Mateo-Garcia, Gonzalo
Veitch-Michaelis, Joshua
Smith, Lewis
Oprea, Silviu Vlad
Schumann, Guy
Gal, Yarin
Baydin, Atılım Güneş
Backes, Dietmar
Funding for this research was provided by:
Ministerio de Ciencia e Innovación (TEC2016-77741-R)
Science and Technology Facilities Council (ST/R002673/1)
Engineering and Physical Sciences Research Council (EP/L016427/1, EP/N019474/1)
Article History
Received: 2 August 2020
Accepted: 4 March 2021
First Online: 31 March 2021
Duplicate publication statement
: A short summary of some material in this paper was previously presented at the Humanitarian Aid and Disaster Response (HADR) workshop at NeurIPS 2019 (Vancouver)<sup>66</sup>. The pre-print was peer-reviewed for inclusion in the workshop, which is not archival and does not form part of the NeurIPS conference proceedings. This paper provides substantially more information and addresses several aspects of the dataset that were not discussed in the workshop paper (such as the the effects of training on S2 only, the effect of temporal misalignment and stratification of results on flood and permanent water). We have also extensively curated the training and testing sets and we include a more thorough discussion of the context of our work, flood mapping, and the satellite we have targeted for deployment. Alongside this paper, we also release the <i>WorldFloods</i> dataset as well as the code used to train and benchmark our models.
: The authors declare no competing interests.