AI-enhanced drought forecasting: a case study in the Netherlands

Drought is a natural phenomenon linked to a temporary but significant reduction in the availability of water resources. Drought usually originates as a deficit in precipitation, with prolonged drought having substantial repercussions on the hydrological, agricultural and socio-economic sectors; making drought one of the most impactful natural hazards modern society faces. The ability to forecast the occurrence of drought events with sufficient lead time, however, allows for the implementation of strategies to reduce drought impacts. Although drought forecasting using both statistical and dynamic techniques has been widely studied, challenges still remain in predicting drought events, especially for sub-seasonal to seasonal forecasts. Because of the increased availability of Earth Observation data, advances in Artificial Intelligence, and progress in computing capabilities in the last decades, drought prediction has received a new impulse. Machine Learning, especially Deep Learning, techniques are now increasingly being used both to improve current weather forecasts and as an alternative to conventional predictions of extreme events.

In this contribution we explore the use of Machine Learning techniques to improve meteorological drought prediction through post-processing of weather forecast analogues. To this aim, we use both ECMWF extended and long-range forecasts, together with reanalysis data, to build a ML-based model that helps correcting forecasts. We then test the model to explore how much current forecasts can be actually improved with the use of AI-based techniques. We apply the method proposed, in the area of the Rhine Delta in the Netherlands, focussing on 1-month lead time predictions. This work is part of the CLImate INTelligence (CLINT) project, which aims at developing AI-enhanced Climate Services for extreme events detection, causation, and attribution.