Plantation species classification for carbon stock differentiation

Monoculture plantations of eucalyptus, acacia, pine and teak carry substantially different carbon stocks, yet most MRV protocols collapse them into one class. Hyperspectral imaging and multitemporal multispectral analysis can separate species at stand level, provided the analyst is honest about where the methods fail.

Sensors

ASI PRISMA hyperspectral: 239 contiguous bands from 400 to 2500 nm at 30 m spatial resolution and roughly 29-day revisit per target. The narrow, contiguous spectral sampling resolves leaf-chemistry features (lignin at ~1680 nm, cellulose at ~2100 nm, chlorophyll red-edge position) that broadband sensors cannot separate, making it the primary instrument for species-level discrimination where archive coverage exists.
ESA Sentinel-2 MSI: 13 spectral bands at 10–60 m resolution with a 5-day revisit at the equator (combined Sentinel-2A and -2B). The four red-edge bands (705, 740, 783, 865 nm) and the SWIR pair carry phenological differentiation power across seasons. A multitemporal stack of 12–24 cloud-free composites is the practical workhorse for large-area species mapping where PRISMA coverage is absent.
Planet SuperDove: 8 spectral bands including two red-edge channels, 3–5 m spatial resolution, and near-daily revisit globally. The fine spatial resolution resolves stand boundaries and within-stand variability that 10 m imagery blurs, though the limited spectral depth means classification relies more on phenological timing than on leaf-chemistry features.
Airbus Pléiades Neo: 4 multispectral bands plus a deep-blue channel at 30 cm native resolution, tasked on demand. Used here not for species classification per se but for stand delineation, canopy gap fraction measurement, and ground-truth polygon extraction to train classifiers on coarser sensors.

Why species identity is a carbon-accounting problem, not just a forestry one

A mature eucalyptus stand in Brazil can accumulate above-ground biomass at 15–25 tonnes of dry matter per hectare per year under favourable conditions, while a teak plantation of the same age in Southeast Asia typically grows at 5–10 tonnes per hectare per year. Acacia mangium sits somewhere between the two. Pine species vary widely by provenance and climate. These are not marginal differences: they translate directly into carbon credits issued, and conflating species within a project boundary can overstate or understate sequestration by a factor of two or more.

Most Verified Carbon Standard and Gold Standard methodologies permit the use of species-specific allometric equations when species identity is documented, but they do not mandate remote-sensing verification of that identity. The practical consequence is that project developers can self-report species composition and auditors rarely have the tools to check. Satellite-based species classification is not a replacement for field inventory; it is an independent check on what the field inventory claims.

What a spectrum gives away that a photograph does not

Broadband RGB or even standard four-band multispectral imagery shows canopy texture and gross colour. That is often enough to separate conifers from broadleaves, but not enough to separate eucalyptus from acacia when both are young, or teak from other deciduous broadleaves mid-season. Hyperspectral data changes the problem. PRISMA's 239 bands allow the analyst to compute the red-edge chlorophyll index, the normalised difference lignin index, the cellulose absorption index and dozens of other leaf-chemistry proxies that differ systematically between species.

Published studies using airborne hyperspectral data have achieved overall accuracies of 85–95% separating four to six plantation species where training samples are adequate and stands are reasonably pure. PRISMA at 30 m ground sampling distance is coarser than most airborne campaigns, which introduces mixed-pixel noise at stand edges. The honest expectation for a PRISMA-based classification across a large project area is an overall accuracy in the range of 75–88%, depending on species pair, stand age and seasonal timing of acquisition. That range is wide because eucalyptus and acacia, in particular, share similar red-edge positions when both are in full leaf flush.

Phenology as a second discriminant: what the calendar reveals

Teak (Tectona grandis) is strongly deciduous: it drops leaves between December and March across most of its planted range, producing a distinctive NDVI trough visible in Sentinel-2 time series. Eucalyptus is evergreen. Acacia mangium is also largely evergreen but shows a subtle leaf-flush pulse in SWIR reflectance during new-growth periods. Pine canopies are dark and structurally consistent year-round but show species-level differences in the NIR plateau.

A multitemporal Sentinel-2 stack, processed to surface reflectance using the Sen2Cor atmospheric correction chain and composited into monthly medians to suppress cloud, produces a time series in which these phenological signatures become separable features. Random forest and support vector machine classifiers trained on this stack consistently outperform single-date classifications by 8–15 percentage points in published benchmarks. The practical requirement is at least 12 months of cloud-acceptable imagery, which is achievable in most tropical plantation regions but can be difficult in persistently cloudy highland zones.

Where the method fails: confusions the analyst must report

Two confusions recur in the literature and deserve explicit disclosure in any MRV deliverable. First, eucalyptus and acacia at ages two to four years share canopy height, leaf area index and spectral response closely enough that even hyperspectral classifiers struggle, with reported producer accuracies for acacia dropping to 60–70% in young stands. Second, mixed or transitional blocks where one species has been replanted into a stand of another create spectral mixtures that no pixel-level classifier handles well.

Cloud cover is a structural constraint in humid tropical regions. A single PRISMA acquisition over a target in Borneo or Central Africa may be cloud-free less than 30% of the time, meaning the hyperspectral observation that anchors the classification may be months old relative to the multispectral time series. Analysts should report the acquisition date of each hyperspectral scene used, the cloud fraction at that date, and the confidence interval on species assignment for each stand polygon. Presenting a single species map without these caveats is not defensible in a third-party audit.

From species map to carbon number: the allometric chain

Once stands are assigned species labels with associated confidence scores, species-specific allometric equations from the published literature convert stand-level basal area or canopy height estimates into above-ground biomass. For eucalyptus, the equations developed by Soares and Tomé (2012) for planted stands are widely cited. For teak, FAO-referenced equations by Kollert and Cherubini provide a starting point, though regional calibration improves accuracy. Acacia mangium equations from Malaysian and Indonesian forestry research are available in peer-reviewed form.

The carbon stock uncertainty budget has three additive components: classification error (species misassignment), allometric model error (typically ±15–25% for individual trees, smaller at stand level), and canopy structure estimation error from the satellite-derived height or basal area proxy. A properly constructed MRV report propagates all three. Satellize's analytics pipeline for the Tonga crop-estimation programme uses a comparable uncertainty-propagation approach, adapted here to biomass rather than yield estimation.

Putting the classification into an audit-ready package

Registry-grade use requires that the species classification be reproducible from archived inputs. That means preserving the raw sensor data (PRISMA L2D surface reflectance, Sentinel-2 L2A), the training sample provenance (GPS-logged field plots or Pléiades Neo-derived polygons), the classifier version and hyperparameters, and the validation confusion matrix. The output layer should carry per-polygon confidence scores, not just a hard species label.

Revisit the classification at each project verification period. Species composition can change through replanting, mortality or species substitution, and a static map from year one is not valid evidence in year five. A practical schedule is an annual Sentinel-2 phenological update with a PRISMA hyperspectral acquisition every two to three years, or whenever a significant replanting event is reported. That cadence is affordable and produces the longitudinal evidence chain that third-party verifiers increasingly expect.

Typical figures

Hyperspectral spatial resolution	30 m (PRISMA); 7.5 m planned for PRISMA follow-on SHALOM
Hyperspectral spectral range	400–2500 nm, 239 bands, ~10 nm sampling (PRISMA)
Multispectral spatial resolution	10 m (Sentinel-2 VIS/NIR), 20 m (red-edge, SWIR); 3–5 m (Planet SuperDove)
Revisit cadence	5 days (Sentinel-2A+B combined); ~29 days (PRISMA per target); near-daily (Planet SuperDove)
Minimum classifiable stand size	~0.5 ha at 10 m multispectral; ~1–2 ha at 30 m hyperspectral (edge-mixing constraint)
Typical species classification accuracy	75–88% overall (PRISMA-based, 4 species); 80–92% (multitemporal Sentinel-2, mature stands)
Archive depth	Sentinel-2: from 2015; PRISMA: from 2019; Planet SuperDove: from 2021 (8-band)
Atmospheric correction	Sen2Cor (Sentinel-2); PRISMA L2D product (ASI-supplied); 6S or ACOLITE for cross-sensor consistency
Delivery formats	GeoTIFF species-class raster with confidence band; GeoPackage stand polygons; CSV uncertainty budget; PDF audit narrative

Analytics Satellize can run

Stand-level species classification map	Random forest or support vector machine trained on PRISMA leaf-chemistry indices and Sentinel-2 multitemporal phenological features; validated against held-out field plots	GeoTIFF raster and GeoPackage polygon layer with species label, confidence score and acquisition metadata per stand
Phenological time-series stack	Monthly median compositing of Sentinel-2 L2A surface reflectance; NDVI, red-edge chlorophyll index and SWIR-based indices extracted per stand polygon	CSV time series per stand polygon; visualisation report showing species-discriminating phenological curves
Species-specific above-ground biomass estimate	Published allometric equations applied per species class; uncertainty propagated from classification confidence and allometric model error ranges	Stand-level biomass table (tonnes dry matter per hectare) with 90% confidence intervals; aggregated project-area carbon stock estimate in tCO₂e
Classification change detection between verification periods	Bitemporal comparison of annual species maps; change polygons flagged for replanting, mortality or species substitution	Change polygon GIS layer with area statistics; narrative section for inclusion in periodic verification report
Confusion-pair uncertainty report	Per-class producer and user accuracy from validation confusion matrix; identification of high-confusion species pairs (e.g. eucalyptus vs. acacia in young stands)	PDF table of confusion matrix and per-species accuracy metrics; recommended field-verification priority list for low-confidence stands
Audit-ready evidence package	Archival of input scenes, training samples, classifier parameters and validation outputs in reproducible pipeline; formatted to Verified Carbon Standard MRV documentation requirements	Zipped archive of raw inputs, processing scripts, output layers and signed chain-of-custody log; human-readable audit narrative PDF

Sources

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.