Hyperspectral imaging

Also known as Imaging spectroscopy, Hyperspectral cameras

Hyperspectral imaging records the light reaching each point in hundreds of narrow wavelength bands, so it can separate materials that look the same to a normal colour camera. Two hyperspectral Earth-observation satellites are in orbit and food-sorting systems have been tested on industrial lines, while crop disease detection and medical diagnosis are still mostly at research stage.

Pilot and early commercialDemonstratedReviewed 21 Oct 2026
AI illustration · Illustrative scene, not a photograph of the specific device or experiment.

In plain language

An ordinary camera records three values for each pixel: red, green and blue. A hyperspectral camera records the light arriving from each point across a long run of narrow wavelength bands, often hundreds of them, reaching from visible light into the infrared. Each material reflects and absorbs light in its own pattern across those bands, so the pattern works something like a fingerprint.

This lets a machine tell apart things that look alike to the eye, such as a healthy leaf and a leaf with an early infection, or a stone and a grain. The trade-off is data. Each image is a stack of many layers rather than three, so the files are much larger, and readings must be calibrated so that the same material gives the same signal on different days and under different light.

Technical explanation

A hyperspectral sensor is usually an imaging spectrometer: light is split into its wavelengths, and a detector array records the intensity in each band. Earth-observation satellites build an image by sweeping the sensor across the ground as the spacecraft moves. The DLR states that EnMAP covers 420 to 2450 nm in 232 spectral bands at a 30 m ground resolution. The eoPortal entry for PRISMA lists 66 VNIR and 173 SWIR channels at 30 m over a 30 km swath. Planet says Tanager-1 images roughly 400 to 2500 nm in one capture, with over 400 bands at 30 m. Band counts and ranges differ between sources, so product metadata should be checked before a specific band list is used.

Analysis usually means turning each pixel's spectrum into a label or a number, using models such as linear classifiers, support vector machines, random forests or neural networks. Reported accuracy depends heavily on the test set-up. A 2023 Scientific Reports study grew 36 tomato plants in phytotron conditions and tested machine-learning detection of infection from measurements aggregated over several days; the authors advise validating the models under real crop conditions. A 2024 Foods study of legume sorting reported 99.36 to 99.67 per cent correct classification for broad beans, lentils and chickpeas on an industrial optical sorter, with no false positives and 0.33 to 0.64 per cent of good legumes discarded.

Medical use is at research stage. A 2022 review of deep learning in medical hyperspectral images describes tissue imaging for skin pathology and for locating tumours during surgery, and says hyperspectral imaging can give broad-area images of tissue that may improve diagnostic accuracy. It also notes that most acquisition devices are relatively large, that real-time analysis has yet to be developed, and that the device it describes is still a prototype needing refinement before clinical use. Calibration is a further constraint: a 2009 comparison of airborne reflectance calibration methods found that a white-panel method was reliable only for very short flights near local solar noon, while a continuous reference panel was the only method found unbiased at any time of day.

What the evidence says

Each statement is typed by the strength of its evidence and linked to the source that supports it.

Demonstrated Shown in a published experiment, test or shipping product.

DLR states that the EnMAP satellite covers 420 to 2450 nm in 232 spectral bands at a 30 m ground resolution, and that it launched on 1 April 2022.

Sources: [1] · Mission summary

The PRISMA satellite has 66 VNIR and 173 SWIR channels at 30 m resolution over a 30 km swath, launched on 22 March 2019, and eoPortal lists its status as operational (extended).

Sources: [3] · Performance specifications

A 2023 Scientific Reports study of tomato plants grown in phytotron conditions (36 plants) reports that machine-learning models detected infection from hyperspectral measurements aggregated over several consecutive days, and that the authors consider diagnosis in less than the accepted seven days possible; they advise validation under real crop conditions.

Sources: [5] · Results and Discussion · medium confidence

The authors of a 2024 legume sorting study reported that an industrial optical sorter gave 99.36 to 99.67 per cent correct classification for broad beans, lentils and chickpeas, recorded no false positives, and discarded 0.33 to 0.64 per cent of good legumes.

Sources: [4] · Section 3.2 and Table 4 (industrial sorter results) · medium confidence

The same 2022 review says most medical hyperspectral acquisition devices are relatively large, that real-time analysis algorithms have yet to be developed, and that the device it describes is still a prototype needing refinement before clinical use.

Sources: [6] · Limitations section · medium confidence

For airborne spectrometer data, a continuous reference panel method was the only one of three reflectance calibration methods that the authors found unbiased at any time of day, with an estimated uncertainty of 0.0025 ± 0.0005 reflectance units outside atmospheric absorption bands.

Sources: [7] · Results and conclusions

Reported, unverified Stated by a company or team but not independently confirmed.

Planet states that Tanager-1 collects over 400 spectral bands at 30 m resolution and images about 400 to 2500 nm in one capture; these are company figures and the page gives no independent calibration details.

Sources: [2] · medium confidence

A 2022 review reports that hyperspectral imaging can give broad-area images of tissue that may improve diagnostic accuracy, and that it has been studied for skin pathology and for locating tumours to guide surgery.

Sources: [6] · Applications section · medium confidence

Proposed A design or plan that has not yet been built or tested.

The authors of the 2023 tomato study argue that the approach could diagnose infection in less than the accepted seven days, but note that models must be validated under real crop conditions, which they describe as more challenging.

Sources: [5] · Discussion · medium confidence

Limitations and open problems

  • Data volume: each pixel holds hundreds of values, so files are far larger than ordinary images. The sources reviewed here did not give verified storage, downlink or processing figures.
  • Calibration: readings must be converted to reflectance so that the same material gives the same signal across dates, times and light conditions. Miura and Huete (2009) found that a white-panel method was reliable only for very short flights near solar noon.
  • Lab accuracy is not field accuracy: the 2023 tomato study used controlled phytotron conditions with 36 plants, and its authors advise validating the models under real crop conditions. The food sorting results come from the authors' own sorter set-ups and a limited range of products.
  • Medical devices described in a 2022 review are relatively large and still prototypes, real-time analysis is not yet developed, and public training data are scarce because of privacy and ethics rules (Cui and colleagues, 2022).
  • Band counts and ranges differ between sources for the same mission, so users should check product metadata before relying on a specific figure.

Possible future uses

Speculative

Ideas, not predictions. None of these has been demonstrated.

  • Satellite monitoring of crop stress across whole fields, flagging disease before symptoms are visible to people on the ground. The current evidence comes from controlled tests on individual crops.
  • Compact hyperspectral probes that help surgeons see tumour edges in real time during operations, if devices become small enough for theatre use.
  • A single food-sorting model that handles many product types on one line, reducing the need to retrain for each crop.

Product concepts: Household bin that sorts waste for recycling

Timeline

  1. 3 Feb 2009Miura and Huete compare three reflectance calibration methods for airborne hyperspectral data and find that only the continuous reference panel method is unbiased at any time of day. Sources: [7]
  2. 22 Mar 2019The PRISMA hyperspectral satellite launches, with 30 m ground resolution and a 30 km swath. Sources: [3]
  3. 1 Apr 2022The EnMAP hyperspectral satellite launches, covering 420 to 2450 nm. Sources: [1]
  4. 11 May 2023A Scientific Reports study reports machine-learning detection of tomato disease infection in laboratory-grown plants, using hyperspectral measurements aggregated over several days. Sources: [5]
  5. 26 Jan 2024A Foods paper reports 99.36 to 99.67 per cent correct classification for three legume types on an industrial optical sorter. Sources: [4]
  6. 19 Sept 2024Planet's Tanager-1 hyperspectral satellite captures first light images over Karachi, Pakistan, according to Planet. Planet says the satellite collects over 400 bands at 30 m resolution. Sources: [2]

Futures this could help bring about

AI illustration
How far away?

Finding every methane leak from space

Pilots · 3 of 7 steps done

Glossary

Spectral band
A narrow slice of the light spectrum, measured as a range of wavelengths, that a sensor records as one layer of data.
Data cube
A hyperspectral image stored as two spatial axes plus one spectral axis, so each pixel carries a full spectrum.
Reflectance calibration
Converting raw sensor readings into the fraction of light each surface reflects, so that results are comparable across dates and conditions.
Ground sampling distance
The size on the ground that one image pixel covers, such as 30 m for EnMAP and PRISMA.

Sources

  1. other
    EnMAP - Environmental Mapping Program
    German Aerospace Center (DLR), 2022. Accessed 21 Oct 2026.
  2. company announcement
    Celebrating one year of Tanager-1
    Planet Labs PBC, 15 Aug 2025. Accessed 21 Oct 2026.
  3. other
    PRISMA (Hyperspectral)
    eoPortal Directory, 31 Jul 2023. Accessed 21 Oct 2026.
  4. peer reviewed paper
    Optimisation of an Industrial Optical Sorter of Legumes for Gluten-Free Production Using Hyperspectral Imaging Techniques
    Roberto Romaniello, Antonietta Eliana Barrasso, Claudio Perone et al.. Foods (MDPI), 26 Jan 2024. DOI: 10.3390/foods13030404. Accessed 21 Oct 2026.
  5. peer reviewed paper
    Early detection of Solanum lycopersicum diseases from temporally-aggregated hyperspectral measurements using machine learning
    Michał Tomaszewski, Jakub Nalepa, Ewa Moliszewska et al.. Scientific Reports, 11 May 2023. DOI: 10.1038/s41598-023-34079-x. Accessed 21 Oct 2026.
  6. review article
    Deep Learning in Medical Hyperspectral Images: A Review
    Rong Cui, He Yu, Tingfa Xu et al.. Sensors (Basel), 13 Dec 2022. DOI: 10.3390/s22249790. Accessed 21 Oct 2026.
  7. peer reviewed paper
    Performance of Three Reflectance Calibration Methods for Airborne Hyperspectral Spectrometer Data
    Tomoaki Miura, Alfredo R. Huete. Sensors (Basel), 3 Feb 2009. DOI: 10.3390/s90200794. Accessed 21 Oct 2026.

Revision history

3 changes since this profile was created.

Changes on 21 Oct 2026
Published on owner instruction (content loop iteration 51)095cfd1
Sources and claims re-checked against originalsfbb434c
Initial profile drafted from primary sources7e66249