Sensors
- Sentinel-2 MSI: 10 m resolution in four visible and near-infrared bands; 5-day revisit at the equator with two satellites. The standard workhorse for forest/non-forest classification at landscape scale. Free and globally archived from 2015.
- Planet SuperDove: 3 m resolution, eight spectral bands including red-edge, near-daily revisit. Resolves small forest patches and narrow riparian corridors that are below Sentinel-2's delineation threshold. Commercial licence required.
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral. Suitable for validating patch boundaries and measuring edge-zone canopy structure in high-priority areas. Tasked on demand; archive coverage is patchy outside frequently imaged regions.
- Landsat 8/9 OLI: 30 m resolution, 16-day revisit per satellite. Coarser than Sentinel-2 for patch delineation but provides a consistent archive back to 1984, enabling long-run fragmentation trend analysis across decades.
What aggregate area statistics quietly ignore
A country can report stable forest cover while its remaining forest is being carved into progressively smaller, more isolated patches. Each new road or clearing adds edge. Edge is not neutral territory: the zone extending roughly 100 to 300 metres inward from a forest boundary experiences elevated wind exposure, desiccation, increased light penetration and higher rates of tree mortality. Published field studies in Amazonia and Borneo have documented measurable changes in microclimate and tree-fall rates across this entire depth. A landscape composed of many small patches can therefore hold the same nominal area as one large block while delivering a fraction of the interior habitat and carbon-stock stability.
Fragmentation metrics make this visible. Patch size distribution, edge density (metres of edge per hectare of forest), and core-area index (the proportion of each patch that lies more than a threshold distance from any edge) are the three quantities that convert a classified binary map into a landscape-level diagnosis. None of them appear in a simple area count.
From classified pixels to patch geometry
The analytical pipeline has two stages. First, a forest/non-forest classification is produced from multispectral imagery, typically using a combination of normalised difference vegetation index thresholding, random-forest classifiers trained on spectral and textural features, and cloud-mask compositing. At 10 m (Sentinel-2), patches smaller than roughly 0.1 ha can be delineated reliably; at 3 m (SuperDove), that floor drops to around 0.01 ha. Below those thresholds, sub-pixel mixing makes boundary placement ambiguous, and any metric derived from it should be treated with caution.
Second, the binary raster is converted to a patch mosaic and landscape metrics are computed. The standard toolset follows the framework established in FRAGSTATS (McGarigal and Marks, 1995), which remains the reference implementation. Core-area calculations require a user-specified edge-influence depth. The 100 to 300 m range is empirically grounded but not universal: it varies with forest type, surrounding land use and the specific ecological process of concern. Any honest fragmentation report should state the depth assumption explicitly and, where possible, test sensitivity across the plausible range.
Resolution determines what counts as a patch
This is where sensor choice has direct consequences for the numbers. A 30 m Landsat classification will merge patches that a 10 m Sentinel-2 classification resolves as separate. Edge density computed from Landsat is systematically lower than the same metric from Sentinel-2 over the same landscape, because narrow clearings and thin roads are absorbed into adjacent forest pixels. Comparing fragmentation trends across sensors without accounting for this is a common and consequential error.
For landscapes with fine-grained clearing patterns, such as smallholder agriculture or selective road networks, Sentinel-2 at 10 m is the practical minimum for credible patch delineation. Planet SuperDove at 3 m adds meaningful resolution for narrow riparian buffers and small isolated woodlots. WorldView-3 sub-metre data is used for ground-truth validation and boundary accuracy assessment rather than wall-to-wall mapping, given cost and archive limitations.
Cloud cover is an unavoidable constraint in tropical regions. A single-date image rarely produces a clean classification. Multi-temporal compositing over a defined season, using the clearest available pixels, is standard practice. Sentinel-2's five-day revisit makes this feasible at no data cost. The compositing window introduces a temporal ambiguity: the resulting classification represents conditions across several weeks, not a single moment.
Carbon consequences that REDD+ accounting can miss
Edge effects have a direct bearing on carbon stocks, and the gap between what aggregate deforestation figures report and what the landscape actually holds is measurable. Interior forest typically carries higher above-ground biomass than edge-zone forest of the same species composition, because edge trees experience greater mechanical stress and mortality. When fragmentation increases without a recorded area loss, the effective carbon stock declines anyway.
Quantifying this requires pairing the fragmentation metrics with a biomass layer. The combination of a core-area map (which identifies interior versus edge-zone pixels) with a published biomass density estimate for the forest type allows an edge-effect carbon penalty to be computed per patch and summed to the landscape. This is not a substitute for direct biomass measurement, but it surfaces a loss term that standard REDD+ area accounting does not capture. The magnitude depends heavily on the edge-influence depth assumed and the biomass gradient data used, so uncertainty ranges should accompany any such estimate.
Honest limits of the method
Fragmentation metrics are only as good as the underlying classification. Spectral confusion between forest and other woody vegetation (tall shrubs, agroforestry, bamboo) introduces patch boundary errors that propagate into every derived metric. Validation against field survey or very-high-resolution imagery is not optional if the outputs are to support policy or finance decisions.
Edge-influence depth is a parameter, not a measured quantity. The 100 to 300 m range cited in the ecological literature comes from studies in specific forest types under specific disturbance regimes. Applying a single fixed depth across a heterogeneous landscape is a simplification. Sensitivity analysis across at least two or three depth values should be standard practice, and the results should be reported as a range.
Finally, landscape metrics are snapshots. A single-epoch fragmentation map shows the current state; it does not explain how the landscape arrived there or predict where it is heading. Satellize pairs fragmentation outputs with multi-year time series derived from Landsat and Sentinel-2 archives to give clients a trajectory rather than a single data point, in the same way its Tonga crop-estimation programme uses seasonal composites rather than single-date imagery to build reliable baselines.
Turning the numbers into a decision
The output of a fragmentation analysis is a set of spatial layers and summary statistics: patch size distribution histograms, edge-density maps, core-area maps and, optionally, a connectivity graph showing which patches are within dispersal distance of one another. These are inputs to decisions, not decisions themselves.
For a conservation agency, the core-area map identifies which patches retain interior conditions worth protecting. For a carbon project developer, the edge-effect carbon penalty quantifies a degradation term that standard area accounting misses. For a national forest authority, the trend in mean patch size over a decade is a more sensitive early-warning indicator of landscape-level degradation than total area loss alone. The metric that matters depends on the question being asked, which is why the analytical specification should be agreed before the classification is run, not after.
Typical figures
| Typical spatial resolution | 10 m (Sentinel-2), 3 m (Planet SuperDove), 30 m (Landsat 8/9) for wall-to-wall mapping; 0.31–1.24 m (WorldView-3) for validation |
| Minimum resolvable patch | ~0.1 ha at 10 m; ~0.01 ha at 3 m; smaller patches are below reliable delineation threshold |
| Revisit / compositing window | 5-day (Sentinel-2, two satellites); near-daily (Planet SuperDove); 16-day (Landsat per satellite). Tropical cloud cover typically requires 30–90 day compositing windows |
| Spectral bands used | Red, NIR, red-edge (Sentinel-2 B4, B8, B5/B6/B7); SWIR optional for shadow/cloud separation. NDVI and textural features standard inputs to classifier |
| Edge-influence depth assumption | 100–300 m inward from boundary; user-specified parameter; sensitivity analysis across this range recommended |
| Archive depth | Sentinel-2 from 2015; Landsat from 1984; Planet SuperDove commercial archive from approximately 2021 |
| Classification accuracy target | Overall accuracy >90%, Kappa >0.85 recommended before metrics are computed; lower accuracy inflates edge-density estimates |
| Delivery formats | GeoTIFF patch rasters, vector patch polygons (GeoPackage or Shapefile), CSV/Excel metric tables, PDF summary report with maps |
| Latency (change epoch) | Days to weeks depending on cloud cover in the compositing window; not a near-real-time product |
Analytics Satellize can run
| Binary forest/non-forest classification | Random-forest or support-vector-machine classifier trained on multispectral and textural features; multi-temporal compositing for cloud removal | GeoTIFF raster layer; accuracy report with confusion matrix |
| Patch size distribution | Connected-component labelling on binary raster; area computed per patch in hectares | Vector polygon layer with per-patch area attribute; histogram chart in PDF report |
| Edge-density map | Perimeter-to-area calculation per patch following FRAGSTATS methodology; aggregated to user-defined grid cells | GeoTIFF raster of edge density (m/ha) at chosen grid resolution; summary statistics table |
| Core-area index layer | Negative buffer applied to each forest patch at specified edge-influence depth (100–300 m); core area computed as proportion of total patch area | GeoTIFF core-area raster; per-patch core-area index in vector attribute table |
| Edge-effect carbon penalty estimate | Core-area map combined with published regional above-ground biomass density estimates; difference between interior and edge-zone biomass applied spatially | Spatial carbon-penalty layer (tC/ha difference); landscape-level summary in tabular report with stated uncertainty range |
| Multi-year fragmentation trend series | Repeat classification at annual or biennial intervals from Sentinel-2 and Landsat archive; time-series of mean patch size, edge density and core-area index | Time-series chart and GIS layers per epoch; trend summary in PDF report |
| Patch connectivity graph | Nearest-neighbour distance between patch centroids; graph constructed at specified dispersal-distance threshold | Vector network layer; table of isolated versus connected patches by size class |
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.