Forest type and floristic composition mapping with hyperspectral data
Contiguous narrow spectral bands from 400–2500 nm reveal leaf biochemistry that broadband sensors cannot resolve, enabling forest-type mapping at the level of dominant species or functional group. PRISMA, DESIS, EMIT and EnMAP make this operational from orbit.
Sensors
- PRISMA (ASI, Italy): 239 contiguous bands from 400–2500 nm at 30 m ground sampling distance; hyperspectral swath 30 km; revisit approximately 29 days at nadir but programmable tasking shortens this. Primary operational spaceborne hyperspectral sensor for vegetation biochemistry.
- DESIS (DLR, aboard ISS): 400–1000 nm in up to 235 bands at 30 m GSD; ISS orbital inclination of 51.6° restricts coverage to latitudes between roughly 52°N and 52°S. Visible and near-infrared only, so misses shortwave-infrared cellulose and lignin features, but excellent for chlorophyll and carotenoid discrimination.
- EMIT (NASA, aboard ISS): 380–2500 nm in 285 bands at 60 m GSD; designed primarily for mineral dust mapping but vegetation absorption features are fully captured. Publicly archived via NASA Earthdata. Same ISS latitude constraint as DESIS applies.
- EnMAP (DLR, Germany): 420–2450 nm in 244 bands at 30 m GSD; dedicated Earth observation satellite launched April 2022; 30 km swath; revisit 27 days at nadir, 4 days off-nadir. Signal-to-noise ratio specified above 400:1 in VNIR, making it well-suited to subtle foliar absorption features.
What a narrow band reveals that a broad band cannot
A Landsat or Sentinel-2 pixel integrates reflected energy across bands tens to hundreds of nanometres wide. That averaging erases the narrow absorption features that identify specific leaf compounds: chlorophyll a and b absorb at 430 nm and 662 nm; carotenoids peak near 490 nm; liquid water has distinct troughs at 970, 1200 and 1450 nm; lignin and cellulose leave fingerprints across 1700–2300 nm. When a sensor samples these regions in bands 10–15 nm wide and contiguous, each pixel becomes a partial biochemical profile of the canopy above it.
This matters for forest typing because species differ in the concentrations of these compounds, not just in their gross greenness. A broadband NDVI will not separate a dipterocarp forest from a peat-swamp forest at the same canopy density. A hyperspectral classifier often can, because the two communities differ in foliar nitrogen, water content and secondary compounds in ways that leave measurable spectral traces. The distinction is not guaranteed, and accuracy degrades when canopy structure rather than leaf chemistry drives the signal, but the information content is categorically higher.
Where the method works well and where it struggles
Structurally diverse tropical forests are the strongest use case. A 2021 study using airborne hyperspectral data in Borneo demonstrated species-level classification accuracies above 80% for dominant canopy trees, a result that broadband methods could not approach at equivalent spatial scales. Spaceborne sensors at 30–60 m GSD cannot match airborne data at 1–5 m, but they can reliably separate forest functional types: evergreen broadleaf, deciduous broadleaf, needle-leaf, and mixed, as well as floristic groups defined by shared biochemical traits.
The method has real limits. Cloud cover is the bluntest one: optical hyperspectral data is useless through cloud, and humid tropical forests, where the technique is most valuable, can have cloud-free fractions below 20% in any given month. A single PRISMA acquisition may need to wait weeks for a clear scene over a target. Atmospheric correction is also more demanding than for broadband sensors because narrow bands amplify the effect of residual aerosol and water-vapour errors. Canopy shadows, mixed pixels at stand boundaries, and phenological timing all introduce ambiguity. Accuracy claims above 85% in the peer-reviewed literature almost always come from airborne campaigns; spaceborne results at 30 m are typically 10–20 percentage points lower for fine-grained species separation, though functional-group classification holds up better.
Spectral indices and radiative transfer: the two analytical routes
Analysts generally take one of two paths. The first is empirical: compute published spectral indices (red-edge chlorophyll index, carotenoid reflectance index, normalised difference lignin index and others), then train a classifier on field-reference plots. Random forest and support vector machine classifiers are the workhorses here. This approach is fast and interpretable, but it is site-specific; a model trained in the Congo Basin does not transfer cleanly to the Amazon without recalibration.
The second path uses physically-based radiative transfer models, most commonly PROSAIL, which couples the PROSPECT leaf model with the SAIL canopy reflectance model. By inverting PROSAIL against measured spectra, analysts can retrieve leaf chlorophyll content, leaf area index, equivalent water thickness and dry matter content as continuous variables rather than class labels. These retrievals are more transferable across sites because they are grounded in physics, but inversion is computationally expensive and requires good atmospheric correction as a prerequisite. For national-scale forest-type mapping, the empirical route is usually faster; for biochemical parameter retrieval intended to feed carbon or nutrient models, radiative transfer inversion is more defensible.
Combining spaceborne hyperspectral with the broader data stack
No single sensor solves the problem alone. PRISMA or EnMAP provides the spectral dimension; Sentinel-2 or Landsat provides the temporal depth needed to account for phenology and to fill cloud gaps with historical context. Sentinel-1 SAR adds canopy structure information that is spectrally invisible. Where GEDI lidar footprints overlap the study area, canopy height can be used to stratify the classification, separating tall closed-canopy forest from shorter secondary growth before applying the hyperspectral classifier within each stratum.
This fusion is where most of the practical accuracy gains come from. A hyperspectral classifier that struggles to separate two forest types of similar biochemistry but different structure can resolve the ambiguity once a height layer is introduced. The data integration is not trivial: co-registration between sensors with different acquisition geometries, resolutions and timing requires careful preprocessing, and the analyst must be explicit about which sensor is driving which part of the classification.
From spectral map to operational product
A forest-type map derived from hyperspectral data is most useful when it is anchored to a national forest inventory or a biodiversity monitoring framework. The map provides spatial coverage that field surveys cannot; the field surveys provide the ground truth that validates and calibrates the spectral classification. Without that loop, a hyperspectral-derived map is a research output, not an operational product.
Satellize runs this integration workflow for government clients, combining open hyperspectral archives with commercial tasking through PRISMA or EnMAP where higher revisit or specific scene timing is required. The Tonga crop-estimation programme demonstrated the general principle of fusing spectral analytics with field reference data at national scale; the same pipeline applies to forest-type mapping, with the classification targets and validation protocols adjusted for forestry rather than agriculture. Deliverables are GIS layers with per-class confidence scores and a documented accuracy assessment, not a map alone.
What to ask before commissioning a hyperspectral forest survey
Four questions determine whether hyperspectral data will improve on what you already have. First, what is the target discrimination: forest versus non-forest, broad functional type, or dominant species? The last is the hardest and requires the most field reference data. Second, what is the cloud climatology over your study area, and can the acquisition window be constrained to a dry season? Third, do you have existing field plots with species records that can serve as training and validation data, or does the programme need to generate them? Fourth, what will the map be used for? A map feeding a REDD+ carbon account needs a different accuracy standard and audit trail than a map informing a national biodiversity strategy.
Hyperspectral mapping is not cheap in analyst time, even when the imagery is free through open archives. The preprocessing chain alone, atmospheric correction, smile correction, bad-band removal, co-registration, is several days of skilled work per scene. Budget accordingly, and be sceptical of any proposal that does not include an independent accuracy assessment against held-out field data.
Typical figures
| Spatial resolution (spaceborne) | 30 m GSD (PRISMA, DESIS, EnMAP); 60 m GSD (EMIT) |
| Spectral range | 400–2500 nm contiguous (PRISMA, EnMAP, EMIT); 400–1000 nm (DESIS) |
| Number of bands | 235–285 contiguous bands depending on sensor; band width typically 10–12 nm |
| Revisit (nadir) | 27–29 days nadir; 4–7 days off-nadir tasking for PRISMA and EnMAP |
| Swath width | 30 km (PRISMA, EnMAP, DESIS); approximately 75 km (EMIT) |
| Latitude coverage | PRISMA and EnMAP: near-global; DESIS and EMIT: 52°N to 52°S (ISS orbit) |
| Minimum discriminable unit | Forest functional group reliably at 30 m; dominant-species separation requires field calibration and is site-dependent |
| Cloud sensitivity | Fully cloud-limited; no signal through cloud or dense smoke |
| Archive depth | PRISMA: from 2019; DESIS: from 2018; EMIT: from 2022; EnMAP: from April 2022 |
| Delivery formats | GeoTIFF classification layers, per-class probability rasters, accuracy assessment report, field-validated legend |
Analytics Satellize can run
| Forest functional-type map | Supervised classification (random forest or SVM) on hyperspectral spectral indices; trained on field reference plots | GeoTIFF raster with per-class confidence scores and confusion matrix; suitable for GIS import |
| Dominant-species probability layer | Spectral mixture analysis combined with red-edge and SWIR absorption indices; validated against national forest inventory plots | Per-species probability raster with documented accuracy assessment; PDF summary report |
| Leaf chlorophyll content map | PROSAIL radiative transfer model inversion against atmospherically corrected PRISMA or EnMAP reflectance | Continuous raster in µg/cm²; uncertainty layer included |
| Canopy equivalent water thickness retrieval | Inversion of PROSPECT-5 leaf model using 970 nm and 1200 nm water absorption features | Continuous raster (g/cm²); flagged where shadow fraction exceeds threshold |
| Floristic change detection between epochs | Bitemporal comparison of classified hyperspectral scenes; change pixels flagged by spectral distance from baseline | Change polygon layer with class-transition matrix; annual or campaign-by-campaign cadence |
| Hyperspectral-SAR fusion forest-type map | Feature-level fusion of PRISMA spectral indices with Sentinel-1 backscatter and texture metrics; stratified by GEDI canopy height where available | Fused classification GeoTIFF; accuracy assessment comparing fusion result against hyperspectral-only baseline |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.