Zerenner, Tanja: Atmospheric Downscaling using Multi-Objective Genetic Programming. - Bonn, 2017. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5n-48408
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5n-48408,
author = {{Tanja Zerenner}},
title = {Atmospheric Downscaling using Multi-Objective Genetic Programming},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2017,
month = oct,

volume = 80,
note = {Numerical models are used to simulate and to understand the interplay of physical processes in the atmosphere, and to generate weather predictions and climate projections. However, due to the high computational cost of atmospheric models, discrepancies between required and available spatial resolution of modeled atmospheric data occur frequently. One approach to generate higher-resolution atmospheric data from coarse atmospheric model output is statistical downscaling.
The present work introduces multi-objective Genetic Programming (MOGP) as a method for downscaling atmospheric data. MOGP is applied to evolve downscaling rules, i.e., statistical relations mapping coarse-scale atmospheric information to the point scale or to a higher-resolution grid. Unlike classical regression approaches, where the structure of the regression model has to be predefined, Genetic Programming evolves both model structure and model parameters simultaneously. Thus, MOGP can flexibly capture nonlinear and multivariate predictor-predictand relations. Classical linear regression predicts the expected value of the predictand given a realization of predictors minimizing the root mean square error (RMSE) but in general underestimating variance. With the multi-objective approach multiple cost/fitness functions can be considered which are not solely aimed at the minimization of the RMSE, but simultaneously consider variance and probability distribution based measures.
Two areas of application of MOGP for atmospheric downscaling are presented: The downscaling of mesoscale near-surface atmospheric fields from 2.8 km to 400 m grid spacing and the downscaling of temperature and precipitation series from a global reanalysis to a set of local stations.
(1) With growing computational power, integrated modeling platforms, coupling atmospheric models to land surface and hydrological/subsurface models are increasingly used to account for interactions and feedback processes between the different components of the soil-vegetation-atmosphere system. Due to the small-scale heterogeneity of land surface and subsurface, land surface and subsurface models require a small grid spacing, which is computationally unfeasible for atmospheric models. Hence, in many integrated modeling systems, a scale gap occurs between atmospheric model component and the land surface/subsurface components, which potentially introduces biases in the estimation of the turbulent exchange fluxes at the surface. Under the assumption that the near surface atmospheric boundary layer is significantly influenced by land surface heterogeneity, MOGP is used to evolve downscaling rules that recover high-resolution near-surface fields of various atmospheric variables (temperature, wind speed, etc.) from coarser atmospheric data and high-resolution land surface information. For this application MOGP does not significantly reduce the RMSE compared to a pure interpolation. However, (depending on the state variable under consideration) large parts of the spatial variability can be restored without any or only a small increase in RMSE.
(2) Climate change impact studies often require local information while the general circulation models used to create climate projections provide output with a grid spacing in the order of approximately 100~km. MOGP is applied to estimate the local daily maximum, minimum and mean temperature and the daily accumulated precipitation at selected stations in Europe from global reanalysis data. Results are compared to standard regression approaches. While for temperature classical linear regression already achieves very good results and outperforms MOGP, the results of MOGP for precipitation downscaling are promising and outperform a standard generalized linear model. Especially the good representation of precipitation extremes and spatial correlation (with the latter not incorporated in the objectives) are encouraging.},

url = {https://hdl.handle.net/20.500.11811/7264}

The following license files are associated with this item: