Uncertainty Quantification in Hydrology

Both panels are on the same discharge scale, and both bands are drawn at their measured widths. Moving the water level costs 356.1× more than moving the rainfall, because discharge is never measured directly — it is inferred from stage through a fitted curve whose local exponent rises from 2.35 to 8.48 as the river rises, so one fixed gauge error buys 39× more error at the flood peak than at baseflow. Refitting the model to each corrupted series recovers 5.76% of the loss and still never beats the reference: you cannot calibrate your way out of an error in the thing you are calibrating against.

The chain behind those two panels: HBV001a, a lumped conceptual rainfall-runoff model with 18 parameters, calibrated by Differential Evolution over 600 generations and 75,480 model evaluations; then sensitivity analysis, locally around that optimum and globally with Sobol indices under Saltelli sampling; then the two perturbation studies above. The two Sobol configurations disagree about which parameter dominates — sl0_fcy sampled narrowly against lrr_dre sampled across the full space — and their total indices sum to 1.856 rather than 1, so most of the variance sits in interactions rather than in any single parameter.

Calibrated baseline

NSE 0.907715

After precipitation noise

NSE 0.907298

After rating-curve error

NSE 0.759226

Rating curve fit

R² 0.998705

• Group courseworkThree contributors across five assignments; individual ownership of any single result is not claimed.
• Course inputs are not redistributableThe forcing series, catchment area and the hmg package holding HBV001A are course-provided. The scientific runs cannot be reproduced from the public repository alone.
• The event shape here is schematicNo observed hydrograph is reproduced. The rating curve, the perturbation and every metric are the repository's own; the trajectory they act on is a placeholder.
• The README and the code disagreeThe README describes a plus/minus 15 cm water-level perturbation. The module bounds and the results file both record plus/minus 25 cm; 15 survives only as a stale default argument. The figures here follow the code.
• Random error onlyThe precipitation study perturbs without systematic bias, which its own summary notes is the more consequential case and was not tested.
• One event, one catchmentA single short high-flow event. Nothing here establishes that the same ordering holds for longer records or other catchments.