This paper deals with the prediction of the total amount of a drought episode under the French "Catastrophe Naturelle" regime. Due to the specificity of this regime, a quick prediction of the final amount of an incident is particularly strategic. The approach that we use is based on a database constituted by the French Federation of Insurers in order to cover approximately 70% of the French market. Linking it with meteorological data and socioeconomic data allows to increase our vision of the exposure. Although the database is large, with a wide vision of the French metropolitan territory, data is imbalanced since a large majority of cities are not stroke by catastrophic events. Machine learning methods are used to compute a prediction.