Improving 2D hydraulic modelling in floodplain areas with ICESat-2 data: A case study in the Upstream Yellow River

Abstract

Reliable flood inundation modelling in complex river systems that are poorly instrumented is often limited by inaccuracies in open source DEMs, particularly near river channels and vegetated regions. This study proposes a methodology to correct and enhance resolution of satellite based DEMs in floodplain areas with ICESat-2 land elevation, Sentinel-2 MSI imagery, and a simple artificial neural network (ANN). FabDEM (30-m) is selected as the base DEM, and the ANN is trained to correct elevation errors at 10-m resolution using spectral bands from Sentinel-2 and ICESat-2 ATL03 elevation. The corrected ANN DEM reduces the mean squared error by 7 cm on average and up to 38 cm in the areas closer to the main river channel. MIKE 21 is used to simulate 2D flood extent maps for four different events, that consider in-situ discharge values at high, medium and low flow, comparing modelled flood extent with observed surface water extent (SWE) maps derived from Sentinel-2 at the selected dates. To ensure that improvements are attributed to DEM corrections rather than hydraulic parametrization, simulations are performed with different uniform values of the Gauckler-Strickler coefficient $K_s$, which are kept consistent across FabDEM and ANN DEM based scenarios. The critical success index (CSI), F1- score and bias are calculated for simulations with FabDEM and ANN DEM. Across all events, the ANN DEM improves flood simulation accuracy, increasing the Critical Success Index (CSI) and F1 score by up to 19 % and 13 %, respectively, and reducing bias by up to 25 %. This workflow demonstrates a scalable and efficient approach to improve hydraulic model inputs in data-scarce floodplain environments, offering valuable insights for flood risk assessment and water resource management in remote regions.