Assessing the spatial spread–skill of ensemble flood maps with remote-sensing observations

Abstract. An ensemble of forecast flood inundation maps has the potential to represent the uncertainty in the flood forecast and provide a location-specific likelihood of flooding. Ensemble flood map forecasts provide probabilistic information to flood forecasters, flood risk managers and insurers and will ultimately benefit people living in flood-prone areas. Spatial verification of the ensemble flood map forecast against remotely observed flooding is important to understand both the skill of the ensemble forecast and the uncertainty represented in the variation or spread of the individual ensemble-member flood maps. In atmospheric sciences, a scale-selective approach has been used to evaluate a convective precipitation ensemble forecast. This determines a skilful scale (agreement scale) of ensemble performance by locally computing a skill metric across a range of length scales. By extending this approach through a new application, we evaluate the spatial predictability and the spatial spread–skill of an ensemble flood forecast across a domain of interest. The spatial spread–skill method computes an agreement scale at every grid cell between each unique pair of ensemble flood maps (ensemble spatial spread) and between each ensemble flood map with a SAR-derived flood map (ensemble spatial skill; SAR: synthetic aperture radar). These two are compared to produce the final spatial spread–skill performance. These methods are applied to the August 2017 flood event on the Brahmaputra River in the Assam region of India. Both the spatial skill and spread–skill relationship vary with location and can be linked to the physical characteristics of the flooding event such as the location of heavy precipitation. During monitoring of flood inundation accuracy in operational forecasting systems, validation and mapping of the spatial spread–skill relationship would allow better quantification of forecast systematic biases and uncertainties. This would be particularly useful for ungauged catchments where forecast streamflows are uncalibrated and would enable targeted model improvements to be made across different parts of the forecast chain.