Efficient Simulation and Downscaling of Large Non-Stationary Fields with Varying Local Anisotropy

Abstract:

Simulation of locally anisotropic, non-stationary random fields is a relatively new topic in geostatistics with applications currently restricted to the construction of an admissible covariance matrix. In this paper, we introduce an efficient algorithm for constructing large non-stationary random fields with arbitrary local covariance structure and anisotropy. At the heart of our approach is a newly developed robust directional multiresolution framework combined with a local tensor anisotropy model. The use of our algorithm is illustrated with local anisotropy analysis, simulation and downscaling of complex pseudo-precipitation (PP) fields* related to tropical and extra-tropical cyclones. The efficiency of the algorithm allows obtaining realistic downscaled global GCM precipitation fields down to a few kilometers resolution in seconds.
* Reference: Unpublished work by Huiling Yuan and Zoltan Toth. PP fields are constructed by taking the precipitation as the positive component of the field and the water vapor saturation deficit as its negative complement.