Flooding is one of the most harmful natural disasters, as it poses danger to both buildings and human lives. Therefore, it is fundamental to monitor these disasters to define prevention strategies and help authorities in damage control. With the wide use of portable devices (e.g., smartphones), there is an increase of the documentation and communication of flood events in social media. However, the use of these data in monitoring systems is not straightforward and depends on the creation of effective recognition strategies. In this paper, we propose a fusion-based recognition system for detecting flooding events in images extracted from social media. We propose two new graph-based early-fusion methods, which consider multiple descriptions and modalities to generate an effective image representation. Our results demonstrate that the proposed methods yield better results than a traditional early-fusion method and a specialized deep neural network fusion solution.