Floods are among the most frequent and destructive natural hazards in the world. Between 1980 and 2017 the World was affected by almost 6000 damaging floods that claimed over 220.000 lives [1].
Currently, most runoff predictions are based on models with fixed process descriptions. Their setup requires a lot of domain knowledge and laborious parameter-calibration from historical data. As a result, the obtained models are highly specific for the site or region and handling them requires a lot of domain knowledge. In this project we examine a different approach, which leverages the power of Machine Learning to build models that will ultimately be usable globally.
Our approach is to use a Long Short-Term Memory network (LSTM) as the main part of the model [2]. The network uses meteorological variables —such as solar radiation, average temperature, or precipitation—and catchment descriptors—such as the percentage of sand in the soil, the vegetation type, or the land-use—and provides runoff predictions. The addition of the latter allows the LSTM to learn a more general understanding of the rainfall-runoff context and should allow them to generalise to new locations [3-5].
As our approach depends on training deep neural networks, it requires large computational resources and thus strong hardware infrastructure.
References