Leonardo Santos, Luiz Fernando Satolo, Ricardo Oyarzabal, Elton Escobar-Silva, Michael Diniz, Rogério Negri, Glauston Lima, Stephan Stephany, Jaqueline Soares, Johan Duque, Fernando Saraiva-Filho, Luiz Bacelar

Background: flash flood modeling faces many challenges since physically-based hydrological models are unsuitable for a small 23 spatiotemporal scale. With the increased availability of hydrological observed data, an alternative approach is to use machine 24 learning (ML) techniques. This work conducts a Systematic Literature Review (SLR) to enhance our comprehension of the research 25 landscape on ML applications for modeling flash floods. 26
Methods: starting with more than 1,200 papers published until January 2024 and indexed in Web of Science, SCOPUS/Elsevier, 27 Springer/Nature, or Wiley databases, it was selected 50 for detailed analysis, following the PRISMA guidelines. The 28 inclusion/exclusion criteria removed reviews, retractions, and papers that were not in the scope of this SLR and included only 29 papers that used data with a temporal resolution finer than 6 hours. From each selected paper, among other information, data were 30 extracted regarding the forecasting horizon, the size of the study area, the different input data, the chosen machine learning 31 technique, and the type of outcome in order to characterize the model applied to flash flood forecasting. 32
Results and Discussion: there has been a notable increase in publications investigating ML techniques for flash flood modeling 33 over the last few years. Most of the studies are performed in China (38%). In 49 out of 50 of the selected papers used as input data 34 just one or an exclusive combination of the following measurements: discharge, rainfall, and water level. From this set, the 35 combination of discharge and rainfall appears in almost 40% of the papers. Notably, 60% of the studies utilize the long short-term 36 memory (LSTM) method. No method consistently outperforms all others in the selected papers. Unfortunately, only 10% of the 37 selected articles provide access to their data. We recommend integration into early warning systems, development and 38 dissemination of benchmarks, publication of successful case studies, and multidisciplinary collaboration.