Development of a Globally Connected, Climate-Driven, Stochastic Drought Model for Hazard Assessment using Machine Learning Techniques

With global temperatures continuing to rise year on year, drought conditions are becoming increasingly frequent and severe, across all continents. More and more, the negative effects of these worsening drought conditions are being experienced by people across the world both directly, through damage to agricultural systems, water scarcity or damage to homes from subsidence, as well as indirectly, through cascading effects on other perils such as heatwaves and wildfires, which in turn may devastate communities and drive great economic losses. For these reasons, drought is of growing concern to the (re)insurance industry, as an emerging peril. It is therefore essential that reinsurers have access to tools which can aid in their understanding of drought hazard and risk in a changing climate. One such tool we present here – a climate driven, globally connected stochastic drought hazard model, which responds dynamically to the climate of any given year, enabling this understanding of how drought conditions change with the climate.

In this presentation, we describe the novel methodology applied to generate this globally connected and climate-driven stochastic drought model. The model is generated in two stages, the first addressing global variability in drought trends and teleconnections, and the second looking at continental scale patterns. In the first instance, we apply a dimensionality reduction to a selection of historical drought indexes over different time scales, allowing extraction of the key modes of variability of drought at the global scale. We then condition the top key modes of variability to the climate state using reanalysis (ERA5) data, allowing us to drive our stochastic set at the global scale, based on the global climate state.

Once these global patterns have been determined, we use the residual drought signal to condition a regional (continental) model using similar reduction and conditioning techniques. This regional layer is then effectively layered onto the global model, allowing us to recreate globally and regionally consistent drought variability in the stochastic set. A Bayesian framework is used to sample a range of realistic drought conditions, aligned with the climate of any given year. Global and regional drought conditions are then combined in order to generate >100K stochastic years of global drought severity as well as duration of drought for three severity levels (moderate, severe, extreme). This framework can also be applied to any other climate model data (for example, CESM LENS2) to generate a stochastic event set up to the year 2100. Here we present initial results from this stochastic catalogue, showcasing the spatial and temporal variation in drought hazard from 1950 – 2100, return periods, and comparisons to historical records. This work also builds upon a previous, continental only version of the drought model.