Simulating daily rainfall fields over large areas for collective risk estimation

- A class of daily rainfall generators for large areas is introduced.
- VGLM/VGAM and GAMLSS are combined with meta-Gaussian random fields.
- Heavy tailed and apparently heavy tailed distributions correct for overdispersion.
- Pairwise upper tail dependence can be spurious.
- Large area circulation indices are incorporated in the model structure.

SummaryLarge scale rainfall models are needed for collective risk estimation in flood insurance, infrastructure networks and water resource management applications. There is a lack of models which can provide simulations over large river basins (potentially multi-national) at appropriate spatial resolution (e.g., 5-25 km) that preserve both the local properties of rainfall (i.e., marginal distributions and temporal correlation) and the spatial structure of the field (i.e., the spatial dependence structure). In this study we describe a methodology which merges meta-Gaussian random fields and generalized additive models to simulate realistic rainfall fields at daily time scale over large areas. Unlike other techniques previously proposed in the literature, the suggested approach does not split the rainfall occurrence and intensity processes and resorts to a unique discrete-continuous distribution to reproduce the local properties of rainfall. This choice allows the use of a unique meta-Gaussian spatio-temporal random field substrate that is devised to reproduce the spatial properties and the short term temporal characteristics of the observed precipitation. The model is calibrated and tested on a 25 km gridded daily rainfall data set covering the 817000km2 of the Danube basin. Standard and ad hoc diagnostics highlight the overall good performance over the whole range of rainfall values at multiple scales of spatio-temporal aggregation with particular attention to extreme values. Moreover, the modular structure of the model allows for refinements, adaptation to different areas and the introduction of exogenous forcing variables, thus making it a valuable tool for classical hydrologic analyses as well as for new challenges of network and reinsurance risk assessment over extensive areas.