Floods are among the most damaging natural hazard in the world. In south of France, most of the flood events are called ‘flash flood’, because of the fast response of rivers to rainfall. This is mainly due to the particular climatic setup of this area which is responsible for intense storm, with large amount of rainfall within a few hours on limited areas. In this project, we study these flood events in basins connected with karst aquifers. These ‘karst flash floods’ are specific because karst terrains will react differently than ‘classical’ basins. Depending on the initial water table level, the karst may reduce or amplify the flood in the river. This specific behaviour of karst basins is a key issue for the French national services of flood prediction which finance the project. Most of flood modelling frameworks are designed for surface processes only and do not allow the explicit representation of the karst processes. This explains the poor performances obtained by these models when applied on karst basins. On the opposite, hydrogeological models do not allow an efficie nt representation of surface processes. The aim of the project is to analyse karst flash flood in order to understand these processes and their influence on surface flood. This project concerns eight basins in two departments (Aude and Herault) of southern France, linked with different karst aquifer (binary and unary karsts, from Cambrian to Cretaceous ages). The first step consisted in gathering a large database on these basins, at a specific spatial (surface basin) and temporal (flood duration) scale, and to pre-process these data. This poster presents our database, the uncertainties associated with each data, and how it will be useful for analyses and modelling. This database contains various kinds of data (rainfall, piezometric level, soil moisture, discharge, geochemical analyses of water during flood, inhabitant testimonies, flood marks and other qualitative data), but our goal is to integrate more data from post crisis enquiries in our database.