On the Spatiotemporal Impacts of Flood Extent Assimilation

Recent studies have demonstrated the potential of assimilating flood extents derived from Synthetic Aperture Radar (SAR) imagery for improved flood forecasts. The increasing number of high resolution SAR sensors which provide a partial coverage of flooded areas can benefit from knowledge on the impacts of spatial location on the reduction of forecast error. This can inform acquisition planning for flood extent assimilation, ensuring optimal utilization of resources at minimum cost. Twin experiments were set up at 90m grid resolution using the 2D hydraulic model LISFLOOD-FP for a 100 year flood event in the Clarence Catchment of NSW, Australia. Synthetic observations were assimilated using the Particle Filter (PF) algorithm, to assess forecast performance sensitivity to spatiotemporal nature of image acquisition. The truth run was setup using a synthetic triangular inflow hydrograph, calibrated channel roughness, and high-accuracy LiDAR floodplain topography supplemented with bathymetric data. Gaussian errors were added to the boundary pixels of observed flood extents, to generate synthetic observations at the assimilation time steps. The open loop was constructed by applying perturbed (i) inflows, (ii) channel roughness, and (iii) topography, represented by four different globally available digital elevation models (DEM). Three regions of interest were selected based on sub-reach flow characteristics and delineated according to flow distances. While Sub-reach 1 could be adequately approximated using a kinematic wave model, Sub-reach 2 and 3 deviated significantly due to meandering and braiding respectively. Modelled binary flood extents were compared with the synthetic observations locally at each assimilation time step and each particle was weighted based on the agreement between them. Preliminary results indicate that assimilating a single observation covering Sub-reach 3 was able to reduce the spatiotemporally averaged RMSE in simulated water depth by almost 36% locally; while assimilating flood extents at Sub-reach 1 and 2 degraded local performance by 198% and 27% respectively. This implies that the spatiotemporal coverage of the observation for flood extent assimilation, needs to be optimized not only for maximum improvement in forecast accuracy but to avoid performance degradation.