Downscaling near-surface atmospheric fields with multi-objective genetic programming

Coupled models of the soil-vegetation-atmosphere systems are increasingly used to investigate interactions between the system components. Due to the different spatial and temporal scales of relevant processes and computational restrictions, the atmospheric model generally has a lower spatial resolution than the land surface and subsurface models. We employ multi-objective Genetic Programming (MOGP) using the Strength Pareto Evolutionary Algorithm (SPEA) to bridge this scale gap. We generate high-resolution atmospheric fields using the coarse atmospheric model output and high-resolution land surface information (e.g., topography) as predictors. High-resolution atmospheric simulations serve as reference. It is impossible to perfectly reconstruct the reference fields with the available information. Thus, we simultaneously optimize the root mean square error (RMSE) and two objective functions quantifying spatial variability. Minimization solely with respect to the RMSE provides too smooth high-resolution fields. Additional objectives help to recover spatial variability. We apply MOGP to the downscaling of 10 m temperature. Our approach reproduces a larger part of the variability and is applicable for a wider range of weather conditions than a linear regression based downscaling.

Original publication: T. Zerenner, V. Venema, P. Friederichs, and C. Simmer. Downscaling near-surface atmospheric fields with multiobjective Genetic Programming. Environmental Modelling and Software, 84(2016), 85--98.