Development of an operational early flood forecasting system and 11-year validation of forecast skill over Japan

Floods are among the most devastating threats to our society in the form of economic damage and casualties, and flood-forecasting systems mitigate these damages by securing time for disaster-preparation and evacuation. We recently developed IGAR-F3 (Integrated Global And Regional Framework for Flood Forecasting) and set it up as an operational flood forecasting system for Japan. IGAR-F3 forecasts hydrological / hydraulic variables such as soil moisture, discharge, and inundated area using a numerical weather prediction dataset, a land surface model, and a hydrodynamic model. IGAR-F3 covers a Japanese domain (123-148 E, 24-46N) with a 39-hour lead-time at 0.05 degree / 1-h resolution using numerical weather predictions issued by the Japan Meteorological Agency Meso-Scale-Model. IGAR-F3 makes a forecast every 3 hours, and disseminates these via web user-interface. In addition to this forecast mode, IGAR-F3 also has a Near-Real-Time (NRT) mode that ingests radar-observed precipitation to monitor current hydrological / hydraulic states. The initial parametrization of each 3-h forecast is derived from an NRT run to avoid propagation of forecast error. To validate forecast skill, we conducted long-term validation on more than 32,000 forecasts during an 11-year reforecast (2007-2017). The validation for discharge forecasts using Pierce's Skill Score (PSS) showed that forecasts with 33-h lead time have a predictability (0.00 < PSS) at more than 90% out of 849 observation stations over Japan. In the forecasts with 12-h lead-time, IGAR-F3 has a relatively high predictability (0.25<PSS) at more than 50% out of 849 stations, but the remaining stations have lower skill. Previous studies tend to have PSS around 0.25 or better, and this typically results from a well-calibrated model developed for a single basin. Our forecasts match this skill on about half of our stations, despite the lack of basin-specific calibration. Further differences in the skill between IGAR-F3 and literature could come from weather forcing, the initialization processes, or our physical representation of hydrologic processes. Finally, we show our forecast results for the catastrophic flood events in July 2018 in Japan, and discuss potential improvements for realizing better early warning and for extending the forecast system to the global-scale.