Reconstruction and optimal modelling of a flash flood in a steep Norwegian river using remotely sensed- and in-situ data

The 2017 flash flood affecting Storelva in Utvik (West Norway), with an average slope of >10% in the ungauged reach, was reconstructed using visual observations during the event, as well as post-event field data and remote sensing. The dataset was then used for i) roughness calibration and sensitivity analysis, ii) validation of a 2D hydrodynamic model (morphodynamic data was insufficient) and reconstruction of the maximum flood extent, critical locations, and preferential flow paths, and its comparison to other modelling studies, and iii) analysis of the impact of mesh refinement on model precision for optimal model design in IberPlus.

Water levels and flow discharge were measured after the flood. The observations were used to calibrate the model in the 400m-long most downstream reach. Similarly, visual flood documentation during the event was used to model the event and validate it in the 800m-long most downstream reach.

To calibrate the model, GIS-classified wet and dry areas in the computational domain were compared with wet and dry areas observed along both banks, calculating the BIAS and RMSE for each calibration Manning. According to the sensitivity analysis, the model with Manning’s roughness coefficient of 0.065 in the upper and middle reach and 0.075 downstream showed the lowest global errors (i.e. RMSE= 1.1cm), although the numerical models generally underestimated the observed water levels (i.e. -8cm <BIAS< -1cm).

Two of the critical locations are located near bridges and the other two near a bank with very fine material, easy to erode. The preferential flow paths indicate that the erosion occurred mainly in the left floodplain. IberPlus simulated satisfactorily the observed maximum flood extent, i.e. F and C indices of 60%–87%. The results for the 2017 flood using IberPlus were compared to the (non-calibrated) hydraulics from literature using TELEMAC-MASCARET and FINEL2D. The IberPlus hydrodynamic model had the highest roughness coefficients from all the modelling studies. This might explain the significantly higher hydraulic values observed, in agreement with those obtained by the morphodynamic models. The paths preferred by the flow during the flood and the flood extent are resembling in all three models.

The F and C indices and the incremental precision between scenarios were estimated for 44,000–11.6 million cells models with uniform and variable mesh sizes. The optimal precision-gain was at model size <150,000 cells for variable mesh (R² =0.65) versus >700,000 cells for uniform mesh (R² >0.94), with a precision gain limited to 5–7% at best when using a finer grid. Uncertainties in the flood mapping used for validation, the hydrodynamic model set-up and input data contributed to the offset. The model precision is limited by the on-site flood protections implemented to protect private property during the flood event. These protections were effective and reduced the flood damage by 43%, yet they could not be implemented in the numerical model. Also, the model validation was carried out against a fully water-covered area, where some local dry cells were considered wet. Remotely sensed data helps understand flood dynamics and monitor flood risk in data-scarce regions.