Résumé
Flash floods are among the deadliest natural hazards. With the increase of worldwide population, many more lives could be at risk. Hydrological modeling methods have significantly evolved in the last years. Also, new methods have emerged like machine learning. However, it is still challenging to anticipate and plan crisis management actions like evacuation. Crisis managers and hydrologists often work individually. Hydrologists will often forecast hydrological variables like discharge. Nevertheless, apart from big cities, crisis management plans are rarely based on this variable. The forecasts are, therefore, unlikely to be used regardless of their informational value. Thus, this study aims to identify relevant information for crisis management using different configurations of forecast bulletins.The study site is the flash-flood prone watershed of Gardon d’Anduze (545 km²) in Anduze, Southern France. Floods can cause the discharge to increase from 100 l/s to 3000 m3/s in about ten hours. Using artificial neural networks, water-level and discharge were forecasted. From these forecasts, six different bulletins were proposed with variant types and levels of information like future rain, discharge at upstream watersheds and threshold markers. To evaluate their relevancy, records of decisions, observation sheets, discussions and surveys were used. The bulletins were then tested in a crisis management simulation setting with operational managers. The main findings are: (1) the usefulness of reference to previous events depended on each crisis manager’s personal experience, (2) discharge was difficult to interpret, (3) confidence in forecast bulletins was medium to high, (4) crisis managers were generally able to make decisions with satisfying degrees of anticipation regardless of the level of information, (5) they also preferred bulletins with the highest level of information due to an increase in visibility. This work showcased the importance of flash-flood forecast bulletins in crisis management and the reliability of neural networks forecasts.