A simple framework for calibrating hydraulic flood inundation models using Crowd-sourced water levels

Floods are the most commonly occurring natural disaster, with the Centre for Research on the Epidemiology of Disasters 2021 report on “The Non-COVID Year in Disasters” estimating economic losses worth over USD 51 million and more than 6000 fatalities in 2020. The hydrodynamic models which are used for flood forecasting need to be evaluated and constrained using observations of water depth and extent. While remotely sensed estimates of these variables have already facilitated model evaluation, citizen sensing is emerging as a popular technique to complement real-time flood observations. However, its value for hydraulic model evaluation has not yet been demonstrated. This paper tests the use of crowd-sourced flood observations to quantitatively assess model performance for the first time. The observation set used for performance assessment consists of 32 distributed high water marks and wrack marks provided by the Clarence Valley Council for the 2013 flood event, whose timings of acquisition were unknown. Assuming that these provide information on the peak flow, maximum simulated water levels were compared at observation locations, to calibrate the channel roughness for the hydraulic model LISFLOOD-FP. For each realization of the model, absolute and relative simulation errors were quantified through the root mean squared error (RMSE) and the mean percentage difference (MPD), respectively. Similar information was extracted from 11 hydrometric gauges along the Clarence River and used to constrain the roughness parameter. The calibrated parameter values were identical for both data types and a mean RMSE value of ∼50 cm for peak flow simulation was obtained across all gauges. Results indicate that integrating uncertain flood observations from crowd-sourcing can indeed generate a useful dataset for hydraulic model calibration in ungauged catchments, despite the lack of associated timing information.