This paper describes the application of deep learning approaches that use textual and visual features to flood severity detection in news content. In the context of the MediaEval 2019 Multimedia Satellite (MMSat) workshop, we test the value of transferring models pre-trained on large related corpora, as well as the improvement brought by dual branch models that combine embeddings produced by mixed textual and visual inputs. We compare these model variants using data distributed in the context of the MediaEval MultiMedia Satellite (MMSat) 2019 workshop. The results disclosed in this paper were presented at the workshop: the present paper significantly extends the concise technical notes bound to the results obtained on the test sets provided by the workshop organizers.