Downscaling near-surface atmospheric fields with multi-objective Genetic Programming

Abstract

We present a new Genetic Programming based method to derive downscaling rules (i.e., functions or short programs) generating realistic high-resolution fields of atmospheric state variables near the surface given coarser-scale atmospheric information and high-resolution information on land surface properties. Such downscaling rules can be applied in coupled subsurface-land surface-atmosphere simulations or to generate high-resolution atmospheric input data for offline applications of land surface and subsurface models. Multiple features of the high-resolution fields, such as the spatial distribution of subgrid-scale variance, serve as objectives. The downscaling rules take an interpretable form and contain on average about 5 mathematical operations. The method is applied to downscale 10 m-temperature fields from 2.8 km to 400 m grid resolution. A large part of the spatial variability is reproduced, also in stable nighttime situations, which generate very heterogeneous near-surface temperature fields in regions with distinct topography.