Habitat fragmentation and patch-connectivity mapping
Patch size, shape, edge density and inter-patch distance tell ecologists what a land-cover map alone cannot: whether a landscape can still sustain viable populations. Satellite-derived fragmentation indices turn that question into a measurable, repeatable number.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands; 5-day revisit at the equator with both satellites. The red-edge bands (B5, B6, B7 at 20 m) improve separation of woody from herbaceous cover. Free and open archive from 2015.
- Landsat 8/9 OLI: 30 m resolution, 16-day revisit per satellite (8-day combined). Consistent archive back to 1984 makes it the standard source for long-term fragmentation trend analysis. USGS Collection 2 surface-reflectance products are analysis-ready.
- Maxar WorldView-3: 30 cm panchromatic, 1.24 m multispectral, 3.7 m SWIR. Resolves individual tree crowns and narrow linear features such as tracks and firebreaks that fragment habitat below Sentinel-2's detection floor. On-demand tasking; cost limits area coverage.
- Copernicus DEM (TanDEM-X derived): GLO-30 product at 30 m posting, GLO-90 at 90 m. Slope and aspect layers constrain least-cost dispersal modelling by encoding terrain barriers that a flat land-cover mask would ignore.
Why a single land-cover map is not enough
A habitat map shows where forest exists. A fragmentation analysis shows whether the remaining forest is functionally connected. Those are different questions, and confusing them has real consequences for conservation planning. A landscape can retain 60 % tree cover and still be ecologically broken if that cover sits in dozens of small, isolated patches separated by roads and cropland.
The standard framework for quantifying this comes from landscape ecology: binary habitat masks derived from classified imagery are fed into metrics libraries such as FRAGSTATS or its open equivalents. The core indices include patch area, perimeter-to-area ratio, edge density, nearest-neighbour distance, and aggregation index. Each captures a different dimension of fragmentation. Patch area predicts local extinction risk through species-area relationships. Edge density measures the proportion of habitat exposed to matrix influence, which matters for interior-specialist species. Nearest-neighbour distance feeds directly into dispersal-probability models.
Building the classification base: resolution choices and their trade-offs
The choice of imagery resolution sets a hard floor on what fragmentation indices can detect. At Sentinel-2's 10 m, a habitat patch must be roughly 100 m² to register as a distinct pixel, but meaningful ecological patches are typically far larger, so the practical minimum detectable patch is closer to 0.1 ha when accounting for classification noise at edges. Narrow linear corridors, say a 20 m riparian strip, are routinely underestimated or missed entirely at this resolution.
Landsat at 30 m is worse for small-patch detection but invaluable for temporal depth. The archive runs to 1984, which means fragmentation trajectories over four decades are computable with consistent methodology. That historical baseline is often more useful to a conservation agency than a single high-resolution snapshot.
WorldView-3 at 1.24 m multispectral resolves the fine-grained features that fragment habitat below Sentinel-2's detection floor: unpaved tracks, isolated trees retained in pasture, narrow watercourse buffers. The trade-off is cost and area. Very-high-resolution tasking is practical for priority corridors or validation sites, not wall-to-wall national coverage.
Accuracy of the underlying land-cover classification propagates directly into fragmentation metrics. A 10 % misclassification rate in a binary forest mask can shift edge-density estimates by a similar margin. Published studies using Sentinel-2 with random-forest classifiers routinely report overall accuracies of 85 to 92 % for broad forest versus non-forest classes in tropical regions, but accuracy drops in transition zones and under persistent cloud. Honest fragmentation analysis should carry that uncertainty forward into the metrics, not suppress it.
What the DEM adds, and what it cannot fix
Topography shapes dispersal in ways a flat land-cover layer ignores. A 400 m ridge between two forest patches is a genuine barrier for many amphibians and small mammals, even if the Euclidean distance between patches looks manageable. The Copernicus GLO-30 DEM, derived from TanDEM-X radar data, provides 30 m terrain elevation suitable for computing least-cost paths and resistance surfaces. Slope, aspect and curvature layers derived from it feed directly into circuit-theory connectivity models such as Circuitscape.
The DEM does not resolve habitat quality within patches. It tells you a slope is steep; it does not tell you whether the forest on that slope is old-growth or a recently logged secondary stand. That distinction requires spectral indices from optical imagery, and in some cases lidar, which is outside the scope of this page.
Fragmentation metrics in practice: what the numbers mean
FRAGSTATS and equivalent tools produce metrics at three levels: patch, class and landscape. Patch-level metrics describe individual polygons. Class-level metrics aggregate all patches of a given habitat type. Landscape-level metrics describe the entire mosaic. For most conservation applications, class-level metrics are the operational currency.
A few indices deserve specific attention. The Effective Mesh Size (MESH) translates fragmentation into a single intuitive number: the area of a hypothetical unfragmented landscape that would give the same probability of two randomly placed points being in the same patch. A MESH of 500 ha in a region that was once a continuous forest block of 50,000 ha is a stark signal of functional collapse. The Patch Cohesion Index measures physical connectedness of patches irrespective of their number. The Aggregation Index measures how clumped or dispersed patches are across the landscape.
These metrics are sensitive to the minimum mapping unit chosen. Halving the minimum patch size from 1 ha to 0.5 ha can substantially increase patch count and edge density, making the landscape appear more fragmented without any real change on the ground. Any reported metric should state the minimum mapping unit explicitly. This is not a technicality; it is the difference between comparable and incomparable datasets across agencies or time periods.
Limits the buyer should know before commissioning the work
Cloud is the persistent enemy of optical fragmentation mapping in humid tropical regions. Sentinel-2 carries no cloud-penetrating capability; a site under persistent cloud cover in the wet season may have only a handful of usable scenes per year. Seasonal compositing reduces the problem but introduces phenological mixing, which can shift class boundaries in the final mask.
Fragmentation metrics are descriptive, not predictive on their own. They quantify landscape structure; they do not directly predict species response without additional ecological modelling. The link between, say, a nearest-neighbour distance of 2 km and the local extinction probability of a given species requires species-specific dispersal data that satellite imagery cannot supply.
Change detection between two epochs assumes consistent classification methodology. If the land-cover legend, minimum mapping unit or accuracy threshold differs between the baseline and current maps, apparent fragmentation change may be an artefact of method rather than reality. Multi-temporal fragmentation analysis should use a single classification pipeline applied consistently across all dates.
Satellize applies this pipeline operationally, using open Sentinel and Landsat archives combined with commercial tasking where patch-level detail demands it. The Tonga crop-estimation programme demonstrated the same underlying workflow: consistent classification, validated against ground truth, delivered as analysis-ready GIS layers rather than raw imagery.
From metrics to a decision
Fragmentation indices become useful to a decision-maker when they are tied to a threshold or a comparison. A protected-area manager needs to know not just that edge density has increased by 15 % since 2015, but whether that change crosses a threshold associated with measurable biodiversity loss in the literature. A national biodiversity strategy needs patch-connectivity maps that are comparable across administrative units and repeatable on a fixed schedule.
The practical output of a fragmentation analysis is a set of GIS layers: the classified habitat mask, the patch polygon dataset with metric attributes, and a summary table of class-level and landscape-level indices for each reporting unit. Those layers feed directly into corridor-prioritisation models, environmental-impact assessments and national biodiversity monitoring frameworks. The analysis is not an end in itself; it is the quantitative foundation that makes subsequent decisions defensible.
Typical figures
| Spatial resolution (classification base) | 10 m (Sentinel-2 visible/NIR), 20 m (Sentinel-2 red-edge/SWIR), 30 m (Landsat 8/9), 1.24 m (WorldView-3 multispectral) |
| Minimum detectable patch (Sentinel-2) | Approximately 0.1 ha under good conditions; narrow linear features below ~20 m width are unreliable |
| Revisit frequency | 5 days (Sentinel-2, equatorial, both satellites); 8 days combined (Landsat 8+9); on-demand (WorldView-3) |
| Spectral bands used | Visible (B2–B4), NIR (B8), red-edge (B5–B7), SWIR (B11–B12) for Sentinel-2; OLI bands 2–7 for Landsat |
| Topographic input | Copernicus GLO-30 DEM at 30 m posting; GLO-90 at 90 m for regional-scale dispersal modelling |
| Archive depth | Sentinel-2 from 2015; Landsat from 1984 (Collection 2 surface reflectance) |
| Classification accuracy (typical, broad forest/non-forest) | 85–92 % overall accuracy with random-forest classifiers on Sentinel-2; lower in persistent cloud and transition zones |
| Delivery formats | GeoTIFF habitat mask, GeoPackage or Shapefile patch polygon layer with metric attributes, CSV summary table of FRAGSTATS-equivalent indices |
| Latency (standard archive-based run) | Days to weeks depending on cloud cover and area extent; near-real-time not applicable for classification-based fragmentation |
Analytics Satellize can run
| Binary habitat mask | Supervised random-forest or support-vector-machine classification on Sentinel-2 or Landsat surface-reflectance composites, with red-edge and SWIR bands as primary discriminators | GeoTIFF raster layer, per-class accuracy report |
| Patch polygon dataset with FRAGSTATS-equivalent metrics | Raster-to-vector conversion of habitat mask; computation of area, perimeter, shape index, fractal dimension, nearest-neighbour distance and core-area index per patch | GeoPackage or Shapefile with metric attributes in the attribute table |
| Class- and landscape-level fragmentation summary | Aggregation of patch metrics to class and landscape levels; computation of Effective Mesh Size, Aggregation Index and Patch Cohesion Index | CSV or Excel summary table, one row per reporting unit and epoch |
| Multi-temporal fragmentation change report | Consistent classification pipeline applied to baseline and current imagery; differencing of class-level indices between epochs | Change map (GeoTIFF), trend table, narrative PDF report |
| Least-cost connectivity surface | Resistance surface derived from habitat mask and Copernicus DEM slope layer; least-cost path or Circuitscape circuit-theory analysis between defined patch centroids | Raster resistance surface and least-cost path vectors, suitable for corridor-prioritisation input |
| Edge-effect zone delineation | Inward buffering of patch polygons at ecologically defined distances (typically 50–200 m) to separate core habitat from edge-affected zones | Core-area and edge-zone GeoTIFF layers, core-area fraction statistic per patch |
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.