Reproducing flash flood warnings with Machine Learning

Flash floods account for a large proportion of flood-based fatalities, and they are becoming more frequent due to climate change. A global flash flood warning system therefore has the potential to be life saving.
Standard approaches to flash flood forecasting – such as the Flash Flood Guidance (FFG) system in the US National Weather Service (NWS), or the European Runoff Index based on Climatology (ERIC) in the European Flood Awareness System (EFAS) – are utilizing recent weather and soil conditions, physiographic characteristics of basins, and weather forecasts, in order to produce a forecast for possible flash floods. These forecasts are not disseminated directly to the public. Instead, they are firstly refined by hydrologists that have intimate knowledge of the relevant basins and of previous flood events. This introduces a difficulty to scaling these methods worldwide, as the training of professional hydrologists in every region is costly and time consuming.
Recent applications of Machine Learning (ML) to hydrology show that a learning system has the potential to train on data-rich basins and generalize to data-poor basins, with a skill that is comparable to state of the art hydrological models. 
In this work we attempt to build a ML model to produce daily flash flood forecasts, based on globally available weather reanalysis and physiographic characteristics from HydroATLAS. We discuss the model architecture, and evaluate it against NWS Flash Flood Warnings (FFW). While such models may not surpass the skill of a professional hydrologist, they have the potential to provide reasonable warnings in regions that do not currently have any such system in place.