Hyperspectral imagery is the third modality in this series on satellite image segmentation, after multispectral optical and radar. It differs from both in one simple respect: where an ordinary image carries a few bands, a hyperspectral image carries hundreds.
That spectral richness changes the nature of the problem. It gives access to a material identification no other satellite modality permits, and it introduces difficulties the preceding articles did not encounter. This article sets them out. It extends the article on segmenting radar imagery.
What hyperspectral brings
A difference of nature separates it from ordinary multispectral imagery. A multispectral image samples the spectrum in a few wide bands, giving a coarse signature sufficient to distinguish broad families of surface. A hyperspectral image samples the spectrum in hundreds of narrow contiguous bands, producing for each pixel a near-continuous reflectance curve. That spectral continuity changes what becomes identifiable. A material presents characteristic absorption features, narrow and located at precise wavelengths, which a few wide bands average away and which fine spectral resolution reveals. Three families of application follow directly. Mineral identification, resting on well documented absorption signatures. Biochemical characterisation of vegetation, water, nitrogen and pigment content. And detection of specific materials, hydrocarbons, plastics and pollutants notably. One observation runs across those three families and it deserves stating. Hyperspectral does not distinguish the same classes better; it permits classes the other modalities cannot define at all.The curse of dimensionality
This theoretical difficulty explains most project failures. It dominates hyperspectral satellite image segmentation. The number of variables per pixel rises from a handful to several hundred, producing a counterintuitive effect: beyond a certain number of bands, adding information degrades classifier performance at constant training volume. Three distinct mechanisms explain it. The volume of the variable space grows exponentially with the number of dimensions, which leaves training examples sparse whatever their quantity. Neighbouring bands are strongly correlated, adding dimension without adding information. And noise accumulates across all bands while useful signal concentrates in a few. One major practical consequence follows. The need for annotated examples grows with the number of bands exploited, which makes the ratio between dimensionality and corpus size a design decision rather than a technical detail. That consequence is commercially decisive. Hyperspectral is the modality where the annotation need is highest per unit area, while its public corpora are the scarcest.Dimension reduction
It constitutes an obligatory step in any hyperspectral project. This operation responds to the preceding difficulty. Three approaches are distinguishable and their effects on annotation differ markedly. Band selection retains a subset of the original bands, which preserves the physical interpretability of each variable and permits justifying a result. Transformation projects the data into a reduced space, which compresses information efficiently and removes the direct link to wavelengths. And unmixing estimates the proportions of pure materials in each pixel, an approach specific to this modality the following section details. One important consequence follows. Reduction by transformation makes annotation difficult, the annotator no longer being able to relate what they see to a physical property, which argues for annotating on the original data and reducing afterwards.Mixed pixels and unmixing
This notion specific to the modality changes the nature of the deliverable. A hyperspectral pixel generally covers several tens of metres, an area that rarely contains a single material. Its signature is therefore a mixture of elementary signatures. Unmixing means decomposing that signature into proportions of pure constituents, producing not one class per pixel but a vector of proportions. Three consequences follow and they sit outside the usual frame of satellite image segmentation. The result is continuous rather than discrete, which makes the metrics set out in the opening article inapplicable. The reference annotation must supply proportions rather than labels, a markedly harder exercise for an annotator. And validation requires quantified field measurements rather than photo-interpretation. One recommendation follows for scoping. Establishing whether the client expects a class map or a proportion map determines the method, the reference type and the cost, and that question does not arise in the other modalities.Spectral libraries in satellite image segmentation
This resource changes a project’s economics and it is poorly known. Libraries of material spectral signatures, measured in the laboratory or in the field, are publicly available and cover minerals, vegetation and artificial materials. Three practical uses follow. Classification by matching compares a pixel’s signature against those in the library, permitting identification without an annotated corpus. Initialising an unmixing supplies the reference pure signatures. And verifying an annotation permits checking that a label is spectrally plausible. That first use deserves particular emphasis. It is a notable exception in this series: the only case where a classification can be obtained without annotated ground truth, which considerably lowers the entry barrier on certain applications. Three important caveats nonetheless frame it. A laboratory signature differs from one observed from space, through the atmosphere, the illumination geometry and the surface state. Natural materials present a variability the library does not capture. And a mixed pixel corresponds to no pure signature.Atmospheric correction
This step takes on particular importance in hyperspectral work. The atmosphere absorbs and scatters radiation in a way strongly dependent on wavelength, which distorts the measured spectrum relative to the surface’s true spectrum. That distortion is tolerable in multispectral work, where wide bands are compared. It is disqualifying in hyperspectral work, where the point is precisely to find narrow absorption features the atmosphere can imitate or mask. Three consequences follow. Some wavelength ranges are unusable, atmospheric absorption there being total, and they must be excluded explicitly. Correction quality conditions comparability between acquisitions, two differently corrected images not being comparable. And a library signature can only be matched against correctly corrected data. One important practical consequence follows for annotation. The correction level must be documented and identical between corpus and production, a requirement identical to the one set out for radar filtering and for the same reasons.The resolution trade-off
This physical compromise structures the hyperspectral modality. It explains why these sensors have modest spatial resolution. A sensor has a limited quantity of light available. Dividing it across hundreds of narrow bands means each band receives little, which degrades the signal to noise ratio. Three levers allow compensation and each is paid for with a counterpart. Widening the instantaneous field of view, meaning degrading spatial resolution, to collect more light per band. Widening the bands, which amounts to giving up the spectral fineness sought. Or lengthening integration time, which reduces coverage. One practical consequence follows and it is decisive at scoping. A project cannot simultaneously hold fine spatial resolution, fine spectral resolution and wide coverage, and the choice between those three terms precedes any consideration of method. That constraint also explains the prevalence of mixed pixels set out above. It is not a defect of current sensors but a physical property of the compromise, which makes unmixing structurally necessary rather than transitionally useful.Available sources
An overview is needed, availability being this modality’s primary constraint in satellite image segmentation. Three regimes coexist with a distribution very different from that of optical imagery. Spaceborne hyperspectral missions are few and recent, with partial coverage and low revisit, which strongly limits regular monitoring uses. Airborne campaigns supply superior spatial and spectral resolution over restricted areas, at high acquisition cost and with control over conditions. And field measurements supply point signatures serving as reference and calibration. One structuring consequence follows for scoping. Most hyperspectral projects concern restricted areas rather than wide territories, which brings them closer to a study logic than to a cartographic production logic. That characteristic carries a commercial implication. Projects are smaller, more technical and better remunerated per unit area than multispectral projects, which suits a provider able to mobilise expertise rather than volume.What annotation requires
Four requirements distinguish it in hyperspectral satellite image segmentation. They justify separate treatment. Annotation often concerns points rather than surfaces, a geolocated field measurement constituting the most reliable reference and exhaustive delineation being rarely possible. Qualification requires domain competence, mineralogy, agronomy or materials science depending on the application, markedly more specialised than object recognition. Checking proceeds through spectral plausibility rather than visual inspection, a label being verifiable by comparing the pixel’s signature against a reference. And the quantity needed is high, for the dimensionality reason set out above, which sits in tension with the high unit cost of each annotation. That tension is the central difficulty of any hyperspectral project. It is generally resolved through a mixed strategy: a few high quality field measurements, extended by spectral matching and assisted photo-interpretation.Uses and their maturity
Five domains employ this modality in satellite image segmentation and their maturity differs markedly, which orients a provider. Mineral prospecting is the historical and most mature use, mineral signatures being well documented and the value of a detection high. Precision agriculture uses biochemical indices for stress and nutrition monitoring, with growing adoption and an acquisition cost constraint. Water quality monitoring detects pigments and suspended matter, an application where hyperspectral brings a fineness multispectral does not reach. Pollution detection, hydrocarbons and industrial discharges, presents high unit value and occasional need. And urban material mapping, roofs and pavements, serves risk assessment and heat island monitoring. One important observation runs across those five domains. All rest on material identification rather than shape recognition, which confirms that this modality answers a question of a different nature.The conventions to fix
Six distinct decisions must be settled before production and they differ from those of the other modalities. The atmospheric correction level and its method. The bands excluded, atmospheric absorption and noisy bands, and the criterion for their exclusion. The nature of the deliverable, class map or proportion map. The material reference used and its link to an existing spectral library. The dimension reduction method and when it occurs relative to annotation. And the field measurement protocol, whose geolocation and representativeness condition the value of the reference.Evaluating a hyperspectral product
It complicates the validation set out in the previous series. Four particularities distinguish it from the other satellite image segmentation modalities. The reference is point-based rather than area-based, which precludes overlap metrics and imposes point evaluation, with the sampling precautions that implies. The reference is expensive, a geolocated field measurement requiring travel, which strongly limits the size of the evaluation sample. The correspondence between a point measurement and a pixel of several tens of metres is approximate by construction, a survey over one square metre not necessarily representing the pixel’s surface. And evaluating an unmixing requires comparing proportions, an exercise for which the usual segmentation metrics do not suit and which calls for quantitative error measures. Those four particularities lead to a recommendation. A hyperspectral evaluation protocol is designed jointly with the field measurement protocol, the two being one operation, where in the other modalities the reference is built afterwards on the imagery.The cost of a hyperspectral corpus
It is the least favourable structure of the three modalities examined. Six factors compose it in satellite image segmentation. Four raise it. The need for examples is higher, for the dimensionality reason set out above. The unit cost of a reference is high, a geolocated field measurement requiring travel where photo-interpretation is done at a desk. The expertise required is specialised and scarce, mineralogy or agronomy depending on the application. And acquisition itself is expensive, an airborne campaign representing a line the other modalities do not carry at the same level. Two offset it. The areas concerned are small, a hyperspectral project rarely covering a wide territory. And spectral libraries permit extending a small number of field measurements, which multiplies the value of each reference acquired. The balance gives a project expensive per unit area and modest in absolute volume, a profile suiting an expertise engagement rather than mass production.Approaching a hyperspectral project
Five questions quickly eliminate infeasible requests. They scope a hyperspectral satellite image segmentation project. Does the information sought concern material composition. If not, another modality suits better and costs markedly less. Do the target materials present documented signatures. Their presence in a library considerably lowers the entry barrier, their absence requires a prior measurement campaign. What spatial coverage is expected. The trade-off set out above makes wide coverage and spectral fineness incompatible, and that constraint sometimes settles feasibility. Is field access possible. This modality’s reference coming from the ground, its inaccessibility radically changes the method and the cost. And does the client expect classes or proportions. That question determines the nature of the deliverable, the validation method and the type of reference to be built. Those five answers determine feasibility before any consideration of model. Asking them in one meeting avoids committing to a project whose failure was predictable.What this use case teaches
This observation closes the examination of the three modalities. It concerns the nature of the question asked. Optical imagery answers the question of what an object appears to be. Radar answers the question of how it is structured. Hyperspectral answers the question of what it is made of. Those three questions are not interchangeable and a misdirected project is not corrected by a better model. Three consequences follow for a satellite image segmentation provider. The choice of modality belongs to scoping rather than to technique, and an error at that stage invalidates everything after it. Annotation competence fragments by modality, each requiring a reading the previous one does not prepare. And the value of an expertise grows with the scarcity of the modality, hyperspectral sitting at the extreme of that scale with the scarcest corpora and the least widespread competences. That third consequence deserves a qualification. Scarcity produces high unit value and a narrow market, which makes it a complementary activity rather than a principal one, unlike radar whose demand grows on substantial volumes.The position of this modality
One final observation situates hyperspectral relative to the other two modalities and it bears on an investment choice. Three criteria permit comparing the three modalities from a provider’s point of view. Demand volume markedly favours optical, followed by radar whose growth is documented, hyperspectral remaining a niche market. Competence scarcity favours hyperspectral, then radar, optical annotation being widely available. And unit value follows the same order as scarcity, a hyperspectral annotation being priced markedly above an optical one. One strategic reading follows. Radar occupies the most favourable position, combining growing demand with a scarce competence, while hyperspectral offers superior unit value in a market too narrow to carry an activity. That reading orients an investment decision. Developing a radar capability is a defensible principal axis, while a hyperspectral capability is justified as a complement permitting occasional high value requests to be answered rather than as the foundation of an offering.Where the three modalities combine
One satellite image segmentation arrangement deserves describing, since the three modalities examined are complementary rather than competing. Each answers a different question, which means a project needing more than one answer needs more than one modality. Three combinations occur in practice. Optical and radar, which the radar article detailed, where radar fills cloud gaps and adds structural information. Optical and hyperspectral, where the multispectral series supplies temporal coverage and the hyperspectral acquisition supplies material identification at key dates. And all three, in demanding applications where structure, appearance and composition all matter. Two difficulties recur across these combinations. Spatial resolutions differ, which requires deciding at which scale the analysis is performed and accepting the information loss that decision implies. And acquisition dates rarely coincide, which is tolerable on stable surfaces and problematic on changing ones. One consequence follows for a data provider. A multimodal corpus, annotated consistently across modalities over the same areas, is among the scarcest resources in this field and among the most valuable, precisely because building one requires competence in each modality rather than in one.Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.- Treating hyperspectral as multispectral with more bands.
- Using all available bands without dimension reduction.
- Sizing a corpus without accounting for dimensionality.
- Annotating after a transformation-based reduction, which removes interpretability.
- Matching library signatures against poorly corrected data.
- Ignoring the unusable wavelength ranges.
- Expecting a class map from a process producing proportions.
- Neglecting the natural variability libraries do not capture.
- Using different correction levels between corpus and production.
- Designing a regular monitoring project without checking available revisit.
- Targeting fine spatial resolution, fine spectral resolution and wide coverage at once.
- Evaluating an unmixing with the usual segmentation metrics.
