Individual tree species identification from very-high-resolution imagery
Sub-metre multispectral and hyperspectral imagery can classify individual tree crowns to species or genus level, but only when collected at the right phenological moment. Sensor choice, acquisition timing and classification method together determine whether you get a map or noise.
Sensors
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral across eight VNIR bands plus eight SWIR bands. The SWIR bands (1195–2365 nm) add lignin and moisture sensitivity unavailable from four-band sensors. Revisit approximately 1 day at mid-latitudes under tasking.
- Airbus Pléiades Neo: 0.30 m panchromatic, 1.2 m native multispectral in six bands including red-edge and deep blue. Crown delineation at this resolution is reliable for crowns wider than roughly 3 m. Constellation of two satellites gives daily revisit at most latitudes.
- ASI PRISMA: Hyperspectral imager covering 400–2500 nm in 239 contiguous bands at 30 m spatial resolution. Too coarse for individual crown delineation in closed canopy, but foliar biochemical signatures retrieved from PRISMA can train and validate species classifiers applied to finer-resolution data.
- Planet SkySat: 0.50 m multispectral in five bands (blue, green, red, red-edge, NIR). Weaker spectral depth than WorldView-3 limits species separability in spectrally similar groups, but high tasking agility makes it practical for capturing narrow phenological windows on short notice.
- Airborne lidar (third-party acquisition): Not a satellite sensor, but point-cloud data at 5–50 pts/m² provides crown height and three-dimensional shape that sharply improves delineation of overlapping crowns before spectral classification. Satellize ingests client-supplied or commercially acquired lidar as a fusion layer.
Why a crown is more than a green blob
At 30 m resolution a forest pixel is a statistical average of dozens of crowns. At 0.3–0.5 m resolution each pixel belongs to a single crown, and the crown becomes an object with measurable shape, texture and spectral character. That shift from pixel to object is what makes species-level classification physically possible rather than merely aspirational.
The spectral fingerprint of a tree species is real but narrow. Chlorophyll absorption at 680 nm and the red-edge inflection between 700 and 740 nm vary by species in ways that are consistent within a phenological stage and a site, but not necessarily across sites or seasons. WorldView-3's eight VNIR bands resolve the red-edge at 1.24 m; its SWIR bands add sensitivity to cellulose and water content that helps separate conifers with similar green reflectance. Pléiades Neo's red-edge band at 1.2 m captures the same inflection. Without a red-edge channel, four-band sensors leave many broadleaf species statistically indistinguishable.
The phenological window is the hardest constraint
Deciduous species are most separable at two moments: leaf flush in spring, when anthocyanin concentrations and leaf area index rise at species-specific rates, and senescence in autumn, when chlorophyll breakdown exposes carotenoid and xanthophyll pigments that differ by species. Outside those windows, many broadleaf species converge on similar mid-summer reflectance and become nearly impossible to separate with multispectral data alone.
The practical implication is that a single acquisition planned for the wrong week can produce a classifier that performs at chance for half the target species. Cloud cover compounds the problem: a persistent overcast during a two-week senescence window can eliminate the opportunity entirely for that year. Hyperspectral data from PRISMA can extend the separable period slightly, because foliar biochemistry varies at finer spectral resolution than broadband sensors resolve, but PRISMA's 30 m pixel still cannot delineate individual crowns in closed forest. The honest answer is that species mapping at crown level often requires two acquisitions per year, not one.
Conifers are generally more stable across seasons because they retain foliage year-round, but separating species within genera such as Picea or Abies remains difficult with multispectral data. Hyperspectral reflectance in the SWIR, particularly the 1700–2200 nm range, improves conifer discrimination by capturing resin and wax chemistry, but this requires airborne hyperspectral or the SWIR bands of WorldView-3, not standard four- or six-band sensors.
From crown delineation to species label
Classification proceeds in two stages. First, individual tree crowns are delineated as objects. Watershed segmentation on a canopy height model derived from lidar or photogrammetric point cloud is the standard approach; without height data, segmentation on the panchromatic band alone produces irregular boundaries that bleed across crown edges and degrade subsequent spectral extraction. Second, spectral and textural features extracted per crown object are fed to a classifier.
Random forest and support vector machine classifiers trained on field-verified reference crowns remain the published baseline for this problem, with overall accuracies in the range of 70–90 % reported in peer-reviewed studies for forests with five to fifteen target species, falling toward the lower end as species count rises or training samples thin. Deep learning approaches, particularly convolutional neural networks applied to image patches centred on each crown, have matched or exceeded those figures in several published comparisons, but they require substantially more labelled training data and are more sensitive to domain shift when applied to new sites.
Accuracy figures from published studies should be read with care. Producer's accuracy for common species with abundant training samples routinely exceeds 90 %; for rare species with fewer than thirty reference crowns it can fall below 60 %. Any honest deliverable should report per-class accuracy alongside the overall figure.
What the method cannot do
Individual crown delineation breaks down in closed-canopy tropical forest where crown overlap is near-total and understorey species are invisible to any optical sensor. In those conditions the method maps the emergent and upper-canopy layer only, which may represent a minority of species diversity.
Spatial resolution sets a hard floor: crowns narrower than roughly twice the pixel size cannot be reliably delineated. At 1.24 m multispectral resolution, that means crowns below approximately 2.5 m diameter are effectively unresolvable as discrete objects. Young plantations and shrub-layer species fall below this threshold. Atmospheric correction quality also matters more at sub-metre resolution than at coarser scales, because small spectral differences between species can be swamped by adjacency effects and uncompensated aerosol scattering. Sites at high aerosol optical depth, common in parts of sub-Saharan Africa and South and Southeast Asia, require careful correction before any spectral classifier is applied.
Practical applications and where Satellize fits
The use cases with the clearest return on the cost of very-high-resolution tasking are those where species identity carries direct economic or regulatory value: timber inventory for high-value hardwood species, invasive species mapping (for example Ailanthus altissima in European urban forests or Eucalyptus in native South American woodland), protected-area biodiversity assessment, and urban tree-canopy management where species-specific mortality risk or carbon sequestration rates matter.
Satellize runs species classification workflows on WorldView-3 and Pléiades Neo imagery under client licence, combining object-based segmentation with red-edge and SWIR feature extraction and optional lidar fusion. The same analytical infrastructure that supports the Kingdom of Tonga crop-estimation programme handles the phenological scheduling and cloud-contingency retasking that species mapping requires. Outputs are delivered as labelled crown polygons in standard GIS formats with per-class confidence scores, not as raster classifications that obscure the uncertainty.
Typical figures
| Best spatial resolution (multispectral) | 1.2–1.24 m (WorldView-3, Pléiades Neo); 0.5 m (SkySat) |
| Panchromatic resolution for crown delineation | 0.30–0.31 m (WorldView-3, Pléiades Neo) |
| Hyperspectral resolution (PRISMA) | 30 m spatial; 239 bands across 400–2500 nm; spectral sampling ~10 nm |
| Revisit under tasking | ~1 day (WorldView-3 at mid-latitudes); ~1 day (Pléiades Neo constellation) |
| Minimum resolvable crown diameter | Approximately 2.5–3 m at 1.24 m multispectral resolution |
| Spectral bands relevant to species discrimination | Red-edge (700–740 nm), NIR (860 nm), SWIR (1195–2365 nm, WorldView-3 only) |
| Typical classification accuracy (5–15 species, closed-canopy temperate) | 70–90 % overall accuracy (published range); per-class accuracy varies widely |
| Archive depth | WorldView-3 archive from 2014; Pléiades Neo from 2021 |
| Delivery format | Labelled crown polygon GeoPackage or Shapefile with per-class confidence scores; optional GeoTIFF classified raster |
| Cloud sensitivity | Optical only; cloud cover during the phenological window can eliminate the acquisition opportunity for that season |
Analytics Satellize can run
| Individual crown species map | Object-based image analysis with random forest or CNN classifier trained on field-verified reference crowns; red-edge and SWIR feature extraction from WorldView-3 or Pléiades Neo | GIS polygon layer of delineated crowns with species or genus label and per-class confidence score |
| Phenological acquisition schedule | Time-series analysis of historical Sentinel-2 NDVI and red-edge chlorophyll index to identify leaf-flush and senescence windows at the target site | Recommended tasking calendar with cloud-risk assessment and contingency dates |
| Invasive species distribution map | Targeted binary classifier (invasive vs. native) using spectral and crown-shape features; validated against known occurrence records | Polygon layer of detected invasive crowns with area statistics and change comparison if multi-date imagery is available |
| Per-species crown density and basal-area estimate | Crown area from segmented polygons combined with published allometric relationships for height-to-diameter; lidar-derived height applied where available | Tabular summary of stem count, crown area and estimated basal area by species per defined management unit |
| Classification uncertainty layer | Posterior probability output from classifier; crowns below confidence threshold flagged for field verification | GIS layer with uncertainty band per crown, prioritised field-check list |
| Multi-date species change detection | Crown-matched comparison across two or more classified acquisitions to detect species composition shift, mortality or recruitment | Change polygon layer with transition matrix (species A to species B, or crown lost) and area statistics |
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.