Evaluation of the performance of hydrological model LISFLOOD using the ECMWF seasonal meteorological forecast at 1arcmin-1day spatiotemporal resolution over German catchments

Floods and their devastating effects on society and economy have increased dramatically in Germany, and Europe in recent years. At the end of 2023, rivers and streams across Germany burst their banks due to heavy rainfall, affecting property, transport and power supplies and necessitating rescue operations and evacuations to protect human lives. One measure to deal with flooding and safeguard lives and property is the implementation of early warning systems, such as the European Flood Awareness System (EFAS), which provides short-term hydrological forecasts in real time. However, preparedness is essential along the responders value chain and longer term forecasts are important to anticipate, take precautions, raise awareness and generally mitigate the effects of flooding. The objective of this study is to evaluate the performance of hydrological forecasting using the seasonal meteorological forecast at a spatio-temporal resolution of 1 arcmin and day over Germany including all transboundary catchments for the period from 1990 to 2020. The hydrological model used was LISFLOOD. In the first step, LISFLOOD was calibrated using the meteorological observations, the EMO 1arcmin dataset and the discharge data from the transnational hydrological portal for all federal states and neighboring countries. The characteristics of land use, land cover, soil, groundwater, and human activity referred to as surface fields for global environmental modelling, were provided by EFAS. The second step, downscaling of the seasonal (long-term) forecast meteorological forcing to 1arcmin, is performed using a Deep Residual Neural Network (DRNN), and a bilinear interpolation approach over the seasonal forecast information of atmospheric conditions up to seven months into the future provided by the European Center for Medium-Range Weather Forecasts (ECMWF), 25 ensemble members in total. In the third step, the discharge is simulated by feeding the LISFLOOD model with two meteorological forcing scenarios, the DRNN downscaled and the bilinear approach of the seasonal meteorological forecast, to finally compare the performance with the observed runoff using the modified Kling-Gupta efficiency criteria (KGE'). The calibrated and validated LISFLOOD parameters showed a good and acceptable performance in all catchments, KGE' between 0.6 and 0.9. The DRNN downscaling technique shows a promising result, providing a good agreement between downscaled and observed dataset. Finally, the hydrological performance, KGE', is expected to be improved by 0.05 to 1 in the hydrological stations with good and poor performance, respectively, by using the DRNN downscaled seasonal forecast.