Karst surface-vegetation mapping as a biodiversity proxy for cave and subterranean ecosystems
Cave-adapted fauna depend on leaf litter and dissolved organic matter falling through karst from surface vegetation. Mapping that surface layer with Sentinel-2, GEDI and terrain roughness indices tells conservationists where subterranean ecosystems are most at risk before a single speleologist descends.
Sensors
- Sentinel-2 MSI: 13 spectral bands from 443 nm to 2190 nm at 10–60 m ground sampling distance; 5-day revisit at the equator with both satellites. Red-edge bands (B5, B6, B7 at 20 m) are particularly sensitive to canopy chlorophyll content and early stress before broadband NDVI responds. Free archive from 2015.
- GEDI (Global Ecosystem Dynamics Investigation): Spaceborne full-waveform lidar operating from the International Space Station. Delivers above-ground biomass density estimates and canopy-height profiles at approximately 25 m footprint diameter along sparse orbital tracks (sampling density is uneven; full wall-to-wall coverage requires fusion with Sentinel-2 or Landsat). Coverage limited to roughly 51.6° N/S.
- SRTM (Shuttle Radar Topography Mission): 30 m global DEM from the 2000 mission. Used here to derive surface-roughness indices (e.g. vector ruggedness measure, topographic position index) that discriminate karst tower-and-depression terrain from non-karst geology without requiring field lithological survey.
- Copernicus DEM GLO-30: TanDEM-X-derived global DEM at approximately 30 m posting, generally more accurate than SRTM in vegetated terrain. Preferred for computing slope, aspect and closed-depression (doline) density, which are the primary terrain proxies for karst surface-to-cave connectivity.
Why the surface matters to an animal that never sees sunlight
Most cave-adapted invertebrates, from troglobitic beetles to obligate cave crayfish, occupy ecosystems that produce almost no energy internally. They depend on allochthonous inputs: leaf litter, woody debris, dissolved organic carbon and invertebrate prey that wash or fall in from the surface above. The karst surface catchment is, in functional terms, the farm that feeds the cave.
This dependency means that deforestation, agricultural conversion or severe drought stress on surface vegetation can collapse subterranean food webs without anyone touching a cave entrance. The threat is invisible to standard protected-area monitoring, which typically tracks forest loss at the boundary of a reserve rather than over the specific karst catchments that drain into known cave systems. Satellite vegetation mapping closes that gap, at least at the surface.
Delineating the karst catchment before any spectral analysis begins
Karst hydrology does not follow surface topography in the way that a standard watershed algorithm assumes. Water sinks through dolines and conduits and resurges at springs that may be kilometres away and topographically uphill from the apparent drainage divide. A useful surface-input proxy therefore requires two delineation steps: first, identifying the karst terrain itself; second, estimating the likely catchment of each cave or spring system.
Terrain identification uses SRTM or GLO-30 derivatives. Closed-depression density, computed by inverting the DEM and running a standard fill-and-difference operation, is a well-documented karst discriminator. The vector ruggedness measure (VRM) captures the fine-scale roughness of tower karst that distinguishes it from structurally similar but non-karstic terrain. These terrain indices are then cross-checked against published lithological maps, because a closed depression in granite is not a doline. Published karst atlases, including the World Karst Aquifer Map (WOKAM), provide the lithological anchor.
Catchment delineation beyond the terrain step is genuinely difficult. Dye-tracing and speleological survey remain the only reliable methods for establishing conduit connectivity. The satellite approach cannot replace them; it can only flag which surface polygons are worth tracing in the first place.
What Sentinel-2 red-edge bands reveal that broadband NDVI misses
Once the karst catchment polygon exists, Sentinel-2 multitemporal composites characterise vegetation condition across it. Standard NDVI (using bands B4 and B8) saturates at moderate-to-high canopy density, which is precisely the condition of interest in intact tropical karst forest. The red-edge bands, B5 at 705 nm and B7 at 783 nm, remain sensitive at higher chlorophyll concentrations and respond earlier to drought or nutrient stress. The red-edge chlorophyll index (B7/B5 minus one) is a practical improvement for dense-canopy monitoring.
Disturbance history is extracted from the time series rather than any single image. A cloud-masked annual composite stack, built from the Sentinel-2 Level-2A surface-reflectance archive going back to 2017 for most scenes, reveals clearance events, selective extraction and recovery trajectories. Phenological metrics, specifically the amplitude and timing of the seasonal greenness cycle, distinguish evergreen karst forest from degraded secondary growth that has similar peak NDVI but a different seasonal shape.
Cloud cover is the principal operational limit. Tropical karst regions, including the major cave systems of Borneo, southern China, Vietnam and Papua New Guinea, sit under persistent cloud. A single-date image is nearly useless; a 12-month composite built from all clear-sky observations is the minimum viable product. Even then, some catchments in persistently cloudy years may have fewer than five usable observations, which degrades confidence in the disturbance detection.
GEDI biomass estimates: what the sparse sampling actually gives you
GEDI's above-ground biomass density (AGBD) product provides canopy-height and biomass estimates at 25 m footprints along ground tracks spaced roughly 600 m apart at the equator. That sampling pattern is too sparse to map a single karst catchment directly, but it is dense enough to calibrate a Sentinel-2 regression model. The standard approach fuses GEDI AGBD samples with Sentinel-2 spectral and textural predictors using a random-forest or gradient-boosting regressor, producing a wall-to-wall biomass map at 20–30 m resolution.
Published validation studies suggest root-mean-square errors of roughly 20–40 Mg ha⁻¹ in tropical forest, which is acceptable for ranking catchments by biomass class but not for precise carbon accounting. For the purpose of this use case, the biomass layer serves as a proxy for organic-matter input potential: high-biomass, structurally complex forest above a karst catchment is assumed to deliver more allochthonous input than low-biomass degraded scrub. That assumption is ecologically reasonable but has not been validated directly against cave food-web productivity measurements in the published literature.
Translating the map into a survey-prioritisation score
The analytic output is a catchment-level score that ranks karst surface units by their estimated input potential and recent disturbance pressure. Four components feed it: current vegetation biomass (from the GEDI-Sentinel-2 fusion), recent disturbance intensity (from the multitemporal NDVI anomaly stack), disturbance trajectory (recovering, stable or worsening over the archive period), and catchment area relative to known cave or spring locations from published speleological databases.
High scores indicate intact, high-biomass catchments with no recent disturbance: these are the systems where subterranean biodiversity is most likely to be in good condition and where survey effort is most likely to find intact communities. Low scores indicate degraded or actively cleared catchments where ground teams should prioritise rapid biological assessment before conditions deteriorate further. The score does not predict cave species richness; it predicts the surface conditions that make richness plausible.
Satellize applies this workflow using open Sentinel-2 and GEDI data, with terrain analysis on GLO-30. The Tonga crop-estimation programme demonstrated that the same Sentinel-2 composite pipeline scales readily to island and archipelago geographies, which include several globally significant karst regions. Clients receive a GIS-ready layer with the catchment scores, the underlying biomass and disturbance rasters, and a written interpretation of the highest-priority survey targets.
Honest limits: what this method cannot tell you
The surface proxy has real predictive value, but several caveats deserve stating plainly. First, karst hydrological connectivity is assumed, not measured. A degraded catchment that drains to a different cave system than the one of interest will produce a misleading score. Ground-truthing with dye traces or speleological survey is not optional for any management decision that depends on the connectivity assumption.
Second, Sentinel-2 at 10–20 m cannot detect small clearings below roughly 0.1 ha, selective vine cutting or understorey removal that leaves the canopy largely intact. These are ecologically significant disturbances in karst forest but are invisible to this sensor at this resolution. Third, GEDI coverage excludes latitudes above 51.6° N or S, which removes temperate karst systems in the British Isles, the Dinaric Alps and parts of North America from the biomass-fusion approach. SRTM-derived roughness and Sentinel-2 spectral analysis remain applicable; the biomass estimate must rely on Landsat-based regional models instead. Finally, the method says nothing about cave microclimate, hydrology or the fauna themselves. It is a surface-input proxy, not a biodiversity survey.
Typical figures
| Spatial resolution (vegetation mapping) | 10–20 m (Sentinel-2 MSI bands) |
| Spatial resolution (terrain analysis) | 30 m (GLO-30 / SRTM) |
| GEDI biomass footprint | ~25 m diameter; track spacing ~600 m at equator |
| Sentinel-2 revisit | 5 days at equator (Sentinel-2A + 2B combined) |
| Spectral bands used | Sentinel-2 B2–B8A, B11, B12 (443–2190 nm); red-edge B5, B6, B7 emphasised |
| Cloud-free composite window | Minimum 12 months recommended for tropical karst; longer in persistently cloudy regions |
| Archive depth | Sentinel-2: 2015–present; SRTM: 2000 (single epoch); GEDI: 2019–present |
| GEDI latitudinal coverage | ~51.6° N to ~51.6° S |
| Minimum detectable clearing | ~0.1 ha at 10 m resolution under clear-sky conditions; smaller clearings not reliably detected |
| Biomass estimation uncertainty | RMSE typically 20–40 Mg ha⁻¹ in tropical forest (published validation range) |
Analytics Satellize can run
| Karst catchment delineation layer | Closed-depression density and VRM from GLO-30; cross-referenced with WOKAM lithological polygons | GIS polygon layer (GeoPackage or shapefile) of karst surface catchments with terrain-class attributes |
| Multitemporal vegetation-condition composite | Cloud-masked Sentinel-2 Level-2A annual composites; red-edge chlorophyll index and NDVI time series | Annual raster stack (GeoTIFF) with per-pixel median, 10th-percentile and amplitude statistics |
| Disturbance-history map | NDVI anomaly detection against archive baseline; breakpoint identification in seasonal time series | Raster layer showing year of first detected disturbance and disturbance magnitude, 2017–present |
| Wall-to-wall above-ground biomass map | GEDI AGBD samples fused with Sentinel-2 spectral and textural predictors via gradient-boosting regression | 20–30 m biomass-density raster (Mg ha⁻¹) with per-pixel uncertainty band |
| Surface input-potential score per catchment | Weighted composite of biomass class, disturbance intensity and disturbance trajectory; ranked by percentile | Catchment attribute table and choropleth map; PDF interpretation report naming highest-priority survey targets |
| Change-monitoring alert | Near-real-time Sentinel-2 scene comparison against baseline composite; threshold-based anomaly flagging | Quarterly email alert with polygon of newly detected disturbance events above 0.5 ha threshold |
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.