Individual urban tree detection and species classification
Satellite imagery at sub-metre resolution can delineate individual urban tree crowns, estimate their diameter, and, with hyperspectral data, classify species or functional type. Each method has a hard resolution floor and a candid set of failure modes.
Sensors
- Airbus Pléiades Neo: 30 cm panchromatic, 70 cm multispectral (4-band plus red-edge and deep-blue), daily revisit at mid-latitudes. The 30 cm pan band is the practical minimum for resolving crowns of street trees with diameters above roughly 2 metres. Tasked commercially; archive depth varies by city.
- Planet SuperDove (PlanetScope): 3 m multispectral, 8 spectral bands including red-edge and NIR, near-daily global revisit. Sufficient for canopy-cover fraction and phenological timing; individual crown delineation is unreliable below roughly 4–5 m crown diameter at this resolution. Useful for change detection between inventory epochs.
- ASI PRISMA: 30 m spatial resolution, 237 contiguous spectral bands from 400–2500 nm (VNIR-SWIR). Resolution precludes individual crown delineation in urban settings, but the hyperspectral signature enables species or functional-type classification when fused with a higher-resolution crown mask. Freely available for research under ESA/ASI data agreements.
- GEDI (Global Ecosystem Dynamics Investigation): Spaceborne full-waveform lidar aboard the ISS. 25 m footprint, non-contiguous sampling along ground tracks spaced roughly 600 m apart at the equator. Provides canopy height and vertical structure profiles, not wall-to-wall mapping. Useful for calibrating crown-height assumptions fed into optical delineation models.
What a floating roof gives away
A tree crown seen from directly above is a roughly circular, slightly domed surface. Its radius, texture and spectral response change with species, season and health state. At 30 cm resolution, a mature London plane with a 6 m crown spans roughly 20 pixels across its diameter. That is enough to fit an ellipse, compute a normalised difference vegetation index value, and separate it from the adjacent pavement. At 3 m resolution, the same crown occupies two pixels. The crown is still detectable as a vegetation object, but its boundary, and any neighbour it touches, cannot be resolved.
This is the central arithmetic of urban tree remote sensing. The 60–70 cm resolution threshold is not a marketing figure; it follows directly from the Nyquist-Shannon logic of spatial sampling. Below that threshold, crowns in closed-canopy streets merge into a single green blob. The blob has an area, but it does not have a count. Any inventory that needs individual stem counts, not just canopy fraction, requires imagery at or below 70 cm per pixel.
Delineating crowns: methods and honest failure modes
The standard computational approach is individual tree crown (ITC) delineation, most commonly implemented as a local-maximum detection step followed by either region-growing or watershed segmentation on a canopy height model or on the NIR reflectance surface. Published studies using airborne lidar at 10–25 cm resolution report commission and omission errors of 10–20 percent in open-grown park trees, rising to 30–50 percent in closed-canopy street canyons where crowns interlock. Satellite-based optical methods operating at 30–70 cm show broadly similar patterns, with errors skewing toward omission (missed trees) in dense rows.
Pléiades Neo at 30 cm panchromatic offers the best currently available spaceborne optical resolution for this task. Fusing the 30 cm pan with the 70 cm multispectral bands via pansharpening (Gram-Schmidt or BDSD methods are standard) produces a 30 cm 4-band or 6-band image that supports both delineation and a limited spectral characterisation. Red-edge reflectance at this scale is a useful proxy for chlorophyll content and, by extension, stress or senescence. It does not, on its own, identify species.
Deep learning approaches, specifically convolutional neural networks trained on labelled crown polygons from airborne surveys, have improved delineation accuracy in published benchmarks. A model trained on one city's tree stock, however, generalises poorly to another city's species mix and street geometry without retraining. This is not a solvable problem at the moment; it is a data-collection cost that buyers should budget for.
Species classification: what hyperspectral actually buys you
Broadleaf species differ in leaf water content, chlorophyll concentration, lignin and cellulose structure. These differences produce diagnostic absorption features across the 400–2500 nm range, most notably around 970 nm, 1200 nm and 1730 nm. A 4-band multispectral sensor captures none of these features directly. PRISMA's 237 contiguous bands capture all of them, but at 30 m resolution, a single pixel in an urban setting typically contains a mixture of canopy, road surface and building shadow.
The practical workflow is a fusion: use Pléiades Neo to delineate individual crowns as polygons, then extract the mean PRISMA spectrum for each polygon using the crown mask as a spatial filter. This works acceptably when the crown is large enough to contain at least one or two relatively pure PRISMA pixels, which in practice means crowns above roughly 8–10 m diameter. For smaller trees, spectral unmixing introduces uncertainty that is difficult to quantify without ground truth. Published classification accuracies for urban tree species using hyperspectral data fused with high-resolution crown masks range from around 70 percent to 90 percent overall accuracy, depending heavily on species richness and the quality of the training dataset.
Functional-type classification (deciduous broadleaf, evergreen broadleaf, conifer) is more reliable than species-level classification and is achievable with fewer training samples. For biodiversity reporting under frameworks such as the EU's Urban Greening Plan or national urban forest strategies, functional type may be sufficient.
Insurance and risk: what the inventory actually needs to answer
Root damage and storm-fall risk assessments require crown diameter, estimated height, species (as a proxy for root architecture and wood density) and proximity to structures. Satellite data can supply crown diameter directly from the delineated polygon, estimated height from GEDI profiles or from shadow-length geometry in very-high-resolution imagery, and a species or functional-type probability from hyperspectral fusion. What it cannot supply reliably is root zone extent below ground, structural defects, or the presence of decay fungi. Those require ground inspection.
The practical value of the satellite inventory is as a prioritisation tool. A city with 200,000 street trees cannot afford ground inspections of every stem. A satellite-derived inventory flags the 8,000 large-crowned trees within 5 m of a building, or the 3,000 trees showing NIR-reflectance anomalies consistent with stress, and directs ground crews accordingly. That triage function is where the economics close.
Seasonal timing and cloud: the constraints that do not go away
Deciduous trees must be imaged in full leaf to support spectral classification; the window in temperate cities is roughly May to September in the northern hemisphere. Within that window, cloud cover over many European and South-East Asian cities can reduce the number of usable acquisition days to single figures in a given month. Pléiades Neo's daily revisit is a theoretical maximum; cloud-free acquisitions over a specific city in a specific month may require waiting several weeks.
Planet SuperDove's near-daily revisit and large constellation size make it far more likely to capture a cloud-free image within a short window, but the 3 m resolution limits its role to change detection between inventory epochs rather than initial crown delineation. The practical architecture for most urban forest inventories is therefore a one-time or periodic Pléiades Neo acquisition for delineation, with SuperDove monitoring for canopy-cover change and stress signals between campaigns.
From pixels to a usable inventory
A deliverable urban tree inventory derived from satellite data typically contains, for each detected crown: a polygon geometry, a centroid coordinate, an estimated crown diameter, a height estimate (with confidence interval), a species or functional-type classification (with posterior probability), a health index derived from red-edge or NDVI, and a proximity-to-structure flag. These attributes are sufficient to populate a GIS-based urban forest management system and to meet the reporting requirements of most national urban forestry frameworks.
Satellize has run satellite-derived vegetation analytics in support of agricultural inventory, including the Kingdom of Tonga crop-estimation programme, and applies the same object-based image analysis and spectral-classification pipeline to urban tree inventories. The methodology transfers; the species training data does not, and that is the first conversation worth having with any city that wants a credible count rather than a canopy fraction.
Typical figures
| Best available spatial resolution (optical) | 30 cm panchromatic, 70 cm multispectral (Pléiades Neo) |
| Minimum detectable crown diameter | Approximately 2 m at 30 cm resolution; individual crowns merge below ~70 cm GSD |
| Revisit (tasked VHR) | Daily (Pléiades Neo, weather-permitting); cloud-free acquisition over a specific city may take days to weeks |
| Hyperspectral bands (PRISMA) | 237 contiguous bands, 400–2500 nm; 30 m spatial resolution |
| GEDI canopy height footprint | 25 m diameter; non-contiguous sampling, ~600 m track spacing at equator |
| Species classification accuracy (published range) | 70–90% overall accuracy; dependent on species richness and training-data quality |
| Archive depth (Pléiades/Pléiades Neo) | Pléiades 1A/1B from 2012; Neo from 2021; coverage varies by city tasking history |
| Delivery formats | GeoPackage or Shapefile polygon inventory, GeoTIFF crown-mask raster, CSV attribute table |
Analytics Satellize can run
| Individual crown polygon inventory | Local-maximum detection and watershed segmentation on NIR or pansharpened reflectance surface (object-based image analysis) | GIS polygon layer with crown diameter and centroid per detected tree |
| Crown height estimation | Shadow-length geometry from VHR imagery or GEDI waveform profile extraction | Attribute column in inventory layer; confidence interval per tree |
| Species or functional-type classification | Hyperspectral signature extraction (PRISMA) fused with Pléiades Neo crown mask; random forest or support vector machine classifier trained on ground-truth spectra | Species probability raster and classified polygon layer |
| Canopy health index | Red-edge chlorophyll index (CIre) and NDVI derived from Pléiades Neo or SuperDove red-edge band | Per-crown health score; flagged anomaly list for ground-inspection prioritisation |
| Proximity-to-structure risk flag | Spatial intersection of crown polygons with building footprint layer; buffer analysis at 3 m, 5 m and 10 m thresholds | Ranked shortlist of high-risk trees for insurance or maintenance triage |
| Canopy-change detection between epochs | Bitemporal NDVI differencing on SuperDove time series; crown-count comparison between Pléiades Neo campaigns | Change map showing new planting, removals and stress expansion; summary statistics report |
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.