Forest connectivity and seed-dispersal corridor identification from canopy structure

Canopy height models and high-resolution tree-cover maps let ecologists quantify forest connectivity using graph-theory metrics, revealing which gaps isolate patches and where restoration planting restores dispersal pathways.

Sensors

NASA GEDI (Global Ecosystem Dynamics Investigation): Full-waveform lidar on the ISS. Footprints of approximately 25 m diameter on a non-systematic grid with roughly 600 m cross-track spacing at the equator. Delivers canopy height (RH95, RH98 metrics) and vertical structure. Latitudinal coverage limited to 51.6° N/S. Not a wall-to-wall raster; requires interpolation or fusion with optical data to produce continuous canopy height mosaics.
NASA/USGS ICESat-2 ATL08: Photon-counting lidar. ATL08 product reports canopy height at 100 m along-track segments. Denser ground-track repeat than GEDI and extends to higher latitudes, making it useful for temperate and boreal corridor work. Complements GEDI by filling spatial gaps in the height record.
Planet SuperDove: Eight-band multispectral imagery at 3 m native resolution with near-daily revisit over most land areas. Enables high-resolution tree-cover mapping and gap-width measurement at scales relevant to published dispersal-distance thresholds for tropical species. Commercial tasking; not freely available.
Sentinel-2 MSI: Thirteen spectral bands, 10 m resolution in visible and near-infrared, 5-day revisit at the equator with two satellites. Freely available. Suitable for landscape-scale canopy-cover mapping and NDVI-based patch delineation. Gap widths below roughly 20 m are not reliably resolved.

What a canopy gap actually costs a seed

Landscape connectivity is not a metaphor. For a seed dispersed by a frugivorous bird or bat, a gap wider than the animal's willingness to cross open ground is a hard barrier. Published dispersal-distance models for tropical forest specialists set that threshold variously at 50 m to 400 m depending on species, with many obligate forest interior birds reluctant to cross gaps exceeding 100 m. Wind-dispersed species have different kernels, often modelled as negative exponential or inverse power-law functions fitted to seed-trap data. The point is that the geometry of the canopy, not just its total area, determines whether populations remain genetically connected.

Graph-theory provides the formal language. Forest patches become nodes; potential dispersal links between patches become edges, weighted by gap width or least-cost path through the intervening matrix. Metrics such as the Probability of Connectivity index and Betweenness Centrality identify which patches and corridors disproportionately maintain network integrity. Remove a high-betweenness stepping-stone patch and the network may fragment into isolated sub-graphs. These calculations are well established in the conservation-planning literature and are directly computable from satellite-derived canopy maps, provided the underlying spatial data are accurate enough.

Building the canopy height mosaic

GEDI's RH95 metric (the height below which 95% of returned energy falls) is the standard proxy for top-of-canopy height in tropical and subtropical forests. The instrument has accumulated tens of millions of footprints since its 2019 deployment, but those footprints do not tile the landscape uniformly. Producing a continuous canopy height raster requires regression or machine-learning fusion of GEDI heights against co-located spectral predictors from Sentinel-2 or Planet imagery. Studies using this approach report root-mean-square errors of roughly 3 to 6 m in closed-canopy tropical forest, rising in structurally complex or fragmented terrain.

ICESat-2 ATL08 adds independent height samples, particularly valuable where GEDI coverage is thin. The two lidar datasets share no sensor design and have different sensitivity to canopy density, so combining them requires careful cross-calibration. The honest position: a fused canopy height mosaic at 10 m resolution is achievable and useful, but it is a modelled surface, not a direct measurement, and uncertainty is higher at forest edges, precisely where gap-width estimates matter most.

Measuring gaps at the resolution that matters

A 10 m Sentinel-2 pixel can detect a gap only if that gap spans at least two or three pixels with confidence. For the 50 m to 100 m gap-width thresholds relevant to many dispersal models, Sentinel-2 is adequate at the landscape scale. For species with narrower crossing tolerances, or for identifying stepping-stone patches smaller than a hectare, 3 m Planet SuperDove imagery changes the analysis materially.

Tree-cover maps derived from Planet imagery use supervised classification against canopy height model training labels, typically producing binary canopy or non-canopy layers at 3 m. Gap width is then measured as the Euclidean or least-cost distance across non-canopy pixels between adjacent patches. The practical floor: gaps narrower than roughly 6 to 9 m (two to three Planet pixels) are unreliably classified, and isolated trees within a gap, which may function as perches and stepping stones for birds, are often missed at any of these resolutions. Field validation remains necessary for species-specific corridor design.

From connectivity map to restoration priority

Once patch nodes and edge weights are assembled, Betweenness Centrality scores identify the corridors whose loss would most increase average path length across the network. Patches with high centrality but small area are the most cost-effective restoration targets: adding canopy there reconnects the largest number of previously isolated patch pairs.

Least-cost path modelling extends this by routing proposed planting corridors through the matrix at minimum cost, where cost is a function of land-cover type, slope and distance from existing forest edge. The output is a ranked list of planting zones with estimated connectivity gain per hectare restored. Restoration planners can then overlay land-tenure data and cost-of-land estimates to produce an investment-ranked corridor map. The connectivity gain is only as reliable as the dispersal kernel assumed. If the kernel is poorly constrained, sensitivity analysis across a plausible range of kernel parameters is standard practice and should be reported alongside any single-scenario result.

Where the method fails, and what to do about it

Three limits deserve explicit acknowledgement. First, cloud cover. Sentinel-2 and Planet imagery over humid tropical forests can be obscured for weeks or months at a time. Annual composites reduce this problem but lose seasonal information about canopy phenology that affects dispersal behaviour. Second, the dispersal kernel. Published kernels exist for perhaps a few dozen tropical tree species and a similar number of vertebrate dispersers. For most species assemblages, the kernel must be assumed or borrowed from a related taxon, introducing uncertainty that no amount of satellite data can resolve. Third, matrix permeability. Graph-theory models often treat non-forest land as a uniform barrier, but a shaded coffee plantation and a concrete road surface are not equivalent obstacles. Incorporating matrix quality requires additional land-cover layers and expert parameterisation.

Satellize's corridor analytics run on open Sentinel-2 archives combined with GEDI and ICESat-2 public datasets, with Planet SuperDove added on client licence for gap-width work below 20 m. The workflow draws on the same fusion methods applied in the Tonga crop-estimation programme, adapted to canopy rather than agricultural structure. Outputs are delivered as GIS layers with per-patch connectivity scores, a ranked restoration-priority table and a sensitivity report showing how results shift across dispersal-kernel assumptions.

What the output looks like in practice

A corridor identification run for a landscape of, say, 500 000 hectares produces a patch network graph with nodes attributed by area, mean canopy height and centrality score, plus a raster of least-cost corridor routes at the resolution of the input tree-cover map. Planners receive this as a GeoPackage or cloud-optimised GeoTIFF alongside a PDF summary showing the ten highest-priority corridor segments, their estimated gap widths, the dispersal-kernel scenarios tested and the confidence intervals on connectivity gain.

The map is a planning input, not a biological guarantee. Corridor design based on canopy structure identifies where connectivity could exist if the right species are present and the matrix is permeable enough. Whether dispersal actually occurs requires monitoring, ideally through seed-trap networks or genetic sampling over subsequent years. The satellite layer tells you where to look and where to plant. It does not replace the ecologist in the field.

Typical figures

Canopy height accuracy (GEDI RH95, closed tropical forest)	RMSE approximately 3 to 6 m against airborne lidar reference; higher at forest edges
Spatial resolution, tree-cover map	3 m (Planet SuperDove) or 10 m (Sentinel-2); gap widths below 6 m not reliably resolved at 3 m
GEDI footprint diameter	~25 m; cross-track spacing ~600 m at equator; coverage 51.6° N/S only
ICESat-2 ATL08 along-track segment length	100 m; extends to polar latitudes unlike GEDI
Sentinel-2 revisit	5 days at equator (two-satellite constellation); cloud cover limits usable acquisitions in humid tropics
Planet SuperDove revisit	Near-daily over most land; commercial licence required
Minimum detectable isolated patch	~0.09 ha at 3 m resolution; ~1 ha at 10 m resolution (practical threshold, not theoretical)
Archive depth	Sentinel-2 from 2015; GEDI from April 2019; Planet from 2016 (varies by region)
Connectivity metric latency	Days to weeks from data acquisition, depending on cloud compositing window required
Delivery format	GeoPackage, cloud-optimised GeoTIFF, PDF summary report with sensitivity analysis

Analytics Satellize can run

Continuous canopy height mosaic	Regression fusion of GEDI RH95 and ICESat-2 ATL08 heights against Sentinel-2 spectral predictors using random forest or gradient boosting	10 m raster GeoTIFF with per-pixel height and uncertainty estimate
Binary tree-cover map	Supervised classification of Planet SuperDove or Sentinel-2 imagery using canopy height model as training label; object-based post-processing to clean patch boundaries	3 m or 10 m GeoTIFF, canopy or non-canopy, with classification confidence layer
Patch network graph with connectivity metrics	Graph-theory analysis (Probability of Connectivity index, Betweenness Centrality) computed from patch adjacency and gap-width edge weights using published dispersal-distance thresholds	GeoPackage with attributed patch polygons and edge table; ranked centrality scores
Least-cost corridor routes	Least-cost path modelling through land-cover resistance surface derived from Sentinel-2 classification; multiple dispersal-kernel scenarios tested	Raster corridor map and vector route lines with per-segment cost and estimated connectivity gain
Restoration priority ranking	Connectivity gain per hectare restored, computed by iterative node addition to the patch graph at candidate planting locations	Tabular priority list with spatial coordinates, estimated area, connectivity gain and kernel-sensitivity scores
Dispersal-kernel sensitivity report	Monte Carlo or scenario-based sensitivity analysis across plausible kernel parameter ranges drawn from published species literature	PDF report showing how corridor rankings shift under low, median and high dispersal-distance assumptions

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.