How EAVES reconstructs elevation-area-volume (EAV) curves from SRTM topography. The full scientific treatment, including validation and uncertainty, is in the accompanying Scientific Data manuscript.
The algorithm searches for a terrain-derived dam wall across the valley at or near the cataloged dam coordinates. Six placement stages are attempted in sequence, from fastest to most exhaustive:
| Stage | Strategy | Description |
|---|---|---|
| 1 | Fast path | Try terrain-derived wall angles at the nominal location |
| 2 | Upstream walk | Walk upstream along the valley thalweg and retry at each position |
| 3 | Quality recovery | Re-search if the initial fill is geometrically suspect (downstream-skewed or too small) |
| 4 | River-direction retry | Shift anchor along the river-network flow vector |
| 5 | Relaxed alignment | Allow wall orientations that would normally be rejected by the flow-alignment filter |
| 6 | Fallback | Multi-direction flood fill without an explicit wall |
Once the footprint is established, elevation bins (0.5 m intervals) are used to compute area at each level, and cumulative trapezoidal integration yields volume. The fill is capped at the catalog capacity, acting at bin resolution: the curve is truncated at the first bin whose cumulative volume reaches the capacity, so capped fills can overshoot by up to one bin. A two-parameter power law (V = c · Ab) is fitted via non-linear least squares, and the released exponent is clamped to [1.1, 2.0] with c re-solved through the recovered full-pool anchor.
Fits passing the reliability gates (quality grades A-B, R² ≥ 0.98, 0.3 ≤ V_SRTM/V_cap ≤ 5.0, n_pixels ≥ 50, b defined) form the trusted set. Of these, only dams built in or after 2000 (verifiably postdating the February 2000 SRTM acquisition) form the training set that the regionalization, the exponent spread b_σ, and the leave-one-out validation are computed on. Pre-2000 and unknown-year dams ship their own SRTM curves (flagged pre_srtm / unknown_year) but do not train the recipe, because their valley floors may already carry sediment. On the Saudi domain: 322 trusted, 200 training.
Dams that fail the trusted gates receive parameters from a single closed-form recipe:
- Exponent b: regional median over the capacity-thresholded training subset (or a multivariate regression on
valley_ratio,channel_slope,mean_catchment_slope,dam_height_mif its leave-one-out R² ≥ 0.25, which rarely holds for arid catchments). - Coefficient c: back-solved as c = V_cap/A_capb from catalog capacity and a multi-feature linear regression that predicts log A_cap from seven log-space features:
capacity_mcm,dam_height_m,spillway_height_m,valley_ratio,channel_slope,mean_catchment_slope,upstream_area_km2. Any feature missing for a given dam is imputed with the training-set median so the regression always returns a finite value.
Leave-one-out cross-validation on the training set quantifies the recipe's accuracy. For the Saudi Arabia deployment: 92% of predictions within a factor of 2 and 99% within a factor of 3 of the SRTM-derived reference, median bias +6%, relative RMSE 47%. See eaves.postprocess.validation and panel p5 for the full per-recipe comparison and the rationale for retiring two earlier candidates (a satellite-anchored recipe and a single-feature log-log regression).
Automated quality gates detect displaced flood centroids and negligible fill volumes, flagging problematic dams for regional parameter assignment rather than propagating unreliable fits.
EAVES reconstructs reservoir geometry from the SRTM surface, not a surveyed bathymetric record. Outputs are a best-effort approximation rather than absolute capacity.
- Valley-geometry approximation, not bathymetry: curves follow the SRTM valley surface up to the spillway, not a measured reservoir bottom. They are sensitive to DEM noise (vertical LE90 ~6 m) in the same way the underlying terrain is.
- Synthetic dam wall: the wall orientation and length come from a terrain-alignment search at or near the catalog coordinates. It is the best-fit crest for that SRTM patch, not necessarily the engineered as-built structure, and small placement shifts can meaningfully change the reconstructed footprint.
- SRTM snapshot (February 2000): dams built after 2000 get a clean pre-impoundment valley. Dams built earlier carry whatever sediment had accumulated by the acquisition date, so their curves describe the as-of-2000 surface. These dams are flagged (
pre_srtm, orunknown_yearwhen the build year is missing) and excluded from regionalization training. - Resolution-limited regimes: sub-pixel reservoirs (
n_pixels < 30), narrow valleys (valley_width_m < 3 x pixel_size), and shallow depressions (spillway_height_m < 5 m) produce curves with elevated uncertainty. See theuncertainty_flagscolumn ineaves_summary.csvfor per-dam tagging.
The full quantitative treatment of these limitations is in the accompanying publication.