A meta-Gaussian stochastic rainfield generator for France

The RAINSIM stochastic daily rainfield generator is based on weather pattern sub-sampling and meta-gaussian models (Ayar et al., 2020). RAINSIM is coupled with an air temperature generator to feed a a distributed hydrological model, allowing to simulate large hydrological chronicles for extreme estimation (both floods and low flows). A large-scale application of RAINSIM to the whole French continental territory (about 500 000 km²) is presented here.

The parameters of the statistical models (both at-site distributions and temporal and spatial covariances) are infered from observed precipitation data at stations, sub-sampled into subsets by seasons and weather types. Before the rainfield generation, sequences of weather types are generated by a Markov model. Here the seasonal transition matrixes are conditionned to observed large-scale climatic indexes such as NAO and WeMO. This conditionning allows a better representation of the year-to-year and decadal variabilities.

The presented application challenges a key assumption of RAINSIM: the stationarity of the spatial covariance.  At the French scale, the diversity of climatology and of the spatial structures of rain fields are significant, thus questioning this hypothesis. To tackle this, an approach based on the deformation of the geographical space (Monestiez et al., 2007) has been tested, thanks to its implementation in the deform R-package (Youngman, 2023). The deformations are computed independently for each subset, illustrating that the spatial covariance structure of the rain fields depends on the weather, and to a lesser extend to the season. Comparisons to observed data with suitable metrics are presented to score this use of covariance-oriented deformations of space.

Perspectives and first developments for application in projected climate are also evoked.

 

References:

Ayar, P. V., Blanchet, J., Paquet, E., & Penot, D. (2020). Space-time simulation of precipitation based on weather pattern sub-sampling and meta-Gaussian model. Journal of Hydrology, 581, 124451.

Monestiez, P., Meiring, W., Sampson, P. D., & Guttorp, P. (2007). Modelling Non‐Stationary Spatial Covariance Structure from Space—Time Monitoring Data. In Ciba Foundation Symposium 210‐Precision Agriculture: Spatial and Temporal Variability of Environmental Quality: Precision Agriculture: Spatial and Temporal Variability of Environmental Quality: Ciba Foundation Symposium 210 (pp. 38-51). Chichester, UK: John Wiley & Sons, Ltd..

Youngman, B. D. (2023). deform: An R Package for Nonstationary Spatial Gaussian Process Models by Deformations and Dimension Expansion. arXiv preprint arXiv:2311.05272.