Plantation species and age-class mapping
Commercial plantations of eucalyptus, acacia and pine look similar from the ground but diverge sharply in their seasonal reflectance trajectories. Multi-temporal Sentinel-2 time-series, particularly the red-edge bands, can separate species and rotation age-classes that standard NDVI misses entirely.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands; 20 m in red-edge bands B5, B6 and B7 (centred at 705, 740 and 783 nm). Five-day revisit at the equator with both satellites. Free archive from 2015. The red-edge bands are the primary reason Sentinel-2 outperforms Landsat for species discrimination.
- Planet SuperDove: 3 m resolution, daily revisit globally. Eight spectral bands including two red-edge channels. Useful for capturing fine crown texture and detecting within-stand age variation that 10 m imagery smooths over. Commercial licence required.
- Landsat 8/9 OLI: 30 m resolution, 16-day revisit per satellite (8-day combined). No red-edge bands, but a deep free archive back to 1972 (Landsat 5 TM onward) makes it the only option for reconstructing rotation histories spanning multiple harvest cycles.
- Maxar WorldView-3: 31 cm panchromatic, 1.24 m multispectral, with eight VNIR and eight SWIR bands. At this resolution individual crown geometry is resolvable, which helps separate species where canopy closure is high and spectral differences are small. Tasked commercially; expensive per km².
Why standard NDVI is the wrong tool for species work
NDVI compresses the visible and near-infrared into a single greenness number. Two stands with identical NDVI can be entirely different species at entirely different ages. The index was designed to track vegetation vigour, not identity. Using it alone to classify plantation species is roughly equivalent to identifying people by their height.
The red-edge region, roughly 700 to 750 nm, is where chlorophyll absorption gives way to the high reflectance plateau of the near-infrared. Different species reach peak red-edge reflectance at different canopy densities and at different points in their seasonal cycle. Sentinel-2 bands B5, B6 and B7 sit precisely in this transition zone. Published studies using these bands consistently report species classification accuracies 10 to 20 percentage points higher than approaches restricted to blue, green, red and broadband NIR.
Phenological trajectory as a species fingerprint
A single image date is rarely enough. The real discriminating power comes from stacking many acquisitions across a full year and treating the temporal profile of each pixel as its signature. Eucalyptus and acacia, for instance, may look spectrally similar in the dry season but diverge sharply after the first rains. Pine plantations in the southern hemisphere show a much flatter seasonal curve because conifers do not flush new foliage in a concentrated burst.
The practical approach is to build a dense time-series from Sentinel-2, filter out cloud-affected observations, and then extract phenological metrics: the timing and magnitude of the green-up peak, the rate of senescence, the minimum reflectance trough. These metrics become the feature vector fed into a classifier, typically a random forest or support vector machine trained on field-truth plots. The classifier is then applied across the full plantation estate.
Rotation age adds another layer. A two-year-old eucalyptus block has a more open canopy, higher soil background reflectance and a steeper red-edge slope than a six-year-old block of the same species. Tracking pixel trajectories across multiple Landsat or Sentinel-2 years allows age-class assignment even where no planting records exist, though accuracy degrades for older stands where canopy closure saturates the signal.
Where classification still fails
Honesty here matters. Species confusion is highest where canopy closure is similar across species and where stands are small relative to pixel size. At 10 m resolution, a narrow strip plantation or a recently replanted block of less than one hectare will be dominated by mixed-pixel effects. WorldView-3 at 1.24 m reduces this problem but introduces its own: the classifier must now handle intra-crown shadow and illumination variation that 10 m pixels average away.
Cloud cover is a persistent constraint in tropical regions where most fast-growing plantation species are managed. A site in Borneo or coastal Brazil may have fewer than 30 clean Sentinel-2 observations per year. Synthetic aperture radar can fill some temporal gaps, but SAR backscatter is not a good species discriminator on its own. It helps most as a complementary input to confirm canopy closure or detect recent harvest.
Field-truth plots are not optional. Remote sensing classification produces a probability surface, not ground truth. Without georeferenced field observations covering the range of species, ages and site conditions present, there is no way to calibrate the classifier or honestly report accuracy. Typical validation targets for published plantation mapping studies are overall accuracy above 85% and a kappa coefficient above 0.80, but these figures are only meaningful if the validation plots are independent of the training data.
Building the age-class map: what the workflow looks like
The input is a Sentinel-2 time-series, typically two to three years of imagery to capture at least one full phenological cycle and ideally one harvest event. Each scene is atmospherically corrected to surface reflectance using the Sen2Cor processor or equivalent. Cloud and cloud-shadow masks are applied, and a compositing step produces monthly or bi-monthly best-pixel mosaics.
From those mosaics, spectral indices are calculated: NDVI, the red-edge chlorophyll index (B7/B5 minus 1), the normalised difference red-edge index, and the SWIR-based moisture index. These are stacked with the raw band values into a multi-layer, multi-date feature cube. A random forest classifier trained on labelled field plots then assigns each pixel to a species-age class. The output is a raster map with an associated confusion matrix and per-class accuracy report.
The deliverable for a plantation manager is typically a GIS layer showing species blocks colour-coded by rotation age, updated annually or after each major harvest. For a government forestry agency the same layer feeds into timber volume projections and replanting schedules. Satellize has applied comparable multi-temporal classification pipelines in agricultural contexts, including the Tonga crop-estimation programme, and the spectral logic transfers directly to plantation forestry.
Sensor choice is a budget decision as much as a technical one
For most plantation estates above roughly 500 hectares, Sentinel-2 is the right starting point. It is free, the archive is deep, and the red-edge bands are genuinely useful. Planet SuperDove adds value where fine spatial detail matters or where the revisit frequency of Sentinel-2 is insufficient to capture a fast-moving phenological event. WorldView-3 is best reserved for small, high-value validation areas or for resolving ambiguous blocks that the coarser sensors cannot separate.
Landsat's value is almost entirely historical. If you need to reconstruct when a block was planted, what species replaced what, and how many rotations a stand has completed, Landsat is the only free archive that goes back far enough. The absence of red-edge bands is a real limitation for current-state mapping, but for temporal reconstruction it remains indispensable.
Typical figures
| Spatial resolution (primary sensor) | 10 m (Sentinel-2 visible/NIR), 20 m (Sentinel-2 red-edge B5–B7) |
| Spatial resolution (high-resolution option) | 1.24 m multispectral (WorldView-3); 3 m (Planet SuperDove) |
| Revisit frequency | 5 days (Sentinel-2 combined); daily (Planet SuperDove); 8 days (Landsat 8+9 combined) |
| Key spectral bands for species discrimination | Red-edge: Sentinel-2 B5 (705 nm), B6 (740 nm), B7 (783 nm); SWIR: B11 (1610 nm), B12 (2190 nm) |
| Minimum mappable stand size | ~1 ha at 10 m resolution; ~0.1 ha at 3 m resolution (subject to mixed-pixel effects at stand edges) |
| Archive depth | Sentinel-2 from 2015; Landsat from 1972 (TM); Planet from ~2017 for SuperDove |
| Typical classification accuracy (published studies) | 85–92% overall accuracy for 3–5 species classes with red-edge features and field truth; lower without red-edge |
| Cloud sensitivity | Optical only; tropical sites may yield fewer than 30 usable Sentinel-2 scenes per year |
| Delivery formats | GeoTIFF raster, vector polygon layer (GeoPackage or Shapefile), confusion matrix report, CSV age-class summary |
| Latency (map update cycle) | Annual or post-harvest; near-real-time species mapping is not the appropriate use case for this method |
Analytics Satellize can run
| Species classification map | Random forest or support vector machine classifier trained on multi-date Sentinel-2 red-edge and SWIR feature stack | GeoTIFF raster and polygon layer showing species identity per stand block, with per-class accuracy statistics |
| Rotation age-class map | Pixel-level temporal trajectory analysis across Sentinel-2 and Landsat archive; change-point detection to identify harvest and replanting events | GIS layer colour-coded by age class (e.g. 0–2, 2–4, 4–7 years); tabular summary of area by class |
| Phenological metrics layer | Harmonic regression or TIMESAT-style curve fitting on annual NDVI and red-edge index time-series | Raster stack of peak timing, amplitude and green-up rate per pixel; useful as classifier input or as standalone phenology report |
| Harvest detection and replanting confirmation | Breakpoint detection (e.g. BFAST or LandTrendr logic) applied to multi-year Sentinel-2 or Landsat time-series | Polygon layer of confirmed harvest events with estimated date; flag layer for blocks showing unexpected canopy loss |
| Accuracy assessment report | Stratified random sampling of validation plots; confusion matrix, overall accuracy, kappa coefficient, producer and user accuracy per class | PDF report with confusion matrix, accuracy statistics and map of validation plot locations |
| Annual estate inventory update | Repeat classification run on new season's imagery, differenced against prior year's map to flag changed blocks | Change polygon layer and tabular estate summary (species area, age distribution) delivered as GIS package |
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.