Machine Learning for Flood and Drought Forecasting

Floods and droughts are the costliest natural disasters in Canada, but impacts on live and livelyhoods can be mitigated with sufficient warning. Currently, flood and drought forecasting is done by govenment agencies and primarily based on semi-empirical hydrological models. This presentation will start with a brief review of current hydrological forecasting methods and their limitations, before introducing a new machine learning approach to streamflow forecasting that has received significant attention in the scientific literature for its versatilty and high forecast skill. We will outline what enabled the success of this new approach and why it yields superior skill to previous methods, in particular in ungauged watersheds (which has been a long-standing challenge in hydrology). Finally, the challenges of operational implementation of this method as a forecasting service will be discussed, using Aquanty's real-time hydrological forecasting platform as an example. Such challenges include real-time processing of numerical weather forecasts, availability of input data, and aspects of stakeholder engagement and gaining trust with the end-user.