Pumped-hydro energy storage site screening from satellite topography
Satellite-derived digital elevation models can screen entire continents for paired pumped-hydro reservoir sites in days, ranking candidates by potential energy storage capacity before any field investigation begins. The method is fast and cheap; it is also blind to geology, land tenure, and ecology.
Sensors
- TanDEM-X DEM: Global DEM at 0.4 arcsecond (~12 m) posting, absolute vertical accuracy better than 10 m (90th percentile), relative vertical accuracy around 2 m over smooth terrain. The primary source for head-difference calculations where sub-metre relative accuracy matters most.
- Copernicus GLO-30 DEM: Globally available at 1 arcsecond (~30 m) posting, derived from TanDEM-X data and freely downloadable. Adequate for first-pass continental screening; volume estimates carry larger uncertainty than TanDEM-X at 12 m.
- SRTM: NASA Shuttle Radar Topography Mission DEM at 1 arcsecond (~30 m) for latitudes between 60°N and 56°S. Older (2000 acquisition) and with higher void rates in steep terrain than GLO-30, but well-validated and widely used for cross-checking.
- Sentinel-2 MSI: 13-band multispectral imager at 10 m (visible/NIR) and 20 m (red-edge/SWIR), 5-day revisit at the equator. Used to generate land-cover masks that exclude existing built areas, permanent water bodies, forests, and protected zones from candidate site lists.
What the Australian National University showed the world
The Global Pumped Hydro Atlas, published by the ANU RE100 group, demonstrated that DEM-based automated screening could identify roughly 616,000 candidate site pairs globally, representing an aggregate potential storage capacity orders of magnitude larger than any plausible grid need. The methodology is public and reproducible: identify all local topographic saddles, pair upper and lower reservoirs within a defined distance and head-difference window, compute inundation volume from the DEM at a given dam-wall height, and convert head times volume to potential energy in gigawatt-hours. The computation is embarrassingly parallelisable across a continental DEM grid.
That result matters not because every site is viable but because it collapses the search space. A developer who once needed months of helicopter surveys and topographic map digitisation can now receive a ranked shortlist in a week. The satellite data does not build the dam; it tells you which valleys are worth arguing about.
How head and volume translate into gigawatt-hours
Potential energy stored in a pumped-hydro system is E = ρ g h V η, where ρ is water density, g is gravitational acceleration, h is the net head between upper and lower reservoir water surfaces, V is the smaller of the two reservoir volumes, and η is round-trip efficiency (typically 0.70 to 0.80 for modern plant). All four physical quantities except efficiency can be estimated from a DEM alone.
Head is computed as the elevation difference between the centroids or spillway sills of the two candidate reservoirs. Volume is estimated by filling the DEM basin to a specified dam-wall height and summing the inundated cell volumes. At GLO-30's 30 m posting, volume errors of 10 to 30 percent are realistic for small basins; TanDEM-X at 12 m reduces that range but does not eliminate it. Narrow gorges and overhanging cliffs are systematically undersampled by radar interferometry, so volume estimates in such terrain should be treated as lower bounds.
The ANU screening methodology applies a set of configurable filters: minimum head (typically 100 m to 700 m depending on application), maximum reservoir area (to avoid inundating large inhabited valleys), maximum tunnel or penstock length between the pair (which drives civil cost), and exclusion of sites where the upper reservoir would sit above a defined altitude. Each filter is a spatial query against the DEM and ancillary layers.
Land-cover masking: where Sentinel-2 earns its place
A DEM alone will happily nominate a city centre or a national park as an excellent reservoir site. Sentinel-2 classification layers fix that. Standard products derived from the Copernicus Land Service or from scene-level classification of Sentinel-2 imagery can mask out urban fabric, existing water bodies, dense forest (where inundation triggers carbon and biodiversity accounting), and agricultural land under active cultivation.
The masking step is not a formality. In many tropical and subtropical regions, the most topographically attractive sites overlap heavily with primary forest or indigenous land. Removing those early, before any engineering assessment, saves time and avoids generating a shortlist that will be immediately rejected on non-technical grounds. The Sentinel-2 derived Dynamic World product (published by Brown et al. in Nature Communications, 2022) provides near-real-time global land cover at 10 m that is well-suited to this masking role.
What satellite screening cannot tell you
This is where candour is commercially useful. DEM-based screening produces a geometric ranking, nothing more. It says nothing about the permeability of the reservoir floor (a karst limestone basin will drain), the seismic hazard at the dam site, the water rights and abstraction law in the jurisdiction, the presence of archaeological sites under the inundation zone, or the transmission distance to the load centre that needs the storage.
Vertical accuracy is the other honest caveat. TanDEM-X achieves relative vertical accuracy of around 2 m over low-relief terrain but degrades in forested areas, where the radar phase centre sits above the ground surface, and in steep terrain, where layover and shadow create voids. A 5 m error in the estimated spillway elevation of a small upper reservoir translates directly into a proportional error in head and therefore in the GWh estimate. Sites with computed storage below roughly 2 GWh should be treated as indicative rather than bankable.
Geology, hydrogeology, and land tenure together probably eliminate 90 percent or more of geometrically attractive sites. The satellite screen is a first filter, not a feasibility study.
Running the screen: practical workflow
A standard continental-scale run ingests the GLO-30 DEM tiles for the region of interest, applies a depression-filling and flow-direction algorithm to identify candidate saddle points, then executes the paired-reservoir search within configurable head and distance bounds. Sentinel-2 derived land-cover masks are applied as exclusion polygons. The output is a GeoPackage or shapefile of candidate site pairs, each attributed with estimated head (m), upper and lower reservoir volumes (Mm³), computed potential storage (GWh), penstock length (km), and a simple cost-proxy score based on civil works distance.
For a country the size of Chile or Norway, a full run on cloud compute takes hours, not weeks. The resulting shortlist, typically a few hundred to a few thousand sites after filtering, is the starting point for the next stage: geological map overlay, protected-area intersection, grid proximity analysis, and ultimately site visits. Satellize runs this workflow on GLO-30 and TanDEM-X inputs, applying the ANU-published methodology with client-configurable filter thresholds. The Tonga crop-estimation programme is a different domain, but the underlying pattern, open satellite data plus reproducible spatial analysis at national scale, is the same approach applied here.
DEM choice and its consequences for project stage
For early-stage national or regional policy screening, GLO-30 at 30 m is sufficient and free. For pre-feasibility work on a shortlist of, say, 20 to 50 sites, TanDEM-X at 12 m is worth the licence cost because the improvement in volume estimation reduces the risk of advancing a geometrically marginal site. For detailed feasibility on a single site, neither product is adequate; airborne LiDAR or drone photogrammetry at sub-metre resolution is required to support dam design and volume certification.
SRTM remains useful as a validation cross-check and for archive studies, but its 2000 acquisition date means it predates significant land-use change in many regions, and its void-filling interpolation in mountain areas can introduce systematic errors in head calculations. Where GLO-30 and SRTM head estimates diverge by more than 20 m for the same site pair, that discrepancy is itself a flag for closer inspection.
Typical figures
| Primary DEM spatial resolution | 12 m (TanDEM-X 0.4 arcsec) or 30 m (GLO-30 / SRTM 1 arcsec) |
| Relative vertical accuracy (TanDEM-X) | ~2 m over low-relief terrain; degrades to 4–10 m in steep or forested terrain |
| Absolute vertical accuracy (GLO-30) | Better than 4 m LE90 over most terrain types per Copernicus documentation |
| Land-cover mask resolution | 10 m (Sentinel-2 MSI visible/NIR bands) |
| Sentinel-2 revisit for land-cover update | 5 days at equator (twin-satellite constellation) |
| DEM archive depth | TanDEM-X acquired 2010–2015; GLO-30 released 2021; SRTM acquired February 2000 |
| Minimum detectable head difference | Practically ~50 m given DEM vertical noise; sites below 100 m head carry high uncertainty |
| Typical volume estimation error (30 m DEM, small basins) | 10–30%; reduces to ~5–15% at 12 m posting |
| Coverage | Global (GLO-30 and TanDEM-X); SRTM limited to 60°N–56°S |
| Deliverable formats | GeoPackage, Shapefile, GeoTIFF (DEM derivatives), CSV ranked site table, PDF summary report |
Analytics Satellize can run
| Candidate site-pair ranked shortlist | ANU Global Pumped Hydro Atlas automated saddle-point and paired-basin search on GLO-30 or TanDEM-X DEM | GeoPackage with head (m), volume (Mm³), potential storage (GWh), penstock length (km), and cost-proxy score per site pair |
| Land-cover exclusion mask | Sentinel-2 MSI classification (urban, forest, water, protected area polygons) applied as spatial filter | GeoTIFF exclusion raster and vector polygon set, compatible with GIS overlay on site-pair layer |
| Basin volume hypsometric curves | DEM inundation modelling at incremental dam-wall heights (area–elevation–volume curves) | Per-site CSV of area and volume vs. elevation, enabling rapid sensitivity analysis on dam height |
| Head uncertainty map | Propagation of DEM vertical error statistics through head-difference calculation; comparison of GLO-30 and SRTM estimates at each site | GeoTIFF of per-site head uncertainty (m) and flag layer for sites where GLO-30/SRTM divergence exceeds 20 m |
| Penstock and tunnel corridor length | Least-cost path analysis across DEM terrain between upper and lower reservoir centroids | Vector line layer with path length (km) and cumulative elevation change, attributable to site-pair table |
| Filtered national or regional atlas | Full ANU-methodology pipeline with client-configurable thresholds (minimum head, maximum area, exclusion zones) applied to country or region of interest | Interactive web map and PDF summary report ranking top N sites by GWh potential after all filters |
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.