material quantification.

3.3 Library Matching: The Core of Identification

Once normalised, the spectrum is ready to be compared against a spectral library. Libraries may be empirical (measured spectra of known materials) or synthetic (generated by radiative transfer models). Matching algorithms can be broadly categorised into distance‑based and statistical methods.

3.3.1 Distance‑Based Matching

The simplest approach is to compute a distance metric between the query spectrum \(q\) and each library spectrum \(l_i\):

def euclidean_distance(q, l):
    return np.sqrt(np.sum((q - l)**2))

Variants include:

• Euclidean distance: Sensitive to absolute differences; works well when spectra are on the same scale.
• Spectral Angle Mapper (SAM): Treats spectra as vectors; only the angle matters, making it invariant to intensity scaling.
• Correlation coefficient: Measures linear similarity; useful when absolute levels vary but shape is preserved.

For example, a leaf spectrum matching a chlorophyll library entry might yield a small Euclidean distance, whereas a stone spectrum would match a silicate entry despite differing absolute levels.

3.3.2 Statistical Matching

More sophisticated algorithms consider the probabilistic nature of spectral data. Common methods include:

• Maximum Likelihood Estimation (MLE): Models each library class as a multivariate Gaussian; computes the likelihood that the query belongs to each class.
• Support Vector Machines (SVM): Trains a classifier on labelled spectra; can handle non‑linear boundaries using kernels.
• Random Forests: Ensemble of decision trees; robust to noise and capable of ranking feature importance.

Statistical approaches generally outperform simple distance metrics when the library contains many similar classes or when the data are noisy. However, they require a representative training set and careful cross‑validation to avoid over‑fitting.

3.4 Case Study: Identifying Urban Materials

Consider an urban street scene acquired with a 400 nm–2500 nm sensor at 10 cm spatial resolution. The goal is to map asphalt, concrete, and painted signage. The workflow proceeds as follows:

1. Extraction: Use ROI extraction to isolate a 20 × 20 pixel patch of asphalt and a 5 × 5 patch of a blue paint stripe.
2. Normalisation: Apply white‑reference division to obtain reflectance spectra.
3. Library: Use the ASTM E 313 2008 spectral library, which contains measured spectra for common construction materials.
4. Matching: Compute SAM for each query against the library entries. Asphalt shows a SAM of 0.015 with the ASTM asphalt entry; the blue paint matches a synthetic paint library entry with a SAM of 0.010.
5. Decision: Assign labels based on the lowest SAM value, with a threshold of 0.02 to flag ambiguous matches.

In this example, distance‑based SAM proved adequate because the materials have distinct spectral shapes. If the scene included a wide variety of paint colours, a Random Forest classifier trained on a larger paint library might be preferable.

3.5 Common Pitfalls and Best Practices

Another subtle issue is the choice of spectral resolution. High‑resolution spectra (1 nm) capture fine absorption features but increase data volume and noise. Coarser resolutions (10 nm) reduce noise but may blur diagnostic features. A pragmatic approach is to resample the spectra to a resolution that balances feature preservation and noise suppression, often 5–10 nm for most materials.

3.6 Emerging Alternatives: Deep Learning

Recent advances in machine learning have introduced convolutional neural networks (CNNs) that operate directly on hyperspectral cubes. Unlike classical library matching, CNNs learn hierarchical spectral–spatial features from labelled data. Studies have shown that a 3‑D CNN can outperform distance‑based methods on complex scenes such as forest canopy discrimination. However, deep learning demands large training sets and substantial computational resources, and its black‑box nature can hinder interpretability.

3.7 Summary

Material identification from hyperspectral data is a multi‑step process: extract representative spectra, normalise to reflectance, and compare against a suitable library using an appropriate matching algorithm. The choice of extraction strategy, normalisation technique, and matching method should be guided by the characteristics of the scene, the availability of reference data, and the computational constraints. By carefully applying these steps, practitioners can reliably translate raw spectral measurements into actionable material maps.

Section 4: Comparative Analysis with Traditional Imaging

In the realm of remote sensing, the choice between hyperspectral imaging (HSI) and conventional RGB imagery is often framed in terms of the trade‑off between spectral richness and spatial detail. This section dissects the three key dimensions that underpin this decision: information content, spatial resolution, and application suitability. By examining concrete examples from agriculture, mineralogy and environmental monitoring, we can appreciate where each modality excels and where they complement one another.

Information Content

Information content refers to the quantity and quality of data that can be extracted from an image. Hyperspectral sensors capture hundreds of contiguous spectral bands, typically spanning from the visible (VIS) through near‑infrared (NIR) into short‑wave infrared (SWIR). RGB cameras, in contrast, record only three broad bands: red, green and blue.

• Spectral Detail: With a typical HSI sensor such as the EnMAP or AVIRIS, each pixel contains a spectrum with ~250 bands at 10 nm intervals. This allows the detection of subtle absorption features associated with specific molecular bonds. For instance, the 1.4 µm band is highly sensitive to water content in vegetation, whereas the 2.2 µm band reveals clay mineralogy in soils. An RGB image, by definition, cannot resolve these narrow features; its three bands are merely broad composites.
• Signal‑to‑Noise Ratio (SNR): HSI data often suffer from lower SNR per band due to photon redistribution across many channels. Advanced processing (e.g. spectral smoothing or principal component analysis) mitigates this, but the inherent trade‑off remains. RGB cameras typically enjoy high SNR because all photons are concentrated into three bands, making them more robust in low‑light conditions.
• Dimensionality: The high dimensionality of HSI facilitates sophisticated algorithms such as spectral unmixing, endmember extraction and chemometric analysis. These techniques can separate mixed signals within a single pixel, revealing the presence of multiple materials at sub‑pixel scales. RGB imagery, lacking spectral depth, relies heavily on texture and shape, limiting its ability to perform such decomposition.

Spatial Resolution

Spatial resolution describes the size of the ground area represented by each pixel. Traditional RGB sensors, whether satellite (e.g., WorldView‑4 at 30 cm) or airborne (e.g., drones with 5 cm resolution), often provide finer spatial detail than most HSI platforms.

1. Sensor Design Constraints: Hyperspectral sensors must accommodate a large number of spectral channels, which typically requires dispersive optics or filter arrays. This increases the sensor size and limits the maximum achievable pixel density. Consequently, airborne HSI systems usually settle for 1–5 m spatial resolution, while satellite HSI missions (e.g., Sentinel‑2 MSI) have coarser resolutions of 10–20 m.
2. Trade‑off with Spectral Coverage: Expanding the spectral range (e.g., adding SWIR) often forces a reduction in spatial resolution to keep data volumes manageable. For example, the NASA EO‑1 Hyperion sensor offers 30 m spatial resolution across 220 bands, whereas the newer PRISMA satellite achieves 5.6 m spatial resolution but with only 153 bands. These design choices directly influence the suitability for point‑level versus area‑level analyses.
3. Practical Implications: In precision agriculture, a 5 cm drone RGB image can reveal individual crop rows and small weed patches, enabling targeted spraying. Conversely, a 30 m HSI image may only delineate broad field boundaries but can detect crop stress signatures across the entire field. Thus, the spatial resolution requirement is often dictated by the scale of the management decision.

Application Suitability

The decision to employ HSI versus RGB hinges on the specific objectives of the mission. Below are illustrative case studies that highlight where each modality shines.

Case Study 1: Crop Health Monitoring

HSI Advantage: The Normalised Difference Vegetation Index (NDVI) can be derived from a single NIR–red pair, but more nuanced indices such as the Red Edge NDVI (RENDVI) require bands around 705–740 nm. HSI data provide these bands, allowing early detection of nitrogen deficiency or disease stress before visible symptoms appear. A field trial in the UK Midlands used the Sentinel‑2 MSI (10 m) to generate RENDVI maps, revealing nitrogen hotspots that guided variable‑rate fertiliser application.

RGB Limitation: While an RGB drone can capture fine spatial detail, it lacks the spectral bands required for RENDVI or water‑stress detection. Therefore, RGB is best suited for mapping canopy structure or identifying visible weed infestations, not for early stress detection.

Case Study 2: Mineral Exploration

HSI Advantage: Mineral mapping relies on detecting characteristic absorption features in the SWIR region. For instance, the 1.25 µm band is diagnostic of orthoclase feldspar, while the 2.73 µm band indicates quartz. A hyperspectral survey of the Scottish Highlands using the AVIRIS sensor produced a mineralogical map that distinguished between quartz‑rich and feldspar‑rich outcrops with centimetre‑scale accuracy.

RGB Limitation: RGB imagery cannot resolve these subtle spectral signatures; it merely presents colour composites that may be influenced by illumination and atmospheric conditions. Hence, RGB is unsuitable for mineral identification but can be useful for contextualising geological features within a broader landscape.

Case Study 3: Environmental Monitoring of Water Bodies

HSI Advantage: Hyperspectral data can quantify chlorophyll‑a concentration, dissolved organic matter, and suspended sediment by exploiting narrow bands in the 680–700 nm and 740–750 nm ranges. A study over the Thames estuary used HSI to create a high‑resolution map of algal blooms, enabling targeted remediation.

RGB Limitation: RGB images provide a visual impression of colour but cannot separate overlapping spectral signals. They are therefore inadequate for precise water quality assessment.

Case Study 4: Cultural Heritage Documentation

RGB Advantage: High‑resolution RGB imagery from drones captures surface detail, texture, and architectural features with centimetre precision. This is invaluable for creating orthophotos and 3D models of monuments.

HSI Limitation: While HSI can detect material composition (e.g., distinguishing between different stone types), its lower spatial resolution may blur fine architectural details, making it less suitable for detailed conservation work.

Complementary Use Cases

In many scenarios, a hybrid approach yields the best results. For example, a UAV can first capture a high‑resolution RGB mosaic to map field boundaries and canopy structure. A subsequent HSI flight at a coarser resolution can then be georeferenced to the RGB base map, allowing the analyst to overlay spectral indices onto precise spatial features. This integrated workflow maximises both spectral and spatial information, enabling more informed decision‑making.

Key Takeaways for Learners

• HSI offers unparalleled spectral depth, essential for material identification and early stress detection, but typically at the cost of spatial resolution.
• RGB provides high spatial fidelity and excellent SNR, ideal for detailed structural mapping and visual interpretation.
• The optimal choice depends on the mission scale: point‑level precision tasks favour RGB; area‑level compositional analysis favours HSI.
• Hybrid workflows that combine both modalities can harness the strengths of each, mitigating their individual limitations.

In summary, the comparative analysis underscores that hyperspectral and RGB imaging are not mutually exclusive but rather complementary tools within the remote sensing toolbox. Understanding their distinct capabilities and constraints empowers practitioners to design missions that align with their specific analytical objectives.

Further Reading

• Jensen, J. R. (2007) — Remote Sensing of the Environment: An Earth Resource Perspective. Provides foundational principles for interpreting hyperspectral data and sensor characteristics.
• Lu, Z., & Weng, Q. (2007) — Remote Sensing Image Interpretation: Theory and Practice. Offers detailed methodologies for classifying and analysing hyperspectral imagery.
• National Geospatial-Intelligence Agency (NGIA) (2015) — Hyperspectral Imaging Handbook. Official guidance on data acquisition, calibration, and interpretation of hyperspectral datasets.
• Wulder, M. A., & Long, J. (2016) — USGS Landsat Data Users Manual, Chapter 4. Discusses spectral data structures and processing workflows applicable to hyperspectral images.
• International Society for Photogrammetry and Remote Sensing (ISPRS) (2018) — ISPRS Commission 4: Hyperspectral and Multispectral Sensors – Technical Guidelines. Covers standardised data formats and metadata conventions.