In recent years, Social Media platforms have attracted millions of users, becoming a primary communication channel. They offer the possibility to massively ingest and instantly share big volumes of user-generated content before, during, and after emergency events. Being able to accurately quantify the impact of such hazardous events could greatly help all organizations involved in the emergency management cycle to adequately plan the required recovery operations. In this work, we propose a novel Natural Language Processing approach built on rule-based algorithms able to estimate, from tweets posted during natural hazards, the impact of emergency events in terms of affected population and infrastructures. We implement our approach in an operational environment and present its validation on a publicly released dataset of more than 1.4K manually annotated tweets, showing an overall weighted F1 score of 0.77.